Synthszr Glossary

The 1260H List is a directory published by the US Department of Defense of Chinese companies that the US considers to be linked to the Chinese military. An entry does not bring a direct sales ban, but it damages the companies' reputations and will exclude them from US military contracts starting in 2026.

3D NAND is a construction method for memory chips in which the memory cells are stacked in many layers on top of each other instead of merely side by side. This allows significantly more data to fit on chips of the same size – the technology is now found in almost every SSD, in smartphones, and in memory cards.

A 3D white model is a rough spatial draft of an object or building that shows only shape and size while remaining uniformly white. Colors, materials, and details are added later, once the basic form is settled.

4-Bit Floating Point is an extremely frugal number format in which a single number occupies just four bits of memory, allowing it to take on only 16 different values. AI models use it to compute with significantly less memory and power, accepting a small loss in accuracy in return.

4D Creation refers to methods that allow computer programs to generate a spatial object from a text or a video that additionally moves over time. The result is not a flat video, but a moving 3D scene that can be viewed from all sides.

800G refers to network connections that transmit 800 gigabits per second — roughly 800 billion signal states per second. Such lines connect the computers in data centers that train AI models.

The A-Series chip is the central computing component that Apple designs itself for iPhones and some iPads. It combines a main processor, a graphics unit, and its own compute part for AI tasks on a single piece of silicon.

An A/B test is a comparison in which two versions of something are delivered simultaneously to randomly split groups of users. This allows measuring which version triggers the desired behavior more often, instead of just guessing.

The Aarhus Convention is an international treaty that guarantees citizens three rights in environmental matters: access to environmental information, participation in decision-making, and access to justice. It was signed in 1998 in Aarhus, Denmark, and is applicable law in the EU and in Germany.

An ablation study is a test in which individual parts of an AI system are deliberately removed or switched off to see how much the performance suffers as a result. This reveals which components truly contribute to success and which are just along for the ride.

Access to Justice refers to the question of whether people can actually enforce their rights – and not fail simply because lawyers are too expensive, proceedings take too long, or forms are too complicated. Digital tools and AI programs are meant to lower these hurdles, but they raise new questions about reliability and accountability.

An Account Executive works in a company's sales department and is responsible for turning an interested prospect into a paying customer. In the tech industry, account executives mainly sell software subscriptions to other companies and carry a fixed revenue target for doing so.

Account-based marketing is a sales strategy in which a company does not try to reach as many customers as possible, but instead specifically targets individual dream clients. For each of these companies, a tailor-made approach is developed.

An ACP integration connects a computer program with an AI assistant via the Agent Communication Protocol, a defined shared language for such exchanges. This allows programs from different vendors to exchange tasks without requiring custom extra work for every combination.

An acqui-hire is the purchase of a company where the buyer is primarily interested in its employees and hardly at all in the product. Especially in the AI industry, large corporations pay high sums this way to get experienced developer teams as a whole.

An acquihire is a company purchase where the buyer is primarily interested in the employees rather than the product. Especially in the AI industry, large corporations pay high sums for small teams with rare expertise.

Action Prediction refers to methods that predict from ongoing observations which action a human or a machine will perform next. Such predictions are found in self-driving cars, in robots, and in cameras that analyze movements.

Activation Capping is a technique used in building AI systems where the intermediate results inside a network are bounded from above and below. This prevents individual extremely large numbers from disrupting the computation or causing training to crash.

Active Directory is a directory service from Microsoft that centrally manages all user accounts, passwords, and computers within an organization. It decides at every login who is allowed to log in and which data that person can access.

Active Parameters refers to the portion of an AI model's adjustable numerical values that actually participates in computation for a single request. In modern language models, this is often just a fraction of the total, which makes operation significantly cheaper.

Activity Metrics are numbers that measure how often and how intensively people actually use a product. They show activity, but on their own say nothing about whether a company is making money from it.

ActivityPub is a common language that lets different social networks exchange posts and subscriptions. Anyone signed up with one service can use it to follow people on a completely different service.

The Adam optimizer is a computational method that, during the training of AI models, determines how strongly the internal parameter values are changed after each learning step. It adjusts the step size individually for each value and has been the standard method for training neural networks since 2015.

Adapters are small additional building blocks inserted into an already trained AI model to adapt it to a new task. Instead of retraining the entire model, only these few building blocks are changed – saving computing time, memory, and money.

The Adapter Pattern is a proven solution in software development: an intermediary piece translates the language of one program part into the language of another, so that both can work together without being changed. The name comes from the travel adapter that lets a German plug fit into a British socket.

Adaptive reasoning means an AI system decides for itself how much thinking work a task deserves: simple things get answered right away, difficult things get more computing time. This saves costs while simultaneously improving quality on tricky questions.

ADAS is the umbrella term for electronic helpers in a car that support the driver with steering, braking, and monitoring the surroundings. These include emergency brake assist, lane-keeping assist, and adaptive cruise control – but the human remains responsible.

An add-in is a small supplementary program that extends an existing program with new features without altering its basic structure. Add-ins are best known from Office programs like Excel or Word, where today they often retrofit AI functions.

An admission handshake is a brief exchange of messages that lets a computer service clarify in advance whether it can accept a new request at all. Only once both sides agree does the actual work begin — otherwise the request is immediately rejected or placed in a queue.

The adoption rate measures what proportion of a group actually uses a new technology. It shows whether an innovation is truly catching on or merely receiving a lot of attention.

The addressable market is the revenue a company could theoretically achieve with a product if it won over every possible customer. The figure is an estimate and describes an upper limit, not a realistic revenue amount.

ADS-B is a radio technology that allows aircraft to continuously broadcast their own position, altitude, and speed. Because these signals are unencrypted, anyone with a cheap receiver can pick them up – which is what flight-tracking sites like Flightradar24 rely on.

AdTech is the umbrella term for all programs and systems used to automatically select, auction, and deliver advertising on the internet. It decides in fractions of a second which ad an individual user gets to see on a website or in an app.

The Advanced Encryption Standard, or AES for short, is a globally used method for making data unreadable and later readable again. It secures banking transactions, WiFi connections, messenger messages, and encrypted hard drives.

Advanced Packaging refers to modern processes used to assemble multiple chips extremely densely into a single component. Because individual chips can barely be shrunk any further, this assembly now plays a major role in determining how fast AI accelerators can compute.

Advantage+ is a family of products from Meta in which software, rather than humans, decides who sees an ad on Facebook and Instagram. The advertiser only specifies the goal, budget, and creative material, while the system handles the fine-tuning automatically.

An adversarial example is an input that looks completely normal to humans but is specifically designed to mislead a computer program for image recognition or text analysis. Even tiny, barely visible changes to a photo can cause the program to recognize something completely different.

Adversarial training is a technique in which an AI system is deliberately trained with manipulated examples that would normally deceive it. This makes it more robust against intentional attacks and against unusual inputs.

An aeroderivative gas turbine is a power generator whose technology is derived from aircraft engines. It is smaller and lighter than classic power plant turbines and delivers full output within just a few minutes.

AES is the world's most widely used method for encrypting data so that only someone with the matching key can read it. It is built into Wi-Fi, messengers, bank cards, and hard drive encryption — usually invisible in the background.

Affiliate marketing is a form of advertising in which a partner sends customers to a provider and only gets paid if this actually results in a purchase or a sign-up. So it's not the ad space that gets paid for, but the outcome.

The Agency Quotient is a proposed metric for how independently a computer program acts – that is, how much it decides and carries out without asking a human. It is not a uniformly standardized measurement, but rather a conceptual figure that appears in debates about autonomously operating AI systems.

Agency-as-a-Service refers to services that used to be provided by an agency with human staff and that are now largely handled by software with artificial intelligence. Customers don't book a tool or hours of work, but a finished result for a monthly fee.

An agent is an AI system that doesn't just respond, but breaks a task down into steps on its own and uses tools such as search engines, programs, or databases in the process. So it keeps working until a goal is reached, instead of stopping after a single answer.

The Agent Client Protocol (ACP) is a common language through which a coding tool talks to an AI assistant. It ensures that the same assistant works across many different editors without having to be adapted anew for each one.

Agent Hooks are predefined checkpoints in the workflow of an AI assistant where custom code is automatically executed. Developers use them to check, log, or deliberately halt the assistant's behavior.

Agent Memory refers to the ability of an autonomously operating AI program to store information from previous tasks and conversations and reuse it later. Without this memory, such a program starts from scratch every time it is launched.

Agent Mode is an operating mode of AI programs in which the software doesn't just answer a task but works through it step by step on its own. In doing so, it may use tools, such as searching the internet, modifying files, or launching programs.

Agent-as-a-Judge means: A computer program using Artificial Intelligence evaluates the work of another such program — and is allowed to take action itself in doing so, i.e. open files, execute code, or ask follow-up questions. It is the further development of methods in which an AI only assesses the finished answer text of another system.

An agent framework is a ready-made collection of software components that developers use to build programs which independently break a task down into sub-steps and work through those steps one after another. It provides the same recurring building blocks every time: memory, access to tools such as search engines, and a loop that controls the process.

An agent scaffold is the program built around an AI language model so that it doesn't just output text but independently carries out multi-step tasks. It determines which tools the model is allowed to use, in what order it works, and when it stops.

An agent harness is the program scaffolding built around an AI language model that turns its responses into actual work steps: reading files, executing commands, reporting back results. The model thinks, the harness acts and controls.

Agent identification refers to procedures by which a computer program that autonomously performs tasks on the internet identifies itself to other systems. It aims to clarify who is behind an automated access and what exactly this program is allowed to do.

The agent runtime is the software that controls an AI program while it works: it calls the language model, executes its tool commands, and returns the results. Without it, a model could formulate answers, but never actually carry out any of them.

An agent loop is the cycle in which an AI program works through a task step by step: it thinks, carries out an action, looks at the result, and then decides anew. This loop repeats until the goal is reached or a limit kicks in.

Agent plugins are add-on modules that let an AI program do things it cannot do on its own: calculate, search the web, place an order. They extend a language model from a pure text machine into a tool that genuinely gets things done in other software.

An agent runtime is the software that controls an AI program while it works: it calls the language model, executes its tool calls, and keeps track of the intermediate state. It is thus the environment in which a text model turns into an acting program.

Agent-to-agent communication refers to the direct exchange between two or more AI programs that independently carry out tasks. Instead of a human triggering every intermediate step, the programs pass on assignments, intermediate results, and follow-up questions among themselves.

Agent-to-system connectivity refers to the technical connection between an AI program that autonomously carries out tasks and a company's software in which the actual data and processes reside. It determines whether such a program merely writes text or actually places orders, enters appointments, and checks invoices.

An agent framework is a ready-made software foundation that developers use to build programs that carry out tasks independently in multiple steps. It provides the building blocks for this: planning, memory, access to tools such as search engines, and the repetition of steps until the goal is reached.

An agent loop is a repeated process in which an AI program thinks, performs an action, looks at the result, and then plans the next step based on it. This loop continues until the task is completed or a limit is reached.

An agent swarm is a group of AI programs that work together on a task instead of a single program doing everything on its own. Each program takes on a sub-step, passes on its result, and sometimes checks the work of the others.

An agent workspace is a sealed-off digital working environment in which an AI program carries out tasks independently: with its own files, its own tools, and its own memory. It separates what the software is allowed to touch from the rest of the system.

Agent coordination describes how several independently operating AI programs divide a shared task among themselves and merge their intermediate results. It governs who does what, in what order the work proceeds, and how a single outcome is produced in the end.

Agentic commerce refers to purchases that a computer program carries out independently on behalf of a human: it searches for products, compares offers, and completes the purchase. The human only sets the goal and the boundaries, such as a budget.

An Agentic Development Environment is a program for writing software in which AI assistants carry out tasks independently across multiple steps. The human describes the goal, reviews the results, and approves them instead of typing every line themselves.

Agentic ransomware is extortion software that uses an AI program as an independent helper: it autonomously searches a foreign computer for valuable files, encrypts them, and demands money. Unlike older malware, the individual steps are not fixed in advance but are decided during the attack.

Agentic Search refers to a search process in which an AI program directs the research itself: it formulates its own search queries, reads the results, notices gaps, and keeps searching until it can compile an answer. Instead of a list of hits, you get a result with source citations.

Agentic-as-a-Service refers to ready-to-use, subscription-based AI programs that independently work through tasks in multiple steps, rather than merely producing text output. The provider runs them in its own data centers, and the customer pays per use or per completed task.

Agentic AI describes computer programs that don't just answer a task, but break it down into steps themselves and carry out those steps one after another. In doing so, they are allowed to use tools, such as searching the internet, opening files, or sending emails.

Agentic models are AI systems that don't just answer a task but independently break it down into steps and work through those steps using tools such as web search or program code. In other words, they act instead of merely responding.

The agentic loop is the basic pattern by which AI programs work autonomously: think, take an action, look at the result, continue. This cycle repeats itself until the task is done.

An agentic harness is the program that turns a language model into an assistant that operates on its own: it repeatedly calls the model, executes its suggestions, and feeds the results back in. Without this framework, a model could only output text but couldn't actually do anything.

An agentic operating system is a software layer that manages independently acting AI programs: it starts them, allocates computing time and permissions to them, and logs what they do. The term transfers the role of Windows or Android to a world in which not humans but AI programs carry out the tasks.

An agentic AI system doesn't just answer questions — it pursues a goal across multiple steps and independently uses tools such as search engines, calendars, or program code. It plans, acts, checks the result, and corrects itself until the task is done.

Aggregation means combining many individual values into a few overall values — for instance, turning a thousand measurements into a single average. In AI and data processing, it is a fundamental tool because it makes large amounts of data manageable while revealing patterns.

An aggregator is a service that collects content or offerings from many external sources in one place and displays them in an organized way. Google, Booking.com, or a news feed are typical examples: they produce almost nothing themselves, but control access to the providers.

AGI stands for an artificial intelligence that doesn't just master one task, but could solve any intellectual task at a human level. So far, no such system exists — the term describes a goal whose very definition even experts argue about.

The AGPL is a license for freely available program code that requires modifications to be disclosed again. Its special feature: this obligation also applies when the software is only offered as an online service and not distributed at all.

AGPL-3.0 is a license for freely available software. It requires that anyone who modifies the program and offers it over the internet must disclose the modified program source code.

The AI Act is the European Union's law for artificial intelligence. It sorts applications according to their risk and specifies which are prohibited and which remain permitted only under strict requirements.

The AI Action Plan is a government program that sets out how a state intends to promote and regulate artificial intelligence. The best-known one is the US government's plan from July 2025, which focuses above all on speed, data centers, and reduced bureaucracy.

AI Before Design refers to a way of working in which it is first clarified what an artificial intelligence can accomplish within a product, and only afterward is the user interface designed. The approach reverses the usual order, in which a product is fully designed first and AI is added later.

The AI Control Problem is the question of how humans can permanently keep very powerful computer systems under control. It concerns ensuring that a system does exactly what was intended – and that it can be corrected or shut down if necessary.

AI Debt refers to the hidden follow-up costs that arise when companies quickly implement AI systems and postpone the cleanup work until later. The term is modeled on debt: you save time today and pay it back later with interest.

AI Discoverability describes how well a company, brand, or piece of text is found, understood, and mentioned by chat assistants like ChatGPT in their answers. It sits alongside classic search engine visibility, because more and more people ask questions directly to a program instead of to a list of results.

The AI Flywheel describes a cycle in which an AI product gets better through its users: more users generate more data, which can be used to improve the product, which in turn attracts more users. The term originates from business and explains why leads held by AI companies often keep growing larger and larger.

An AI Gateway is a software intermediary through which all of a company's AI requests pass. It routes requests to the appropriate AI providers, tracks costs, and blocks unwanted content.

An AI Gigafactory is an extremely large data center built specifically for training very large AI systems. The term originates from EU policy and describes facilities with tens of thousands of specialized chips, their own power connection, and investments in the billions.

AI Governance refers to all the rules, responsibilities, and controls with which a company or a state steers the use of artificial intelligence. It is not about the technology itself, but about who decides, who is liable, and what is permitted.

AI Infidelity describes the case where someone in a relationship forms a close, often romantic or sexual bond with a computer program that mimics human conversation. The term is not a legal category, but a description of a new kind of conflict in partnerships.

AI Mode is an operating mode of Google Search in which you don't get a list of links, but a fully formulated answer from a program that generates text itself. Afterwards you can ask follow-up questions, just like in a chat.

AI Overviews are automatically generated summaries created by Google that appear at the top of many search queries and answer the question directly. They replace part of the classic results list and thereby change how many visitors still click through to websites.

AI Slop refers to texts, images, or videos mass-produced by software with no discernible value, intended primarily to generate clicks and ad revenue. The term is meant pejoratively and describes a volume problem: such content clogs up search engines, social networks, and online shops.

An AI crawler is a program that automatically fetches web pages and collects their content so companies can use it to train AI systems or answer user questions. Because this uses other people's text and images without compensation, media companies, authorities, and AI firms have been arguing about these bots for years.

An AI factory is a data center built almost exclusively for artificial intelligence: thousands of specialized chips work there around the clock on a single task. The term emphasizes that such facilities function like a factory that produces finished AI answers and AI models out of electricity and data.

AI-First describes a company's stance of thinking about artificial intelligence first for every product and every workflow, rather than bolting it on afterward. The term is not a technical procedure but a strategic decision – and in corporate communications it is often also a marketing promise.

AI high-performance chips are especially fast computing components tailored to the mathematics behind artificial intelligence. They perform extremely large numbers of simple calculations simultaneously, making them the most expensive and scarcest ingredient of modern AI systems.

AI-Native describes products or companies that were built around artificial intelligence from the very beginning, rather than adding it on later. The term distinguishes such offerings from classic software that has an AI feature bolted on afterward.

An AI PC is a computer with an additional specialized chip that processes AI tasks directly on the device. This allows functions like translation, image editing, or voice assistants to run without a detour through data centers on the internet.

AI persona badges are visible markers on profiles or chat windows indicating that a character is not operated by a human but by software. They are meant to prevent users from mistaking an artificial conversation partner for a real person.

AI referrals are visitors who arrive at a website via an AI chatbot because it included a link to the site in its answer. They count as a new source of website visits alongside the classic search engine – though in very small numbers so far.

AIME is a demanding American mathematics competition for high school students, whose problems are now also used to test AI systems. Because every answer is an integer between 0 and 999, it can be checked automatically and unambiguously whether a model calculated correctly.

Active parameters are the portion of an AI model's adjustable numerical values that actually participate in the computation for a single request. For some models, this is only a fraction of the total, which makes operation significantly cheaper.

Active parameters are the part of an AI model that actually does the computing for a single request. In models with expert partitioning, this number is much smaller than the total parameter count — and it determines how fast and how expensive the model's responses are.

Activated parameters are the portion of an AI model's learned settings that actually factor into computation for a single request. This number reveals more about speed and cost than the often-cited total size of the model.

The activated parameter count indicates how many of an AI model's set values actually factor into a single request. It can be significantly smaller than the total count and helps determine how fast and how costly a model is to operate.

An activation is the numerical value that a single computational unit in an artificial neural network is currently outputting. These intermediate results travel layer by layer through the network and ultimately determine what response comes out.

An activation is the numerical value that a single computational unit in an artificial neural network is currently outputting. All activations together form the intermediate state through which the network gradually transforms an input into a response.

Activation energy is the minimum energy a chemical reaction needs in order to get started at all. It explains why wood doesn't spontaneously burn at room temperature, even though burning overall releases energy – and why a single match is enough to make it happen anyway.

The activation space is the numerical space in which an AI system deposits its intermediate results while processing an input. Similar content ends up in similar locations there, which is why researchers study this space to understand what a model actually thinks internally.

The acceptance rate measures what proportion of a computer program's suggestions are actually adopted by users. It is especially, for coding assistants like GitHub Copilot, the most important metric for whether a system is genuinely useful in everyday work.

Alexa is Amazon's voice assistant: a program that understands spoken commands and responds to them, for example by playing music or switching on lamps. It became well known through the speakers of the Echo line, and since 2025 a newer AI technology has additionally been working behind it.

Algebraic number theory studies integers and their divisibility by deliberately extending the number range. Among other things, it provides the mathematical foundation for modern encryption.

Algorithmic Bias refers to a systematic skew in automatically made decisions: a computer program regularly treats certain groups of people worse than others. The cause is usually the data the program learned from, not malicious intent on the part of the developers.

Algorithmic Monoculture describes a situation in which many companies or authorities use the same automated decision rule. Because everyone then judges by the same standards, the same people run into the same rejection everywhere — and a flaw in the rule takes effect immediately and across the board.

Alignment refers to the task of aligning an AI system so that it actually does what humans want it to do — not just whatever happens to be written in its task specification. The term covers both technical methods and the question of which values a system should be aligned to in the first place.

Alignment pretraining means that an AI system is already shaped toward desired behavior during the first major learning process — not only through later refinements. Desired and undesired examples are treated differently from the very beginning.

An allow list is a list that precisely defines who or what is permitted – everything else is automatically rejected. It is the counterpart to the block list, which conversely only prohibits individual items and lets everything else through.

The alpha phase is the early stage in the development of a program in which it already runs but is still unfinished and prone to errors. Testing is usually done only internally or by a small selected group.

AlphaGenome is an AI model developed by Google subsidiary DeepMind that predicts what effect individual sites in the human genome have. It is intended to help researchers understand why certain deviations in the genome can trigger diseases.

An AlphaGo moment refers to the point in time when a computer program suddenly solves a task better than the best human – and in doing so finds solutions that experts would never have come up with. The expression goes back to the board game program AlphaGo, which beat world-class player Lee Sedol in 2016.

Age verification refers to all technical methods by which an online service checks how old a user actually is. It ranges from a simple click on "I am over 18" to matching against an ID card or an estimate via facial analysis.

An always-on agent is a computer program with artificial intelligence that runs permanently in the background and decides for itself when to become active. Unlike a chatbot, it does not wait for a question, but instead observes events and acts on its own initiative.

Always-on sensing refers to sensors in devices like phones, watches, or speakers that measure continuously instead of only after being switched on. They use extremely little power and wake up the rest of the device only when something actually happens.

An ambient agent is an AI program that runs in the background and notices on its own when it needs to do something – instead of waiting for input. It reacts to events like new emails, calendar entries, or code changes and only speaks up when it becomes important.

Ambient computing refers to the idea that technology runs invisibly in the environment and reacts in the background, rather than being consciously operated. Instead of picking up a device, one simply speaks, walks, or acts – and the technology draws its conclusions from that.

Ambient Intelligence describes spaces equipped with sensors and computing technology that adapt to the people within them without being asked. The technology remains as invisible as possible: you don't operate a device, you simply live in the environment.

Amdahl's Law describes how much faster a task becomes when it is distributed across multiple processing units. The core idea: the portion of work that must inherently run step by step limits the gain — no matter how many processors are added.

An American Depositary Receipt is a security traded on a US stock exchange that represents shares of a company based outside the United States. It allows American investors to invest in companies like SAP or Alibaba without having to buy on a foreign exchange.

An amicus brief is a written submission filed with a court by someone who is not themselves a party to the case. In AI-related proceedings, companies, associations, and researchers use this instrument to alert judges to the technical consequences of a ruling.

Android Auto is a service by Google that displays content from a phone on the screen inside a car, presenting it in a heavily simplified form. The car's screen thus becomes a control surface for navigation, music, and messages, while the phone itself can stay in your pocket.

The Android Open Source Project, or AOSP for short, is the freely available base version of the mobile operating system Android. Google publishes the blueprint of the system there, allowing manufacturers and developers to use and modify it free of charge.

An attack vector is the concrete route by which someone gains unauthorized access to a computer system – such as a fake email, a stolen password, or a flaw in software. Professionals catalog and assess these routes in order to close off the most dangerous ones first.

Annotation means adding extra information to data so that a machine can learn from it. On a photo, for example, a human marks where a car can be seen — providing the software with the correct answer to practice on.

Annual Recurring Revenue, or ARR for short, is the revenue a company reliably expects from ongoing subscriptions per year. The metric shows how much money comes in dependably if no customer cancels and no new one is added.

Annual Recurring Revenue, or ARR for short, is the revenue a company expects from ongoing subscriptions over the course of a year. The metric is considered the most important yardstick for how fast software and AI companies are growing.

Annualized recurring revenue, or ARR for short, projects a company's regular payments up to a full year. It is the most important metric for companies that sell software by subscription – and in AI news the number by which growth is measured.

An Answer Engine is a search system that delivers a directly formulated answer to a question instead of just displaying a list of links. Well-known examples are Perplexity, ChatGPT with web search, and the AI Overviews in Google Search.

Answer Engine Optimization (AEO) refers to all the measures companies take to ensure that programs like ChatGPT mention their content in ready-made answers. It is the evolution of classic search engine optimization for a world in which many users no longer click on a list of results.

Anthropomorphic AI refers to computer programs that are deliberately designed to appear human – through language, voice, name, or appearance. The term also describes people's tendency to attribute feelings and intentions to such programs that they do not actually have.

Antibiotic resistance means that bacteria are no longer killed by a drug that used to work against them. It arises through random changes in the bacteria's genetic material and spreads when antibiotics are used too often or incorrectly.

The application layer is the topmost level of a layered computer system – the place where the programs people work with directly reside. In the AI industry, the term refers to the companies that build finished products for end customers on top of third-party foundation models.

Apache 2.0 is a widely used usage agreement for software, set in advance by the author. It allows anyone to use, modify, and redistribute the program code for free, including commercially, as long as the original author is credited.

The Apache 2.0 License is a standard contract that developers use to release their software for free use – including commercial purposes. It is especially widespread in the AI world because many freely available language models are published under it.

An API is a defined interface through which one program can use the data or functions of another program. It specifies exactly which requests are allowed and in what form the response comes back.

An API call is a request that one program sends over the internet to another program in order to retrieve a service from it. For AI services, the API call is the standard way to use a model — and the unit by which usage is billed.

API documentation is the instruction manual for a programming interface: it describes which requests a program is allowed to send to a foreign service and which responses it gets in return. Without it, every developer would have to find out by trial and error how a service works.

An API endpoint is a fixed internet address that a program sends a request to in order to get a response from another program. Anyone building an AI model like GPT or Claude into their own app today is talking to exactly this kind of endpoint.

An API extension adds new functions to an existing programming interface, through which external programs can use a service. It is the usual way providers like OpenAI or Google open up their AI services to developers without breaking existing applications.

An API function is a single, clearly defined service that one program offers to another program — for example, "translate this text." You call it over the internet or within a computer, send along data, and get back a result.

An API gateway is a central entry point through which all requests to the building blocks of an online service pass. It checks, distributes, throttles, and logs these requests so that the underlying programs remain relieved and protected.

An API key is a long string of characters that a program uses to identify itself to an external service. It lets the provider recognize who is making the request, what that person is allowed to do, and who to bill for usage.

The API economy describes a business model in which companies rent out their software functions to other firms via technical interfaces instead of selling them to end customers. Payment is usually made per call, which is why large parts of today's apps consist of purchased building blocks from other providers.

An API interface is a defined way for one program to use data or functions from another program. It is the foundation that allows apps, websites, and AI services to work together.

The Apollo Program was the American space program of the 1960s and early 1970s that first landed humans on the Moon. Between 1969 and 1972, twelve astronauts walked on the lunar surface.

Apoptosis is the orderly self-degradation of a cell according to a fixed internal program. The body uses it to remove old, superfluous, or damaged cells without harming the surrounding tissue.

Apple Silicon is the name for the computer chips that Apple has designed itself since 2020 and builds into Macs, iPads, and iPhones. The chips combine the main processor, graphics unit, and memory connection on a single component and are considered especially energy-efficient.

An Applicant Tracking System is software that companies use to collect, sort, and guide incoming applications through the selection process. It automatically reads resumes, matches them to job postings, and filters or ranks candidates according to predefined criteria.

The Application Layer is the topmost level of a technical system – the place where humans actually use it. In the AI world, the term refers to products that build on other companies' models instead of training their own.

An Application Programming Interface, or API for short, is a defined interface through which one program can use the services of another program. It specifies exactly which requests are allowed and which responses come back.

Approximate Nearest Neighbour is a search method that very quickly finds entries similar to a queried entry within huge amounts of data. It deliberately forgoes the guarantee of finding the absolute best match, thereby becoming hundreds of times faster.

The Apps SDK is a collection of tools published by OpenAI that lets developers integrate their own applications directly into ChatGPT. Instead of opening a separate website, users invoke the app within the chat window, where it can also display its own user interface.

AR-Diffusion refers to AI models that combine two generation principles: they produce content section by section in order, but refine each section over multiple revision steps. The goal is to create text, images, and videos faster without losing the overall thread.

ARC-AGI is a test for computer programs consisting of colorful grid puzzles. It checks whether a program can recognize a completely new rule from just a few examples, rather than simply recalling what it has learned.

The Arena Ranking is a leaderboard of AI chatbots created through direct comparisons: users ask a question, see two answers without names attached, and choose the better one. Out of millions of such votes, a score is calculated, similar to chess ratings.

ARM is a blueprint for processors, that is, for the computing chips in computers and phones. The British company Arm designs this blueprint and sells licenses for it instead of manufacturing chips itself — almost every smartphone in the world runs on it.

ARM64 is the 64-bit instruction set of the ARM processor family, i.e. the language in which programs talk to the chip. Almost all smartphones use it, and for some years now it has also been conquering notebooks and data centers.

Artificial General Intelligence (AGI) refers to a computer system that could solve any intellectual task a human can handle – not just individual specialized tasks. Whether such a thing is possible, and when, remains open; the term is therefore above all a goal and a point of contention, not an existing product.

Article 40 is the section of the European Digital Services Act that obliges large online platforms to disclose data to vetted researchers. It is meant to enable independent investigations into the risks platforms create for society.

arXiv is a free online platform where researchers publish their papers before a journal has reviewed them. In AI research, most important papers appear there first.

Ascend is a series of computer chips made by the Chinese conglomerate Huawei, built specifically for the computational work of artificial intelligence. It is considered the most important Chinese alternative to the chips of US market leader Nvidia.

Ascend is a series of specialized chips from the Chinese conglomerate Huawei, built for computing artificial intelligence. They are considered China's most important homegrown answer to the dominant AI chips made by the US manufacturer Nvidia.

ASI stands for "Artificial Superintelligence," meaning an artificial intelligence that clearly surpasses humans in virtually all cognitive tasks. To this day, nothing like this exists – ASI is a future scenario that researchers and companies argue about.

An ASIC is a computer chip that is permanently built for a single task and can't do anything else. Because it contains nothing superfluous, it performs that one task faster and more efficiently than a regular processor.

An asset pipeline is the automated chain of work steps that turns raw material such as images, 3D models, music, or text into exactly the format that a game, an app, or a website can ultimately load. It ensures that huge original files become fast, checked, and consistently named end products.

The AT Protocol is an open rulebook for social networks in which user account, content, and app are separated from one another. It became known through Bluesky, the first major service built on it.

Atomic Answers are very short, self-contained answers that directly answer a single question without requiring further reading or clicking. They have become the preferred format because search engines and chatbots specifically extract and display exactly these kinds of text snippets from websites.

Atomic Design is a method for assembling the interface of a website or app from small, reusable building blocks. The building blocks are organized into five stages, ranging from a single button to a finished page.

Attention is a computational method that allows an AI program to decide, for each word in a text, which other words are important to it. It is the core of modern language models like ChatGPT and the reason these programs are able to understand longer contexts.

Attribution answers the question of which advertising contact triggered a purchase. It distributes the credit for a sale across the various ads, emails, and search results that a customer saw beforehand.

The term attribution crisis refers to the growing difficulty of reliably determining who or what caused a piece of content or an attack. Because AI systems replicate texts, images, and voices with deceptive accuracy, the usual evidence of authorship loses its probative value.

An audio summary is a spoken condensed version of a longer text, automatically generated by a computer program. From a document, a study, or several news articles, it creates an audio track lasting a few minutes, often in the style of a conversation between two voices.

An audit log is an ongoing record in which a computer system keeps track of who did what and when. It serves to allow processes to be reconstructed and reviewed after the fact.

Augmented Reality means that a device overlays additional images or text onto the real environment while you are looking at it. You still see the real space, supplemented with content that a computer adds in.

Augmented reality glasses are eyewear that overlay additional images and text into the field of view while the real environment remains visible. They combine transparent lenses, cameras, and sensors with a small computer that understands where the wearer is currently located.

Alignment refers to the task of building an AI system so that it truly does what humans want it to do – and not something that merely looks like it on the surface. This is one of the central topics of debate in AI research.

An auth token is a string that a program presents as proof that it has previously logged in successfully. Instead of sending a password or key again with every request, this time-limited piece of text is included instead.

Auto-dubbing refers to the automatic translation of spoken language in videos, where software generates the new audio track. The voice often sounds like that of the original speaker.

Auto-labeling means that a computer program labels training data itself instead of humans doing it by hand. The results are faster and cheaper, but less reliable than human work.

Auto layout is a feature in design and app programs where elements adjust their own position and size instead of being fixed to set coordinates. When a text or window size changes, the layout automatically rearranges itself.

Autograd is a technique that lets a computer program automatically calculate how much a result changes when you turn a specific knob. In AI, this is the foundation that allows a model to learn from its mistakes.

An automation benchmark is a standardized test designed to measure how well a computer program independently accomplishes real work tasks. Instead of answering individual questions, the program must complete an entire task from start to finish — such as evaluating a spreadsheet or completing a booking in a browser.

Autonomous agents are AI programs that don't just answer a goal but work through it in their own steps. They plan, use tools such as search engines or calendars, and check their own intermediate results.

Autonomous driving means that a vehicle perceives its surroundings itself and steers, accelerates, and brakes without human intervention. How much the human still needs to do is described by an internationally used scale ranging from level 0 to level 5.

An Autonomous Knowledge Platform is a software system that gathers a company's scattered knowledge, organizes it, and turns it into answers on demand. Unlike a classic search tool, it doesn't just wait for search terms but keeps working independently: it sorts new documents, checks for contradictions, and initiates tasks on its own.

An autoregressive model generates its output piece by piece, always using what has already been generated as the basis for the next step. Almost all of today's chatbots work this way: they write text word by word instead of completing an answer in one go.

An Availability Zone is a self-contained location of a cloud provider, consisting of one or more data centers, with its own power, cooling, and network connectivity. Multiple such zones within a region make it possible to distribute services so that they keep running even if one location fails.

AWS Trainium is a specialized chip that Amazon developed itself to train AI systems. It is Amazon's attempt to become less dependent on the expensive chips of market leader Nvidia when it comes to computing for artificial intelligence.

AX Design means designing websites, programs, and services so that not only humans can use them well, but also autonomously operating computer programs. The term is a response to the fact that increasingly software rather than humans is clicking through pages, filling out forms, and triggering orders.

The axion is an as-yet undetected, extremely light elementary particle that physicists proposed to solve a puzzle in the structure of atomic nuclei. It is also considered one of the most promising candidates for dark matter, the invisible substance that makes up most of the mass in the universe.

An AI agent is a computer program that independently breaks a goal down into individual steps and also carries out those steps. So it doesn't just respond—it acts: it searches the web, writes files, or sends messages.

An AI agent is a computer program that independently breaks down a goal into multiple steps and also carries out those steps. So it doesn't just respond — it uses tools like search engines, calendars, or other programs while doing so.

An AI agent is a computer program that independently breaks down a goal into multiple steps and also carries out those steps. Unlike a pure chatbot, it doesn't just answer — it uses tools such as search engines, calendars, or program code.

An AI accelerator is a computer chip specifically built for the computational tasks of artificial intelligence. It performs the ever-recurring number calculations of such programs much faster and with much lower power consumption than a regular main processor.

An AI chip is a computing component specifically built for the tasks of artificial intelligence. It performs vast numbers of simple numerical operations simultaneously, making it significantly faster and more energy-efficient for AI programs than a regular computer processor.

An AI coding agent is a program that handles a programming task in multiple independent steps: it reads the existing code, modifies files, tests the result, and corrects itself when errors occur. Unlike a mere suggestion service, it works over extended periods largely without needing to check back.

An AI detector is a program that estimates whether a text or an image was created by a human or by a machine. It only ever provides a probability, not proof – and it is wrong more often than many users assume.

"AI-native" describes products or companies that were built around artificial intelligence from the very start — rather than bolting it onto a finished product later. The term is not a technical standard, but a description of how central the technology is to the way something is built.

"AI-native" describes products or companies that were built around artificial intelligence from the very start. Take away the AI, and no functioning product remains—unlike software that merely receives AI bolted on afterward as an add-on feature.

The AI stack is the totality of layers needed for an AI product to work: from the chips through the data centers and the models to the app you use yourself. The term helps to structure the market, because different companies make money in each layer.

An allowlist is an enumeration of what is explicitly permitted – everything else is automatically considered forbidden. In technology, it controls, for example, which programs may start or which websites an AI is allowed to read.

An application programming interface is a defined way for one computer program to use the functions of another program. It specifies exactly which requests are allowed and in what form the response arrives.

An asterisk field is an input field on a website or in an app that does not display the characters typed but replaces them with dots or asterisks. Its main purpose is to protect passwords from being read by bystanders.

Access control governs who is allowed to view which data and perform which actions in a computer system. It consists of two steps: verifying who someone is, and deciding what that person is allowed to do.

B2B stands for "Business-to-Business" and refers to deals in which one company sells to another company rather than to private individuals. In the tech industry, B2B is the most common business model because corporate customers pay more reliably than individual users.

The B300 is a specialized chip made by Nvidia, built for computing artificial intelligence. It is the improved version of the predecessor chip B200 and is among the most expensive and sought-after components in modern data centers.

A backdoor is a hidden malfunction that someone deliberately teaches a learning computer program: normally it works correctly, but upon a secret trigger it delivers the result the attacker wants. Because the program behaves inconspicuously in all normal tests, such backdoors are extremely hard to detect.

The backend is the part of an app or website that runs invisibly on other people's computers, where it stores, checks, and processes data. What's visible is only the frontend – the interface on your own screen.

A background model is a computational model that describes what the normal, uninteresting substrate looks like — that is, everything that is not currently being searched for. Only through this benchmark can one recognize what is truly conspicuous about an observation.

A background agent is an AI program that independently works through a task in the background while the user does something else. Instead of responding turn by turn in a chat, it works on a task over minutes or hours and reports back with the result.

Backtracking is a problem-solving strategy in computer science: you try a path, and as soon as it turns out to be wrong, you step back and try the next possibility. This way, a huge number of possible solutions can be searched systematically without having to check each one completely.

A base model is the raw version of an AI language model: it has read enormous amounts of text and can continue text, but has not yet been trained to follow instructions or respond politely. From this raw version, additional training steps later produce the chat programs familiar from everyday life.

A baseline is a simple benchmark against which one measures whether a new method is truly better. Without this reference point, a success figure like "92 percent correct" says nothing at all.

Bash is a program used to control a computer by typing commands instead of using windows and icons. It is the standard control interface on many Linux servers and is therefore a basic tool in software development.

BASIC is a programming language developed in 1964 that was deliberately kept so simple that beginners could learn it in just a few hours. For decades it was the standard language on home computers and made programming accessible to millions of people for the first time.

Batch processing means that a computer collects many tasks and works through them together in one go, instead of handling each one immediately as it arrives. This is significantly more efficient and cheaper, but you have to wait for the result.

Bauhaus Emigration refers to the emigration of the teachers and students of the Bauhaus art school after its closure by the National Socialists in 1933. Those who were driven out spread their ideas of simple, functional design in the USA, Israel and many other countries.

Bayesian prioritization is a method in which a ranking of tasks or options is continuously adjusted in light of new observations. It is based on Bayes' rule, which combines existing expectations with fresh data into an updated assessment.

The Bayesian approach is a way of thinking in which an initial guess is corrected step by step through new observations. Instead of a fixed yes-or-no answer, it delivers a probability that changes with each new piece of evidence.

A BDR is a sales employee who reaches out to prospective new customers and figures out whether a sales conversation is even worthwhile. They don't close deals themselves, but instead hand promising contacts over to colleagues — and it's precisely this work that is increasingly being supported by AI software.

The beam width determines how many possible sentence beginnings a language program pursues simultaneously before settling on an output. A larger value often delivers better results, but costs more computing time.

Behind-the-Meter refers to power generation located directly at the consumer's site, with electricity that does not pass through the public grid's meter. Data centers use this principle to place power plants directly next to their server halls, bypassing the long detour through the electrical grid.

A benchmark is a standardized test used to measure and compare the performance of computer programs or AI systems. All candidates work on the same tasks, and the result is usually given as a score or percentage.

A benchmark harness is a program that automatically runs an AI model through a fixed collection of test tasks and tallies the results. It ensures that different models are tested under exactly the same conditions and that their scores are comparable.

Benchmark contamination refers to a situation where the tasks of a test for AI systems were already present in the system's training material. The tested program can then partially reproduce the solutions from memory, and the test result says less about its actual capability.

Benchmarking means: different computer programs are made to solve the same test tasks, and the results are compared. In the AI industry, such comparisons decide which program is considered the best.

Best-of-breed integration means that a company selects the best individual program for each task and then connects these programs with one another. The opposite approach is a single all-in-one system from one provider that covers everything but is often weaker in individual areas.

Beta refers to a version of a program that is already usable but has not yet been fully tested. The maker deliberately releases it to users to find bugs before the final version is published.

A beta header is a short additional piece of information that a program sends when calling an online service in order to unlock a not-yet-finished test feature. Without this information, the service behaves as usual; with it, the caller deliberately receives an experimental version.

The beta phase is the stage in the development of a program in which a nearly finished version is released to real users to find bugs in everyday use. It lies between internal testing and the official release.

A beta version is a not-yet-finished version of a program that is already distributed to users so that bugs can be found. It works largely as intended but is explicitly marked as unfinished and can crash or still change.

BF16 is a format in which computers store decimal numbers using 16 digits instead of 32. It has become the standard for AI models because it halves memory usage and computation time without making training unstable.

Bias in AI models refers to systematic distortions: a computer program consistently treats certain groups or cases differently, without any factual reason for doing so. The cause usually lies in the data from which the program learned.

A bidding agent is a computer program that automatically places bids in auctions on behalf of a buyer or seller. It decides in fractions of a second whether and how much to bid – based on rules or on learned predictions.

A bidirectional transformer is a language model that looks at every word of a sentence in context with all other words at once – including the ones that come after it. This makes it strong at understanding text, but unsuitable for writing new sentences.

Bidirectionality means that a language model looks at a text simultaneously from left and right, interpreting each word based on its entire surrounding context. This makes such models strong at understanding texts, but unsuitable for freely continuing writing.

A Bill of Materials is a complete list of all the components that make up a product. In technology today, it is used above all for software, so that manufacturers and customers know which third-party components are contained in a program.

Binary noise is a random disturbance that knows only two possible values: on or off, 0 or 1. In technology it arises as an error during data transmission; in AI it is deliberately generated to make models more robust.

The binary system is a way of writing numbers that uses only the digits 0 and 1. Computers compute and store data in this form because two states can be reliably distinguished at the technical level.

Biometric authentication verifies a person's identity using physical characteristics such as fingerprint, face, or voice. Instead of entering a password, you show yourself — which is convenient, but carries its own risks.

Biometric data are measured values of physical or behavioral characteristics that can be used to identify a specific person — such as fingerprint, face, iris, or voice. Because, unlike a password, they cannot be changed, they are legally considered especially worthy of protection.

Biometric categorisation refers to software that sorts people into groups based on physical characteristics such as face or voice, for instance by age or gender. It does not recognise who someone is, but rather claims what someone is – and is therefore partly banned in the EU.

Biosecurity encompasses all rules and precautions intended to prevent pathogens from causing harm – whether through accidents or deliberate intent. In the AI debate, the main question is whether language models could help laypeople build dangerous pathogens.

Biotech is short for biotechnology: the use of living organisms and their building blocks to produce medicines, food, or materials. On the stock market, biotech refers to an industry with high research costs and very large price swings.

The Bitter Lesson is a well-known thesis by AI researcher Richard Sutton from 2019. It states: methods that simply exploit massive computing power and large amounts of data almost always beat, in the long run, methods into which humans have hand-built their expert knowledge.

Black Hat refers to someone who breaks into other people's computer systems to enrich themselves or cause damage. The term distinguishes these criminal attackers from professionals who use the same techniques with permission to secure systems.

The blackboard architecture is a software design in which several specialized program parts work on a shared data store. Each part reads what is already there, adds its own intermediate results, and thereby advances the solution step by step.

A blacklist is a list of things that are explicitly forbidden or blocked – such as web addresses, senders, or words. Anything not on the list remains allowed; this is precisely where the strength and the weakness of this principle lie.

Blackwell is the name of a chip generation from the company NVIDIA, built for training and running large AI systems. Starting in 2024, it succeeded the previous generation, Hopper, and is considered the most important gauge for the boom around artificial intelligence.

The BLEU score is a metric that measures how similar a machine-generated translation is to a human-made reference translation. It has been used since 2002 to automatically compare translation programs without human evaluation.

A blind box is a sealed package where the buyer doesn't know which figure or item is inside before purchase. The principle is the foundation of a billion-dollar collectible toy market and resembles, in its effect, random purchases in video games.

A blind test is a comparison in which the test subject does not know which of the items being tested they currently have in front of them. This way, expectations and brand names do not influence the judgment — in the AI industry, this method is used to evaluate language models against each other.

A blind test is a comparison of AI responses in which the person evaluating does not know which program wrote which answer. This is meant to prevent a well-known name from influencing the assessment.

Bloatware refers to programs that are already installed on a phone or computer at the time of purchase, even though the buyer never asked for them. They take up storage space, often run in the background, and frequently cannot be fully removed.

A Bloom filter is a very space-efficient method for checking whether an element has already been stored. It answers either "definitely not present" or "probably present" – so an error is only possible in one direction.

Blueprints are the visual programming system of the game development software Unreal Engine: instead of typing lines of code, you connect boxes with lines to form a flowchart. This lets you build game rules, characters, and menus without needing to know a programming language.

Boilerplate code is program text that always looks the same and has to be written anew in many projects, even though it doesn't accomplish anything unique. It's necessary so that a program can start at all or fit together with other software – no original thought went into it.

The book-to-bill ratio compares a company's newly received orders with what it has shipped and billed in the same period. A value above 1 means: more work is coming in than is being worked off — an early signal of growing demand.

Bookings are the total value of all contracts a company has newly signed in a given period — regardless of whether the money has actually been received yet. The metric shows how well sales are currently performing, and among software and AI companies it is considered a leading indicator of future revenue.

Bootstrap refers to the process by which a system pulls itself up from a tiny startup program until it becomes fully operational. The best-known case is starting a computer, but the principle also appears in programming languages, companies, and AI models.

A bootstrapping compiler is a translation program for a programming language that is itself written in that very same language. For such a compiler to come into existence at all, an initial, simple version built in another language is required.

A bot is a program that automatically carries out a task on the internet without a human at the screen. Bots crawl websites, answer questions in chat, buy tickets, or spread false reports – useful and harmful at the same time.

A bot is a program that performs a task independently and repeatedly, without a human triggering every step. Modern bots powered by AI in the background are often called AI agents: they plan intermediate steps themselves and, in doing so, use other programs such as search engines or calendars.

Bot traffic refers to all visits to a website that come not from humans but from automatically running programs. On many websites, these machine-generated visits now account for roughly half of all page views.

A bounding box is a rectangle that encloses an object in an image, indicating where it is located. It is the most common way to mark up image content for computer programs designed to detect objects.

A brain-computer interface is a connection between the brain and a device that translates neural activity directly into control commands. This allows computers, prosthetics, or wheelchairs to be operated without any muscle movement.

A branch is a diverging line of work in a project where you can try out changes without endangering the existing version. The term comes from software development, where it is the standard way for multiple people to work on the same program at the same time.

A branch is a separated working state of a software project in which changes can be tried out without damaging the previous version. If the work proves successful, the branch is merged back into the main line.

Breach response refers to the organized process a company follows after unauthorized parties have infiltrated its computer systems or stolen data. This includes stopping the attack, investigating the incident, legally mandated reporting, and notifying those affected.

Break-even is the point at which a company's revenue exactly covers its costs – meaning neither profit nor loss is made. In the tech and AI industry, the term is important because many companies spend more than they earn for years.

A break-up fee is a payment agreed upon in advance that becomes due if a planned company acquisition or merger ultimately does not go through. It is meant to compensate the side that invested a great deal of time and money into the failed deal.

A breaking change is a change to a piece of software after which previously working third-party programs no longer run. Such changes are usually announced in advance and bundled into major version jumps, because all users have to make adjustments.

Bring Your Own Key (BYOK) means: A customer uses an online service, but manages the secret code for encrypting and decrypting their data themselves. The provider stores the data but cannot read it without this code.

Broken Object Level Authorization is a security vulnerability in which a program checks who someone is, but not whether that person is actually allowed to see the requested data. Attackers then simply change a number in the request and receive other people's records.

The broken windows theory states that visible small deficiencies lead to bigger problems. Originally rooted in criminology, it is now also used in software development and AI systems as an image for creeping quality decline.

BrowseComp is a test by OpenAI that checks how well an AI system can track down hard-to-find information on the internet. The questions are constructed so that the answer is short and unambiguous, but finding it costs many search steps.

A browser extension is a small add-on program that you install into a browser like Chrome or Firefox and that changes its behavior. Extensions can, for example, block ads, manage passwords, or place an AI assistant over every page you visit.

A browser extension is a small add-on program that you install into a browsing program like Chrome or Firefox, which changes its behavior. It can hide ads, manage passwords, or bring AI features directly into any webpage — but in return often gets far-reaching access to everything you do online.

Brute-force analysis solves a problem by simply trying out all possibilities one after another, instead of thinking cleverly. It always succeeds if given enough time – except that this time grows into the unaffordable for large tasks.

Brute-force search is a solution method that simply tries out all conceivable possibilities one after another until one fits. It is always correct, but often prohibitively slow — and that's precisely why it also protects passwords and encryption.

A book gain is a profit that exists only on paper: the value of an asset has risen, but nothing has been sold yet. Only upon sale does it become real money — until then, the gain can just as easily disappear again.

A bug bounty program is a public offer made by a company: whoever finds a flaw in its software and reports it confidentially gets paid for it. This way, the company deliberately seeks outside help instead of waiting for criminals to discover the vulnerability first.

The Build-Fix-Validate loop is a recurring workflow in which something is built, then searched for errors and improved, and finally checked — with the result becoming the starting point for the next round. In software development and AI systems, it is the fundamental pattern by which programs and models get better step by step.

A build layer is a single, immutable intermediate stage in the assembly of a software package: each build step lays down a new layer on top of the previous one. Because unchanged layers can be reused, this shortens build times and saves storage space.

The bull case is the optimistic scenario for a stock or company: the story of how things go if many things go well. It is not a forecast but a calculation based on deliberately favorable assumptions — and is usually considered alongside a pessimistic counter-scenario.

Bun is a program that lets software written in the programming language JavaScript run on a computer. It is regarded as a fast alternative to Node.js and bundles several tools that developers would otherwise have to install separately.

The burn rate indicates how much more money a company spends per month than it earns. It shows how quickly a young company depletes its reserves — and how many months it can survive without new funding.

Business Email Compromise is a fraud scheme in which criminals pose via email as a superior or business partner in order to trigger money transfers. The attack does not rely on malware, but on deceiving employees.

Business Process Outsourcing means that a company has entire workflows handled by an outside firm, such as the customer hotline or bookkeeping. This industry is considered especially strongly affected by AI, because many of the outsourced tasks are routine.

A buyout firm raises money from wealthy investors and large institutional backers and uses it to buy entire companies. It holds onto the companies for a few years, tries to increase their value, and then sells them on, ideally for a higher price.

BYOD stands for "Bring Your Own Device" and refers to the rule that allows employees to use their private devices, such as phones or laptops, for work. This saves companies devices and costs, but raises questions about data protection and security.

BYOK means that a customer manages the secret code for encrypting and decrypting their data themselves, rather than leaving it to the provider. Anyone who disables the key makes the stored data unreadable for the provider.

Bytecode is an intermediate form of program code: it sits between the readable text that humans write and the instructions that a processor executes directly. A helper program on the target device reads the bytecode and converts it into real machine instructions there.

Behavioral biometrics recognizes people by how they do things: how they type, swipe, hold their phone, or move their mouse. Banks and online services use this in the background to detect fraud without requiring users to enter anything.

Behavioral drift describes the phenomenon where a computer program that learns from data gives different answers over time than it did before – even though the question stays the same. Causes include updates to the program or a world that develops differently than the examples it learned from.

The C-Band is a range of radio frequencies between roughly 3.4 and 4.2 gigahertz. It is considered especially valuable for fast mobile networks because it can carry large amounts of data while still covering long distances.

C-End refers to the business with private customers: products and services that individual people use themselves and usually also pay for themselves. The term comes from Chinese tech usage and frequently appears in reports about companies like Alibaba, ByteDance, or Tencent.

C-level refers to the top leadership tier of a company – the handful of people with titles like Chief Executive Officer or Chief Financial Officer. They make the decisions about strategy, money, and personnel that apply to the entire company.

C2PA is a technical standard that stores the origin of images, videos, and audio files within the file itself in a tamper-proof way. This makes it possible to trace which device or program created a recording and who edited it afterward.

CAC stands for Customer Acquisition Cost, meaning the average amount a company must spend to acquire a single new paying customer. The metric shows whether advertising and sales efforts pay off — especially when compared to what a customer generates over time.

The cache hit rate indicates how often a computer already has a requested piece of information in its cache and doesn't need to fetch it anew. The higher this share, the faster and cheaper a system operates – from websites to AI chatbots.

A cache hit means that an AI language model has already processed part of the input before and can reuse the stored intermediate result. This saves computing time and, with many providers, significantly reduces the price per request.

A cache miss occurs when a computer looks for data in its fast intermediate storage but doesn't find it there. It then has to fetch the data from slower memory, which noticeably delays the process.

Cached Input refers to the part of a request to an AI system that the provider reuses from an intermediate cache because it appeared in exactly the same form in an earlier request. Because this saves computational effort, such reused text segments cost only a fraction of the normal price with many providers.

CAD stands for "Computer-Aided Design," meaning the design of components and buildings on a computer instead of on a drafting board. The resulting models are precisely dimensioned and can be used directly for simulations, manufacturing machines, and 3D printing.

A call-to-action is a targeted prompt to visitors of a website or readers of an advertisement to do something specific right now – such as clicking a button or signing up. It is the element by which it can be measured whether a page achieves its goal.

Cambricon is a Chinese company that develops specialized computer chips for artificial intelligence. It is considered the most important domestic competitor to the US corporation Nvidia in the Chinese market.

camelCase is a naming convention in which multiple words are joined together without spaces, with each new word starting with a capital letter. It is used primarily in programming languages, since names there cannot contain spaces.

A Canaries Dashboard is an overview page where a team can see the results of many automated continuous tests of a computer system at a glance. If one of these tests fails, this becomes immediately visible there – often before real users notice the error.

A canary release is a new software version that is initially rolled out to only a small portion of users. If errors occur there, the rollout is stopped before the majority is affected.

CANN is the software used to drive Huawei's AI chips. It translates AI programs into instructions for Ascend processors, making it Huawei's counterpart to Nvidia's software package CUDA.

A capability threshold is a predetermined limit on what an AI system is allowed to be capable of before additional safety measures kick in. If a model reaches this limit during testing, the developer must respond – for example with stricter safeguards or by halting release.

Capex refers to a company's spending on long-lived assets such as buildings, machinery, or data centers. In the AI industry, capex has become the most important number, because building data centers now costs hundreds of billions of dollars.

A capex forecast is a company's announcement of how much money it plans to spend in the coming year on long-lasting acquisitions such as buildings, machinery, or data centers. For tech companies, it is considered the most important indicator of how strongly they believe in the success of AI.

Capture the Flag is a competition in which participants deliberately break into intentionally insecure computer systems and find hidden codewords stashed inside. Today, these challenges have also become a standard test for how well AI systems can independently find security vulnerabilities.

A cargo cult is the imitation of outward forms without understanding the underlying causes. In technology and science, the term refers to practices that look correct but are ineffective.

Causal Attention is a rule in language models that ensures each word, during computation, may only look at previous words, never at later ones. Without this rule, a model could copy the answer during learning instead of practicing genuine prediction.

CBRN is an umbrella term for the four most dangerous types of weapons: chemical, biological, radiological, and nuclear. In the AI debate, the abbreviation stands for the concern that language models could assist in building such weapons.

CBRN risks are dangers posed by chemical, biological, radiological and nuclear substances that can harm very large numbers of people at once. In the AI debate, the term refers to the concern that a language model might help someone build such weapons.

CBRN weapons are chemical, biological, radiological, and nuclear weapons – that is, weapons that can harm very large numbers of people at once using toxins, pathogens, or radioactivity. In the AI debate, the acronym stands for the principle that AI systems must not provide instructions for building such weapons.

Chain-of-Thought refers to the technique of getting an AI language program to work through a task in intermediate steps instead of immediately stating the result. For tasks involving logic, mathematics, or multiple conditions, this significantly increases the accuracy rate.

A Change Data Feed is a running list of all changes made to a data table: what was newly inserted, what was deleted, what was modified? Downstream systems only read these changes instead of loading the entire table every time.

A chatbot is a computer program that you can converse with via text or voice and that responds automatically. Modern chatbots like ChatGPT are based on language models that have learned from vast amounts of text which word typically follows next.

A checkpoint is a saved intermediate state of an AI model: a file that captures all learned numerical values at a specific point in time. This makes it possible to resume interrupted training, restore an earlier state, or pass a model on to others.

Checkpointing means regularly writing the current intermediate state of a learning computer program to disk. If the computing system crashes, you resume from this saved state instead of starting all over again.

A chiplet is a small silicon building block that is assembled together with other such building blocks in a shared package to form a large processor. This construction method is cheaper and more flexible than manufacturing an entire processor from a single piece.

Chips Act is the name of government funding programs through which the US and the EU pay for the construction of computer chip factories on their own soil. The goal is to become less dependent on manufacturers in Asia.

Chunking refers to breaking down long texts into smaller, self-contained sections called chunks. It is a standard step when AI systems need to search large document collections for relevant passages.

A CI/CD pipeline is an automatic chain of work steps that checks new program code, assembles it, and delivers it through to the finished product. It ensures that software teams can publish changes daily without anyone having to run the checks by hand.

Circular Financing refers to deals in which a company gives money to a customer, who then uses that same money to buy the company's own products. The money moves in a loop, but the resulting revenue looks like genuine external growth on the balance sheet.

Citation rules determine when and how an AI system must name the sources its answer is based on. They decide whether a statement can be verified and whether the original authors remain visible.

A claim model is a computational model used by insurers that predicts how many claims will occur and how costly they will be. Modern versions use machine learning and thereby influence prices, reserves, and the processing of individual claims.

Class A shares are one of several share classes issued by a company, typically differing in voting rights. Holders of such shares often have more or less influence over corporate decisions than holders of other classes within the same company.

A classifier is a computer program that sorts things into predefined categories — for example, an email as spam or not spam. It learns this sorting from many examples, rather than following fixed rules.

CLAUDE.md is a simple text file in a software project that tells Anthropic's AI assistant Claude how work is done in that project. It is automatically read by the AI as soon as it starts up in the project folder, thereby replacing instructions that would otherwise have to be typed out anew each time.

A clean room is a working method in which a team recreates something without ever having seen the original code or original documentation. This makes it possible to prove that nothing was copied – an important safeguard against copyright lawsuits.

Clean Room Design is a working method in which a team recreates a foreign product without ever having seen its original documentation. It serves to refute the accusation of copying in court.

In a clean-room reimplementation, a team rebuilds a piece of software without ever having seen the original source code. It works only from a description of what the program is supposed to do — the aim being to prove that nothing was copied.

CLI stands for Command Line Interface, meaning you operate a computer by typing text commands instead of using windows and mouse clicks. Programmers use it every day because typed commands can be repeated exactly, saved, and automated.

The click-through rate measures how many people click on something that has been shown to them. It is one of the most important metrics in the online advertising business and indirectly controls what recommendation systems present to us in the first place.

Client-Side Encryption means: data is made unreadable on the user's own device before it is ever sent to a provider on the internet. The provider only stores the unreadable jumble of data and does not possess the key.

A client timeout is a time limit that a requesting program sets for itself: if no response arrives from the server within this period, it cancels the request. The server often keeps working in the meantime — it initially has no idea the request was abandoned.

The Clipper Chip was an encryption chip for telephones introduced by the US government in 1993, intended to keep calls secret – but with a built-in duplicate key for authorities. After massive criticism from experts and civil rights advocates, the project failed in 1996, and it is still regarded today as a textbook example of the debate over government backdoors.

In advertising, closed loop refers to a self-contained cycle: an ad is displayed, the user's reaction is measured, and the result immediately feeds back into the decision about who sees which ad next. Because the feedback is automatically processed further, the system continuously improves itself.

Closed Source means that a company keeps the blueprint of its program secret. Users are allowed to use the program, but they cannot look inside it, copy it, or modify it themselves.

Closed-loop cooling is a cooling system in which the same fluid continuously circulates in a loop, carrying heat away from computer chips. Because nothing evaporates and nothing is drained off, such a loop consumes almost no water – an important argument in the debate over data centers' water consumption.

The CLOUD Act is a US law from 2018. It obligates American providers such as Microsoft, Google, or Amazon to hand over customer data to US authorities — even when that data is stored on servers outside the United States.

Cloud Run is a service from Google that runs programs without requiring you to set up your own server. The service automatically starts the program when a request comes in and shuts it down again when there is nothing to do.

A Cloud Summit is a professional conference where providers of internet-based computing power present their latest developments. Today, such events are primarily stages for announcements related to artificial intelligence and are accordingly closely followed by customers, journalists, and investors.

The cloud backlog is the sum of all amounts that customers have contractually committed to a cloud provider but that have not yet been paid out, because the service will only be delivered in the future. Investors read the figure as a preview of how strongly revenue is likely to grow in the coming quarters.

Cloud deployment means: a program doesn't run on your own computer, but on rented machines in a large data center, reachable via the internet. For AI services this is the normal case, because the computing power needed would be far too expensive for individual users.

A cloud lab is a real research laboratory where robots carry out the experiments and researchers submit their trials over the internet. Instead of standing at the lab bench themselves, they send an experimental instruction as a file and receive the measurement results back.

A cloud region is a geographic area in which a provider such as Amazon, Microsoft, or Google operates multiple data centers. Customers select the region and thereby determine where their data physically resides and how quickly their applications respond.

Cloudflare Workers is a service that lets developers run small programs directly in the worldwide distributed data centers of the company Cloudflare. The code starts right where the visitor happens to be, which is why it responds very quickly.

A compute cluster is a group of many individual computers connected via a fast network that work together on a single task. For training large AI models, such clusters are the foundation, because a single machine would be far too slow for that.

A CMS is software that lets you publish text, images, and videos on a website without touching the site's underlying code. Well-known examples include WordPress, Typo3, and Shopify.

Co-design means that two things that will later have to work together are also designed together – in technology, this usually means a chip and the software that runs on it. Instead of developing one after the other, both sides are adapted to each other right from the start.

Co-evolution refers to the development of two sides that drive each other forward: every change on one side alters the conditions for the other. In the world of technology, the term describes, for instance, how AI systems and their users, attackers and defenses, or software and hardware mutually escalate one another.

Co-Intelligence describes the collaboration between humans and AI programs as a team in which both sides contribute different strengths. The term was coined by business professor Ethan Mollick and now serves as a guiding concept for how people should meaningfully use AI at work.

CO2 neutrality means that a company or a country no longer releases any carbon dioxide into the air on balance. Emitted amounts are either avoided or removed from the atmosphere again elsewhere.

COBOL is a programming language from 1959, designed for accounting, payroll, and banking operations. It is considered outdated, but continues to run in banks, insurance companies, and government agencies today, managing enormous sums of money on a daily basis.

The cobra effect describes a measure that produces exactly the opposite of its intended goal, because people or machines exploit the rule in unexpected ways. In the AI debate, the term stands for systems that perfectly fulfill their stated objective and yet do something undesirable.

Code Arena refers to an online platform where different AI systems solve the same programming task, after which people choose the better solution. From many such duels emerges a ranking that shows which system is most convincing at coding.

Code Churn measures how much newly written program text is changed or deleted again shortly after its creation. The metric is considered an indicator of how mature a piece of software is and how clear the requirements for it are.

A Code of Practice is a voluntary set of rules through which an industry demonstrates how it implements legal requirements in practice. In the AI world, the term mainly refers to the guide that the European Union negotiated with major AI providers for handling their models.

A code review is the examination of new lines of code by other people before that change becomes part of the finished product. In software development, it is the most important filter against bugs, unnecessary complexity, and security vulnerabilities.

Code execution means that an AI system actually runs program lines it has written itself and reads off the result, instead of merely formulating an answer. This makes calculations, tables, and charts more reliable — but the executed code can also be dangerous if it doesn't run in isolation.

The codebase is the entirety of all text files from which a program is built. It is the actual working material of software developers and, for large products, can grow to millions of lines.

Codex is the name OpenAI gives to its systems for writing program code. From a description in plain language, Codex generates finished code, tests it, and proposes changes to existing projects.

Codex agents are programs that write and modify software on their own: you describe a task in plain language, and the agent then works on real project files and tests its own work. The name comes from OpenAI's Codex system, but today it is often used generally for programming assistants that work this way.

The Codex app is a program by OpenAI that handles programming tasks independently: you describe in plain language what the computer should do, and the app writes and tests the necessary program code. It is available as a standalone application, as a command-line tool, and as a service in the browser.

A Coding Agent Index is a searchable directory covering a project's entire source code, which an AI assistant builds before helping with programming. It ensures the assistant can find the right files without having to read millions of lines of code every time.

A coding agent is an AI program that doesn't just suggest code line by line, but works on a task independently in multiple steps: reading files, changing code, running tests, fixing errors. The human describes the goal and reviews the result at the end.

A coding benchmark is a standardized test used to check how well an AI system solves programming tasks. The written programs are automatically executed and compared against predefined checks, resulting in a percentage score.

Cognitive load refers to the amount of information a person must hold and process in their mind at the same time in a given moment. Because this working memory in the brain is very small, this load determines whether we understand something or give up overwhelmed.

Cognitive offloading refers to shifting mental work onto aids outside one's own head – such as a notepad, a navigation device, or an AI chatbot. The term originates from psychology and is now used above all in the debate about the influence of AI tools on thinking.

Cognitive Surrender describes people handing over their own thinking to computer programs and barely checking the answers anymore. The term is used above all where chat programs pre-formulate decisions in everyday life.

COGS stands for "Cost of Goods Sold" and refers to the costs that arise directly because a product is sold or a service is provided. At AI companies, these are mainly the compute costs incurred in data centers, which is why COGS has become the key metric for the question of whether AI businesses can be profitable.

Cold Start refers to the situation in which a system does not yet have any data or experience and therefore performs poorly. The term is mainly used for recommendation systems, but also appears in the context of servers and voice assistants.

A cold wallet is a storage method for cryptocurrencies in which the secret access key permanently remains without an internet connection. This means attackers from the network cannot steal it, but paying becomes less convenient.

Colocation means: A company places its own computers in a building owned by someone else, purpose-built for this – with power, cooling, internet connectivity, and security. The equipment belongs to the company, while the building and its supply infrastructure belong to the provider.

Colossus is the name of the massive data center operated in Memphis by Elon Musk's company xAI. It houses hundreds of thousands of specialized chips used to train the AI models of the Grok series.

Colossus 2 is the second major computing facility of the AI company xAI in Memphis, Tennessee, in which hundreds of thousands of specialized chips work together on a single AI model. The facility is considered one of the largest training data centers in the world and at the same time represents the enormous power and capital demands of current AI development.

Command and Control refers to the remote control of computers that an attacker has secretly taken over. Through a central server, the attacker sends commands to the infected devices and retrieves data from them.

Command Injection is a security vulnerability in which an attacker uses a normal input field to make their own system commands run on someone else's computer. The cause is almost always that a program passes user input to the operating system without checking it.

A command line interface is a user interface where commands are typed as text instead of clicking on icons. It's considered old-fashioned, but it remains the most important tool in software development and the operation of AI systems to this day.

Command Validation refers to the checking of a command before a computer executes it. It ensures that only permitted and sensible instructions are let through — especially important when AI programs generate commands autonomously.

A command-and-control server is the machine from which attackers control their malware on other people's devices. Through it they send commands and collect stolen data — and that's exactly why it's investigators' favorite target.

A command-line interface is a way of operating a computer: you type commands as text instead of clicking on icons with a mouse. It is considered old-fashioned, but remains the most important tool in software and AI development to this day.

A commit is a saved intermediate state of a software project, together with a short note about what was changed. All commits strung together form the complete development history of a program.

A commit hook is a small program that starts automatically the moment someone saves a change to program code. It checks the change and can abort the save if something is wrong.

Commoditization describes how a product turns from a distinctive offering into an interchangeable mass-market good. When many providers deliver nearly the same thing, price alone decides — in the AI industry, this now applies to many language models.

Commoditization describes how a once-special product turns into interchangeable mass merchandise because many providers deliver almost the same thing. Price then becomes the most important distinguishing factor – in the AI industry, this currently affects language models and computing power.

A commodity is a product that is essentially the same everywhere, so that price alone almost decides the purchase. In the tech and AI industry, it's the buzzword for a once-special technology turning into interchangeable, mass-market goods.

Commodity hedging refers to protection against fluctuating commodity prices. Companies use contracts to lock in the price of oil, copper, or wheat in advance, so they aren't hit by expensive surprises later on.

Community Notes are short additional annotations that users themselves write beneath misleading posts on social networks. A note only becomes publicly visible if people with otherwise opposing views jointly find it helpful.

Compaction refers to consolidating and cleaning up accumulated data so that a system doesn't choke on its own volume of data. The term comes from database technology and today appears mainly in AI assistants that need to condense long conversations.

Companion AI refers to programs that present themselves as permanent conversation partners and simulate a personal relationship with the user. They remember preferences and past conversations, often have a name and a character – and for many providers they are a subscription business.

A compiler is a program that translates human-written program code into a computer's command language. Without this translation, a processor could make nothing of instructions like "if" or "print".

A compliance checker is a program that automatically verifies whether a text, a piece of software, or a business process complies with applicable rules. It does not replace humans, but finds violations faster and reports them for review.

A compliance agent is a computer program with artificial intelligence that independently checks whether a company complies with laws and internal rules. It reads documents, monitors processes, and reports anomalies to humans, who then make the decisions.

Compliance infrastructure is the sum of software, processes, and responsibilities with which a company demonstrates that it adheres to laws and its own rules. For AI products, this includes things like logs of training data, checks of outputs, and documented approvals.

Composer is a feature in the code editor Cursor in which an AI modifies several files of a program at once based on a simple instruction. Instead of individual suggestions, it delivers a complete rework, which the human then reviews and accepts or rejects.

Compositing refers to combining multiple image layers into a single, seemingly real image. Films, commercials, and today also AI-generated videos are almost always created from such overlaid layers.

Compound Engineering is a way of working in software development in which every solved task leaves a lasting trace: a note, a rule, an automated test. The goal is for teams using AI tools not just to work faster, but to repeat the same mistakes less and less from week to week.

Compressed Sparse Attention is a computational method in language models that partly summarizes and partly selectively focuses on long texts, instead of comparing every word with every other word. This allows programs like chatbots to process very long inputs without the computational cost exploding.

Compute is the amount of computational work needed for a task – in AI, primarily for training and generating responses in large models. In the tech industry, compute is simultaneously a cost center, a scarce resource, and a factor of power.

Compute allocation refers to the decision of how much computing power a company assigns to which project. Because powerful computing chips are scarce and expensive, this distribution is one of the most important strategic decisions in the AI industry.

A compute cluster is a large group of computers connected by fast cables that work together on a single computing task. In the AI industry, such clusters are the machines on which large language models are trained.

Compute discipline refers to a company's deliberate, frugal handling of expensive computing power for AI. It describes the stance of treating compute time like a scarce budget rather than deploying it without limit.

Compute futures are contracts that let you buy or sell computing power for a later period today at a fixed price. They are meant to make compute time as tradable as electricity, wheat, or oil.

Computer vision is the field in which computers learn to recognize the content of images and videos. Software converts raw pixels into a statement, such as "there is a pedestrian here."

Computer-Generated Imagery, or CGI for short, refers to images and moving images that are created entirely within a computer rather than in front of a camera. The technology is behind feature films, video games, commercials, and product images in online shops.

A computer-use agent is an AI program that operates a computer the way a human does: it sees the screen, moves the mouse pointer, and types on the keyboard. Rather than just writing text, it carries out tasks itself in real programs and websites.

A computing voucher is a credit that lets research groups, start-ups, or government agencies pay for computing time on other people's high-performance computers instead of buying it with money. Such vouchers are usually issued by governments, the EU, or large cloud providers to specifically promote AI development.

Conditional branching means that a program checks at a certain point whether a condition is true, and continues differently depending on the outcome. This conditional branching is a fundamental building block of any software and today also controls which path an AI assistant takes through its work steps.

Constitutional AI is a training method in which an AI text program critiques and improves its own responses based on a written list of rules. Instead of thousands of human ratings, this list of rules provides the standard for what counts as a good answer.

Constraints are fixed conditions that a solution must satisfy in order to be admissible at all. In computer science and AI, they narrow down the space of possible answers before anyone even starts searching for the best answer.

A container is a package that bundles a program together with all the auxiliary files it needs to run. This means the software behaves the same way on every computer, whether on a developer's laptop or on a server in a data center.

Container pooling means that a data center continuously keeps a set of ready-to-use, pre-prepared work environments on hand instead of starting each one up from scratch. When a request comes in, it immediately receives a free environment from this pool and doesn't have to wait for a startup.

Container virtualization packs a program together with all the files it needs to run into a self-contained package. This package runs the same on any suitable machine – from the developer's laptop to the server in the data center.

Containerization is a technique in which a program, together with everything it needs to run, is packed into a self-contained package. This package runs the same way on any suitable computer – on the developer's laptop just as on a server in a data center.

Content moderation is the control of content that users upload to a platform. Humans and software check whether a post violates laws or the platform's rules and, in doubtful cases, remove it.

A content management system is software that lets you manage and publish text, images, and videos on a website without programming yourself. Well-known examples are WordPress, Shopify, and TYPO3.

Context bleed refers to the unwanted leakage of information from an earlier part of a conversation into a later, supposedly unrelated response. AI-powered chat programs mix up topics, roles, or user data that should have remained separate.

Context Engineering is the art of giving an AI text system exactly the information it needs for a good answer – no more and no less. It often determines the quality of an answer more strongly than the choice of system itself.

A Context Machine is a system that automatically assembles the right information for an AI language model before it answers. Instead of better questions, it delivers better material: documents, notes, search results, and tool outputs, neatly sorted and trimmed down to what matters most.

Context poisoning refers to an attack in which someone secretly slips manipulated text to an AI chatbot that the chatbot reads along while formulating its answer. The program treats the hidden instructions as genuine information and acts on them.

The context window is the amount of text an AI language model can keep in view at the same time. Everything within it flows into the response, everything outside it simply doesn't exist for the model.

Continual Learning refers to methods that allow a computer program to keep learning from new examples over time without losing what it has already learned. The problem: programs of this kind very easily overwrite old knowledge.

A Continuous Glucose Monitor is a small sensor worn on the body that measures the sugar level in tissue around the clock and sends the readings to a phone or a reader device. For many people with diabetes it replaces repeated finger-prick testing and delivers data series that are also analyzed by software and AI systems.

Continuous Integration, or CI for short, is a way of working in software development: every change to the program code is immediately merged with the team's current state and automatically tested. This way, errors are caught within minutes rather than weeks later.

Continuous Integration means that programmers merge their changes several times a day and have them automatically checked. This way, errors surface within minutes instead of only after weeks.

The control plane is the governing part of a computer system: it decides who is allowed to do what and when, but it does not carry out the actual work itself. The work is performed by the data plane, the layer that actually processes and forwards data.

Controllable Effort refers to the ability to configure how much computational effort an AI model puts into a single answer. Low effort means fast and cheap answers, high effort means more thorough results at higher cost.

A Convolutional Neural Network, or CNN for short, is a computational method that independently extracts patterns from images. It examines an image in many small sections and gradually assembles whole objects out of simple shapes.

Conway's Law states: The structure of a product mirrors the structure of the organization that built it. Those who work in separate teams that barely talk to each other end up with separate, poorly matching product parts.

A Coordinator Agent is an AI program that breaks a large task down into subtasks and distributes them to other programs. It does little substantive work itself, instead planning, delegating, and assembling the results at the end.

Copyleft is a licensing condition for software and other works: you may freely use, modify, and redistribute them – but only under the same terms. This keeps the source code open in all later versions as well.

Core Web Vitals are three metrics Google uses to assess how fast and pleasant a webpage feels to visitors. They measure the loading time of the main content, responsiveness to clicks, and unwanted shifting of elements.

A cornered resource is a valuable asset that a company owns exclusively and that competitors cannot simply replicate — such as a patent, a raw material deposit, or a unique data archive. The term originates from strategic analysis and is currently often used to explain why some AI providers are barely challengeable.

A counter-metric is a second measure that is tracked to check whether improving a primary number causes harm elsewhere. It acts as a kind of warning light against optimization running in the wrong direction.

A Covered List is a register of companies or securities that a bank, research firm, or regulator regularly monitors and to which special rules apply. The term appears mainly in financial analysis and compliance regulations, where it must be defined which securities are "covered."

CPA optimization refers to the automatic management of online advertising with the goal of spending as little money as possible per customer action achieved – such as per purchase or sign-up. A learning system of the advertising platform decides anew for each individual ad delivery how much to bid.

CPM is the price an advertiser pays for one thousand impressions of their ad. The metric makes ad placements on different websites, apps, and platforms comparable.

The CPU is the main processor of a computer: the component that carries out computational steps one after another and controls the rest of the device. It is extremely versatile, but falls short when it comes to the massively parallel calculations of modern AI – there, graphics chips take over.

The crawl budget is the number of pages a search engine fetches from a website within a given period of time. It plays a role in determining how quickly new or changed content even appears in search at all.

The crawl-to-refer ratio compares how often a search engine's or AI provider's program automatically reads a website with the number of visitors that provider sends back in return. It has become the central metric in the debate over whether AI providers treat publishers fairly.

A crawler is a program that autonomously visits web pages, stores their content, and follows the links found on those pages. Search engines and AI companies use crawlers to build massive collections of text and images.

Credentials are the proofs with which a person or a program identifies itself to a computer system – for example a username and password or a digital key. Whoever gets hold of them can pose as the legitimate user, which is why stolen credentials are the most common entry point in hacker attacks.

A credit system is a billing method in which, instead of money, you buy an internal balance currency and spend these points when using a service. Among AI providers, it is the common way to charge very differently priced tasks with a single price tag.

CRM stands for Customer Relationship Management and refers to software in which a company collects everything about its customers and prospects. Every call, every email, and every offer is recorded there so the sales team keeps an overview.

Cron is a service on servers that automatically starts programs at fixed times – for example every night at three o'clock. The individual tasks are called cron jobs and are described in a text file using five time specifications.

A cron job is a task that a computer automatically executes at set times, for example every night at three o'clock. The name comes from the program "cron," which has been managing such schedules on servers for decades.

Crowdsourcing means assigning a task not to individual employees but publicly to a large group of volunteers or paid casual workers. Many small contributions together add up to a result that a single company could hardly manage alone.

A CRT monitor is a display that generates its image using a large glass tube, inside which an electron beam excites a luminescent coating. This technology dominated televisions and computer screens until around 2005, when it was replaced by flat-panel displays.

A CSS parser is the part of a browser that reads styling instructions for web pages and translates them into a form the program can compute with. Only after that does the browser know which font, color, and size an element should get.

CSV is a very simple file format for tables: each line is a row of the table, and commas separate the individual fields. Because almost every program can read CSV, it is the standard way to exchange data between systems.

CTR (Click-Through Rate) measures how many people click on something that was shown to them. It is the most important metric in online advertising and the training benchmark for many recommendation systems.

CTSS was a control program for mainframe computers that ran at MIT starting in 1961 and allowed multiple people to use the same computer simultaneously. It is regarded as the first practically deployed system of its kind and shaped nearly everything that came after it.

CUDA is a software environment from Nvidia that allows programmers to use graphics cards for general computing tasks. It is the reason why almost all AI research runs on Nvidia hardware.

CUDA is a software collection from the company Nvidia that lets programmers use graphics cards for general-purpose computing tasks. Because nearly all AI programs are built on it, CUDA is considered the most important reason for Nvidia's dominant position in the AI business.

Cumulative Layout Shift is a metric for how much content on a webpage still shifts around while it's loading. The higher the value, the more restless the page feels, and the more likely you are to accidentally click the wrong thing.

curl is a small program used to fetch or send data over the internet from the command line. It is the standard tool for testing web addresses and programming interfaces without a browser.

Curriculum Learning is a training strategy for AI systems: examples are not shuffled randomly but shown in a planned order, usually from easy to hard. The goal is a program that learns faster and ultimately works more reliably.

Cursor is a program for writing software that has a language AI built directly into it. Developers describe in plain English or German what their program should do, and the AI suggests the matching code.

A custom agent is a self-configured AI program that performs a fixed task for a specific person or company. For this, it is given its own instructions, its own data, and access to selected tools such as calendars, databases, or email.

A custom chip is a computing component that a company has designed for a single, precisely known task instead of buying an off-the-shelf chip. Because it doesn't need to do anything else, it works faster and more efficiently at that task — but development is expensive and time-consuming.

Custom code is self-written program code that a company develops precisely for its own purposes instead of buying a ready-made standard solution. It fits perfectly, but permanently causes effort for maintenance and upkeep.

Custom Instructions are permanently stored settings that a user provides once to an AI chatbot, which then automatically apply to every subsequent conversation. They specify things like language, length, tone, or background knowledge about the user, so you don't have to type it all out again each time.

Custom Silicon refers to computer chips that a company has specially designed for its own tasks, rather than buying standard off-the-shelf chips. In the AI industry, almost every major corporation now builds such chips of its own in order to cut costs and reduce dependencies.

CVE is a globally uniform numbering system for publicly known security vulnerabilities in software. Each vulnerability receives a unique identifier such as CVE-2024-3094, so that everyone involved can reliably talk about the same problem.

The CVSS score is a number from 0 to 10 that indicates how dangerous a security vulnerability in software is. It helps companies decide which problem they need to fix first.

Cybench is a standardized test that researchers use to measure how well AI systems solve real hacking tasks. It consists of competition tasks from the field of IT security that the system must complete independently on a practice computer.

The Cyber Kill Chain is a model that breaks down an attack on a computer network into seven consecutive steps. Anyone who knows these steps can disrupt an attack before it reaches its actual goal.

Cyber Safety refers to protecting people from harm caused by the use of digital services – such as fraud, harassment, or dangerous content. Unlike technical cybersecurity, it is not about systems but about the well-being of users.

A cyber classifier is a program that automatically sorts incoming texts or requests according to whether they relate to attacks on computers. Providers of AI systems use such sorting programs to detect dangerous requests before a response is sent out.

CyberGym is a test procedure that checks how well AI programs find real security vulnerabilities in software. Instead of quiz questions, the programs are presented with actual program code from well-known open-source projects and are supposed to demonstrate flaws within it.

A cybersecurity clearinghouse is a central body where companies and government agencies collect and share information about internet attacks and security vulnerabilities with one another. The goal is for all participants to learn about a threat faster than if each only knew about its own incidents.

Containment is the step in a computer attack where the spread is stopped before cleanup begins. Affected devices are disconnected from the network or isolated so the damage doesn't keep growing.

A convolutional layer is a component in image recognition programs that scans an image piece by piece with a small grid, thereby finding patterns such as edges or corners. Several such layers stacked one after another gradually build complex shapes from simple patterns.

Color space conversion means recalculating the same color from one numbering system into another, for example from the description used for screens into the one used for printing presses. It ensures that a photo looks as similar as possible on a phone, monitor, and paper.

A closed model is an AI system whose inner workings the manufacturing company keeps to itself. Users can only use it via an interface on the internet, but cannot download, copy, or modify it.

Calibration describes whether the confidence values given by a learning computer program match reality. If a system says "80 percent certain" in a thousand cases and is correct in about 800 of them, it is well calibrated.

Candidate generation is the first step in recommendation and search systems: out of millions of possible entries, a small preliminary selection is drawn in a flash. Only afterward does a second, more precise step decide what the user actually gets to see.

A catalog query is a search within an ordered directory that describes what data exists and where it is located. It does not deliver the data itself, but information about the data – making it the first step in many analyses and AI projects.

A call option is a contract that gives the buyer the right to purchase something at a later date at a price fixed today — but not the obligation to do so. For this right, they pay a fee upon signing, which they lose in any case.

Classification means that a computer program assigns a case to one of several predefined categories, such as labeling an email as spam or not spam. The program learns this assignment beforehand from many examples in which the correct category was already known.

A classifier is a computer program that sorts things into predefined categories — for example, classifying an email as "spam" or "not spam." Such programs usually learn this sorting from many examples rather than following fixed rules.

Coding theory studies how to write down data so that it remains readable even when disturbances occur. It provides the mathematical methods that let phones, hard drives, and space probes detect and repair transmission errors themselves.

A cognitive architecture is a blueprint for how the individual parts of a thinking system work together: perception, memory, decision-making, and action. It describes not a single program, but the fixed underlying framework within which various tasks are solved.

The command line is a way of operating a computer where you type commands as text instead of clicking on icons. It's considered old-fashioned, but it remains the most important tool for programmers and in AI development to this day.

A command line interface is a way of operating a computer by typing commands instead of clicking on icons. It looks old-fashioned, but in software development and on servers it remains the fastest and most precise way to work to this day.

Commoditization describes how a product moves from being a rare special case to interchangeable mass-market goods. In the AI industry this means: language models increasingly resemble one another, prices fall, and the provider becomes interchangeable.

The competency trap describes how a company clings to a method because it has become very good at it – and therefore overlooks a better new method. Past success thus becomes the cause of later failure.

Connectionism is the idea that mental achievements such as recognition or speech emerge from the interplay of very many simple, interconnected computing units. It is the theoretical foundation of today's artificial neural networks and thus of nearly all modern AI systems.

A connector is a ready-made link between two programs: it enables one piece of software to read data from another system or write data to it. In the AI world, connectors are what link a language model to company data such as emails, databases, or cloud storage.

A connector is a ready-made link between two programs that ensures data gets from one system into another. In the AI world, connectors are what link a chatbot to emails, company documents, or databases.

A consensus output is the answer that multiple computation runs of a program agree on by majority. Instead of taking a single output, the same task is solved multiple times and the result that occurs most frequently is chosen.

Consumerization describes how new technology today reaches private individuals first, and only afterward makes its way into companies and government agencies. It used to be the other way around: large organizations got new technology first, and private individuals much later.

Context tokens are the small chunks of text that an AI language program has in view at the same time while writing a response. Their number limits how much text the program can take into account at once – and it also determines what usage costs.

The context window is the amount of text an AI language program can take in at once. Anything outside this window doesn't exist for the program at the moment it responds.

Continual learning refers to the ability of a computer program to keep learning new things without forgetting what it learned before. In today's AI systems, this is precisely an unsolved problem: new knowledge often overwrites old knowledge.

Contrastive pairs are two examples that are deliberately shown together to a learning computer program: one that is desired, and one that is not. From the difference between the two, the program learns faster and more accurately than from single examples.

Contrastive pruning is a method for removing superfluous components from a trained AI model while using counterexamples to check which parts are truly important. The goal is a smaller, faster model that still recognizes the differences between similar inputs.

Conversation-based advertising formats are ads that appear within a conversation with a computer program, for instance in a chat window. Instead of a rigid advertising banner, the advertising becomes part of the dialogue and is tailored to what the user has just asked.

The conversion rate indicates what share of visitors to a website or app actually carries out a desired action, such as a purchase or a sign-up. It is one of the most important metrics in online business because it shows how well attention is turned into actual business.

The cortex is the outer, folded layer of the human brain, where perception, language, and thought are processed. In the world of AI, it serves as a model and namesake: many technical terms and product names draw on its structure of uniform nerve cells organized in layers.

Cost per token indicates how much money a single unit of text costs that an AI language program reads or writes. This metric is the standard yardstick for how expensive it is to operate such programs.

Credit card data is the information needed to pay with a credit card: card number, name, expiration date, and security code. Because these few numbers are enough to make a payment, they are a popular target for data theft and are therefore strictly protected.

Cryptanalysis is the science of reading encrypted messages without possessing the matching key. It uncovers weaknesses in encryption methods and thereby decides which methods we can trust on the internet.

A cryptographic key pair consists of two related sequences of numbers: one that anyone may know, and one that stays secret. Whatever was secured with one can only be opened or verified with the other – the foundation of HTTPS, digital signatures, and crypto wallets.

A coding agent is an AI program that independently handles a programming task in multiple steps: it reads files, writes code, tests it, and fixes errors. Unlike a pure suggestion assistant, it keeps working over an extended period without needing a new instruction for every step.

Compute capacity describes how many computational steps computers can perform in a given amount of time. In the AI industry, it is the scarcest and most expensive resource: without enough of it, large models can neither be trained nor operated.

Computing power describes how many computational steps a computer can perform per second. In the AI industry it is the scarcest and most expensive resource: without sufficient computing power, no large model can be trained or operated.

A compute agreement is a contract through which a company secures long-term access to third-party computers in order to compute AI programs. Such contracts often have a volume of several billion dollars and run for years.

A cleanroom is a workspace in which the air is kept extremely clean, because even tiny dust particles would disrupt production. Without such rooms, computer chips, medications, and displays could not be manufactured with consistent quality.

A code of practice within the meaning of the EU AI Act is a voluntary set of rules with which AI providers demonstrate that they comply with the requirements of the law. Those who sign have to provide less proof themselves — but in return commit to concrete rules on transparency, copyright, and safety.

The centaur configuration refers to a way of working in which a human and a computer program jointly solve a task, rather than either one deciding alone. The name comes from chess, where such teams temporarily played stronger than the best human and the best computer alone.

Circularity describes deals in which money flows in a circle between a small number of companies: a company invests in a customer, who then uses that same money to buy the company's own products. In the AI industry, the term is used critically because it creates revenue that looks like genuine demand but is partly self-financed.

The depreciation cycle is the period of time over which a company spreads the cost of an expensive purchase for accounting purposes. In AI data centers, it has become a central point of contention: do the graphics cards last three years or six?

The D-Band is a radio frequency range of roughly 110 to 170 gigahertz. It offers enormous room for fast data transmission, but has only a very short range and is considered a candidate for future mobile network generations.

A daemon is a program that runs permanently in the background of a computer and waits for tasks, without any window or user interface being visible. Web servers, print queues, and automatic updates are typically handled by such programs.

A Daily Driver refers to the AI model that someone uses by default in everyday life, without much deliberation. It is rarely the strongest model available, but rather the one with the best balance of quality, speed, and price.

Dark Compute refers to computing power in data centers that has been installed and paid for but is not being used. The term has become important because companies are pouring billions into chips for artificial intelligence, and part of that capacity remains unused.

A dark factory is a factory that runs entirely without humans in production: machines and robots handle all the work steps themselves. The name comes from the fact that in such a hall, light wouldn't even be necessary, because no one needs to see anything there.

Dark fiber refers to laid fiber-optic cables through which no light has yet been sent and which therefore lie unused in the ground. Companies lease such fibers and operate them with their own equipment instead of buying data lines from a network operator.

Dark Output refers to texts, images, or code that an AI has generated but that no one ever reads or checks again. Such content accumulates in companies and across the web, causing costs and risks without providing any benefit.

Dark patterns are design tricks in apps and on websites that push users toward decisions they didn't actually want to make. Typical examples include hidden cancellation buttons, pre-selected add-on costs, or artificial time pressure during a purchase.

A data clean room is a shielded environment in which two companies combine their customer data without either side getting to see the other's data. Only predefined analyses are permitted, and only aggregated results are allowed to come out.

Data governance is the totality of rules, responsibilities, and controls with which a company organizes its data. It defines who may view, change, and share which data, and who is accountable for its accuracy.

A data loader is the piece of software that fetches training data from storage, prepares it in portions, and passes it on to a learning program. It plays a key role in determining whether expensive computing chips are kept busy or left waiting unnecessarily.

A data pipeline is a fixed chain of work steps that automatically transports data from its point of origin to its point of use, cleaning it up along the way. Without such chains, companies would have data, but no usable figures and no AI models.

Data poisoning refers to attacks in which someone deliberately smuggles false or manipulated examples into an AI's learning materials. The program learns the intended errors from this and later behaves exactly as the attacker planned.

Data Residency refers to the question of in which country the data of a company or public authority is physically stored. Many laws and contracts require that certain data must not leave the home country or the EU.

A data warehouse is a central repository in which a company collects data from many individual systems and organizes it in a consistent way. It is not meant for day-to-day operations, but for evaluation: reports, metrics, and analyses spanning long periods of time.

Database branching means creating an independent copy of a running database in seconds, one in which you can experiment safely. The copy initially takes up almost no storage space, because it shares the existing data with the original.

A data flywheel describes a self-reinforcing cycle: a product gets used, this usage generates data, that data makes the product better, and this attracts even more users. The term explains why large providers in the AI market often extend their lead rather than lose it.

A database migration is the planned move of stored data from one storage system to another, or a change to its internal structure. It is considered one of the riskiest undertakings in IT, because it involves moving a company's most valuable possession: its data.

Data exfiltration through models refers to the unwanted leakage of confidential information via an AI system — for instance when a chatbot blurts out company secrets or attackers get it to send data outward. It is one of the central security risks in the corporate use of AI.

A data moat is an advantage a company has because it possesses data that competitors cannot easily obtain. The image comes from the moat of a castle: it keeps attackers at a distance.

Data scarcity refers to the state in which there are no longer enough suitable texts, images, or measurement values available to train an AI. It is considered one of the central limits to the further progress of large AI systems.

Data localization means that a state mandates that digital data about its citizens be stored within the country itself. Such rules force tech companies to build expensive data centers on-site instead of managing everything centrally.

Data portability means that you can take your own data from one provider and transfer it to another. In the EU, a legal right to this has existed since 2018, applying to services such as social networks, streaming providers, or online shops.

A data contract is a binding agreement about what data must look like when one team passes it on to another. It defines fields, formats, meanings, and responsibilities so that downstream programs don't break unnoticed.

DCN v2 is a model architecture that recommendation systems use to predict whether a user will click on something or purchase something. Its distinctive feature is a component that systematically combines features like age, time of day, and product category with one another, rather than considering them only individually.

A DDoS attack deliberately overwhelms a website or online service with a flood of requests from thousands of sources until normal users can no longer get through. The attacker doesn't break into anything or steal data, they simply make the service unreachable.

De-skilling refers to the loss of abilities when machines or software take over a task that people previously had to master themselves. The term plays a major role in the debate about AI tools, because many professions are currently handing off parts of their routine work.

The dealer principle describes a business strategy in which a product is first offered for free or very cheaply so that users get used to it. Later, this dependency is converted into money, for example through rising prices or paid add-on features.

A Decentralized Autonomous Organization is a community that manages shared money and shared rules through a public computer program instead of through a management board. Decisions are made through votes of the members, and the outcome is executed automatically.

Deceptive Behavior refers to cases where an AI system claims or displays something to humans that differs from its actual internal computations. The term plays a major role in safety research because a deceptive system can pass normal tests without raising suspicion.

A decoder is the part of an AI model that generates text by always predicting the next piece of a word. Almost all of today's chatbots, like ChatGPT, consist entirely of such a decoder.

Decoding is the process by which a language model turns pure probabilities into finished text. It determines which word is actually selected next — and thereby decides whether a response sounds sober, creative, or incoherent.

Deduplication means removing duplicates from a data collection so that each piece of content remains only once. In AI development this is a standard step when cleaning training data, while in storage technology it saves space on hard drives.

Deep Learning Super Sampling is a technique from Nvidia in which a video game is rendered at a lower resolution and then upscaled to a higher resolution by a trained program. This allows games to run significantly smoother without the image visibly looking worse.

Deep Packet Inspection is a technique in which network devices look not only at where data is headed, but also at what's inside it. The technology is used for security, network management, and in some countries, also for censorship.

Deep Tech refers to companies whose products are based on a scientific or technical discovery and which therefore require long development times and a lot of money. Typical fields are quantum computing, new batteries, biotechnology, semiconductors, and space travel.

A deepfake is a fake photo, video, or audio recording created by a computer program that looks or sounds real. It usually shows people saying or doing things that never actually happened.

Defense in Depth is a security principle in which multiple independent protective measures are arranged one after another. If one layer fails, the remaining ones catch the attack instead of leaving everything immediately exposed.

The Defense Production Act is a US law from 1950 that allows the government to direct companies to prioritize the production of certain goods for the state. Today it is used not only for weapons but also for vaccines, batteries, chips, and data center equipment.

DeFi stands for "Decentralized Finance" — financial activities like lending, exchanging or saving that take place without a bank, directly through computer programs in an open network. These programs are publicly viewable and execute every rule automatically – which replaces intermediaries but also creates new risks.

Decoding refers to the step in which a language model actually selects one word from probabilities calculated for many possible next words. The chosen strategy determines whether a response sounds sober and predictable or varied and surprising.

The Delphi method is a procedure in which experts assess a question about the future across several anonymous rounds of questioning. After each round, they see the other participants' answers and may reconsider their own estimate.

Demand aggregation means: many individual buyers join forces and appear as one large order. Because large quantities are cheaper and easier to deliver, prices drop for everyone involved.

A demand underwriter is a company that contractually guarantees future demand for a product by committing to a fixed purchase before production even exists. In the AI space, large cloud providers take on this role when they make data centers or chip fabs financeable through long-term offtake agreements.

A Dense Transformer is an AI model in which all learned computational building blocks are used for every request. This design is simple and reliable, but quickly becomes expensive to operate as size increases.

A denylist is a list of things that are explicitly forbidden — such as words, internet addresses, or programs. Everything not on the list is allowed; this is what distinguishes it from its opposite, the allowlist.

Dependency Confusion is an attack on software companies in which an attacker publishes a public package with the same name as an internal company package. The company's build process then mistakenly downloads the foreign version and executes its malicious code.

DeployCo refers to a company that does not develop finished AI models itself, but instead builds them into concrete products for customers. The term distinguishes such companies from the labs that train the models.

Deployment refers to the step in which finished software moves from the developers' machines to the servers that serve real users. Only after this step does a new feature actually become usable for the public.

A deployment company is a company that integrates finished AI systems from other providers into the concrete workflows of businesses, rather than developing AI models itself. Its business is application: adaptation, rollout, operation, and support at the customer site.

A Deployment Safety Board is a standing body within an AI company that decides whether a new model or a new feature may be released. It reviews test results and risks beforehand and can delay, restrict, or completely stop a release.

A design document is a written plan in which a development team records, before coding begins, what is to be built and why exactly in that way. It describes the goal, the approach, rejected alternatives, and open risks, and is commented on by colleagues before work starts.

A design system is a binding collection of building blocks, rules, and ready-made code elements that a company uses to give all its apps and websites a consistent look and feel. It ensures that a button looks and works the same everywhere, so no team has to reinvent it.

A design token is a named value for a design decision, such as a color or a spacing amount. Instead of typing the color individually everywhere, the app just contains the name — and the value can be changed centrally.

In a design system import, a program adopts a company's established design rules – such as colors, fonts, and spacing – as a machine-readable specification. The result: newly generated screens automatically look the way the brand requires.

Distillation is a process in which a large, expensive AI program teaches a smaller one to behave as similarly as possible. Afterwards, the small program needs far less computing power and often even fits on a phone.

A detached task is a task that a program starts in the background and then no longer monitors. The caller doesn't wait for the result and continues immediately – this is practical, but can silently swallow errors.

Determinism means: given the same input, the exact same result always comes out. Classic software mostly behaves this way, whereas AI chatbots often don't — and that's precisely what makes them hard to verify.

Determinism in training means: if you repeat an AI learning process with exactly the same data and settings, you get exactly the same result. In practice, this is hard to achieve because small randomness in the computing hardware slightly alters each run.

A deterministic system always produces exactly the same result given the same starting conditions. The opposite are systems involving randomness, which include many AI applications.

Device Code Authentication is a sign-in method for devices without a convenient keyboard, such as TVs or game consoles. The device displays a short code that you enter on a website using your phone or computer to complete the sign-in there.

A DevTool is a tool program that helps people write, test, and publish software. It is not a product for end customers, but a work instrument for the people who build programs.

Diarization is the automatic segmentation of an audio recording according to who speaks when. It does not answer what was said, but instead separates the voices of multiple people into individual speech segments.

Die size refers to the area of a single computer chip on the round silicon disc from which it is cut. It is given in square millimeters and largely determines how expensive a chip is to manufacture.

A diffusion model is a computer program that generates images by gradually working out a clear motif from pure image noise. It powers well-known image generators such as Stable Diffusion, Midjourney, or DALL·E.

A diffusion model is an AI technique that generates images by gradually working out a clean subject from pure image noise. It powers well-known image generators such as Stable Diffusion, Midjourney, and DALL·E, as well as newer video and audio generators.

The Digital Markets Act is a law of the European Union that imposes fixed rules on the largest internet corporations. It aims to prevent them from exploiting their market power and to secure fair opportunities for smaller providers.

The Digital Services Act is a law of the European Union that regulates how online platforms must handle illegal and harmful content. The larger a platform is, the stricter its obligations – violations can result in fines of up to six percent of global annual turnover.

A digital worker is a software program that takes on a fixed role within a company and independently handles recurring tasks there. Unlike a simple helper program, it is given its own area of responsibility, access to company systems, and a person who oversees its work.

Digital biology refers to the approach of treating life processes as readable and writable information and studying them on the computer. Instead of only experimenting in the lab, programs predict how molecules, cells, or drugs will behave.

A digital signature is a kind of computationally generated stamp on a file or message. It proves who created the file and that not a single character has been changed since.

A digital avatar is an artificial figure that represents a human on screen – as an image, cartoon character, or deceptively realistic-looking video person. Modern avatars are controlled by software, speak, move their mouths, and thereby seem like a real counterpart.

A digital twin is a computer model of a real object that is continuously fed with measurement data from reality. This makes it possible to observe, calculate, and try out on screen what is currently happening or could happen to the real object.

Direct Preference Optimization is a method for teaching a language model human preferences: it is shown pairs of responses, with the better one marked. It replaces the older, considerably more elaborate approach involving an additional reward model.

Direct-to-Device refers to mobile communication in which a regular phone connects directly to a satellite without any additional device. The goal is to eliminate dead zones where no cell tower is within range.

DirectX is a collection of programming interfaces from Microsoft through which games and other programs access the graphics card, sound chip, and input devices of a Windows PC. It saves developers from having to adapt their software to every piece of hardware individually.

A disassembler is a tool that translates a finished computer program from numeric code back into readable machine instructions. This makes it possible to understand what a program does, even when the original program text is missing.

Discovery refers to the phase in which a team figures out which problem it should actually solve – before anything gets built. In legal disputes, the same term refers to the obligation to disclose all relevant documents to the opposing party; for AI companies, this often involves millions of files.

Discovery is the phase of a US civil lawsuit in which both sides must disclose to each other which documents, emails, and data they hold relevant to the dispute. For tech companies, this means internal chats, emails, and model data can end up in court and become public.

A discovery layer is the intermediate layer in a digital offering that suggests the right options to users from a vast pool of content or services. Search bars, recommendation lists, and homepage tiles are all part of it — and whoever controls this layer decides what gets found at all.

A Discovery Sprint is a short, tightly time-boxed block of work in which a team clarifies whether a product idea is even worth pursuing. Instead of coding right away, the idea gets tested – often within a single week.

Discovery vs. Delivery describes the separation of two distinct ways of working in product development: figuring out what should be built, and then building it cleanly. Both tasks require different methods, different error cultures, and usually different timelines as well.

The discount window is an emergency lending facility offered by the central bank to commercial banks: banks can borrow money there on a short-term basis when they cannot obtain it anywhere else. It functions as a safety net for the banking system and becomes especially important during crises.

Discrete geometry is a branch of mathematics concerned with finitely many, clearly countable geometric objects – such as points, line segments, or polygons. It provides the foundations for computer graphics, robotics, and many methods computers use to process spatial data.

The Discrete Cosine Transform is a computational method that decomposes an image or a sound into overlapping waves. It is the mathematical foundation of JPEG images, MP3 music, and almost every video on the internet.

Discrete tokens are countable building blocks from a fixed list into which a computer breaks down text, images, or sounds before an AI model computes with them. Instead of arbitrarily fine intermediate values, there are only whole units – much like the letters of an alphabet.

Distillation is a technique in which a large, expensive AI program teaches a smaller one to give similar answers. The small model then runs faster and cheaper, often on a phone.

The DiT architecture is a blueprint for programs that generate images and videos. It combines the step-by-step process of extracting an image from image noise with the computational technique that also underlies modern language programs.

DLRM stands for Deep Learning Recommendation Model and refers to a model architecture published by Facebook (now Meta) that predicts which content or products a particular user will be interested in. Such recommendation models determine what appears at the top of feeds and shops, and they are among the most economically important AI systems there are.

A DMA decision is a formal decision by the EU Commission based on the Digital Markets Act, an EU law targeting very large digital corporations. Through it, the Commission determines who falls under the law, whether a company has violated it, and what penalty or obligation follows.

The DNA supply chain is the path from an ordered genetic blueprint to the finished piece of DNA a lab receives in the mail. It is considered the most important control point for preventing AI-designed dangerous genetic material from actually being manufactured.

DNA synthesis is the artificial production of genetic material in the lab, building block by building block, without copying a living organism as a template. It is the technical foundation for vaccines, genetically engineered microbes, and data storage in molecules – and a field in which AI software is increasingly designing the blueprints.

DNS prefetching means that a browser looks up a website's address data before the user even clicks. This makes page loading start a few milliseconds to tenths of a second faster.

A DNS resolver is the service that translates internet addresses like "tagesschau.de" into the numeric address computers need to connect. It usually runs at the internet provider or with providers like Google and Cloudflare and is used with almost every click on the web.

Docker is a program that packages software together with everything it needs to run into a self-contained package. This package runs identically on any computer – whether on a developer's laptop or on a server in a data center.

A Docker container is a program together with everything it needs to run, packaged into a self-contained unit. This makes it behave the same way on every computer – whether on a laptop or on a server in a data center.

A docstring is a short descriptive text placed directly within program code that explains what a particular section of that code does. Unlike an ordinary side note, it remains in the finished program and can be read out by tools and AI systems.

The DOM is the structure into which a browser translates a loaded web page in memory: a tree representation made up of headings, paragraphs, images, and buttons. Programs can read and modify this tree, which allows a page to change without being reloaded.

Domain squatting refers to securing internet addresses that match someone else's name or brand, usually with the intent to resell them later at a high price or to intercept visitors. Because addresses are assigned on a first-come, first-served basis, the practice is technically simple and often legally problematic.

Domain data is data from a single specialized field, such as medicine, law, or mechanical engineering. For computer programs that learn from examples, it is often more valuable than general texts from the internet.

DOS is a family of simple operating systems that ran on PCs from the 1980s onward and was operated entirely through typed commands. The best-known representative is Microsoft's MS-DOS, the predecessor of Windows.

Double Precision refers to a number format in which a computer stores a decimal number using 64 bits, allowing it to accurately represent a very large number of decimal places. It is considered the standard in science and simulation, but plays hardly any role in modern AI models.

In the doubleshot process, a plastic part is manufactured from two separately injected plastic portions that fuse firmly together. It is best known from keyboard keycaps, whose letters consist of a second plastic and therefore cannot wear off.

A draft model is a small, fast AI model that produces initial text suggestions, which a larger model then merely checks. This interplay speeds up chatbots without making the answers worse.

DRAM is the most common type of working memory in computers and smartphones. It stores data only as long as power is flowing, but in exchange it is cheap and can be manufactured in huge quantities.

A drone swarm is a large group of unmanned aerial vehicles that coordinate with one another and thereby act as a single unit. Instead of steering each drone individually, a human only specifies the goal, and the software distributes the tasks itself.

Dual-use refers to technology that can be used both peacefully and militarily or harmfully. With AI, this is especially tricky because the same program can design medicines or suggest toxins.

A Durable Object is a small piece of software on the internet that has its own identity and its own persistent storage, of which there is always only ever a single running copy worldwide. This design comes from the provider Cloudflare and solves a classic problem: multiple servers wanting to modify the same data at the same time.

Durable Streams are continuous data streams that are stored permanently instead of disappearing after delivery. This allows programs to read a stream again later, resume from an old position, or cleanly recover after a crash.

Durable Workflows are program sequences that persistently store their state and, after a crash, resume exactly where they left off. They are used primarily for long, multi-step processes, such as orders, payments, or AI agents.

Dynamic Workers are computing processes that a provider automatically starts when work arises and shuts down again once it's done. This principle ensures that AI services don't collapse under sudden surges and still cost little during quiet hours.

A developer API is a fixed access point through which one program can use the functions of another program. For AI providers, it is the usual way to embed a language model into your own app without having to run it yourself.

A degree of freedom is a quantity of a system that can change independently of all others. The term describes how many directions of movement a robotic arm has or how many adjustable parameters a statistical model possesses.

A data center is a building full of computers that compute and store data for others around the clock. Without these facilities, there would be no streaming, no online banking, and no AI services like ChatGPT.

Data center capacity describes how much computing technology can be operated in large server halls. It is usually measured not in the number of computers, but in megawatts of electrical power.

Distributional geometry describes the spatial shape taken on by the data an AI system works with. The term helps explain why models perform reliably in some areas and suddenly fail in others.

An Employment Contract Act is a draft law intended to bundle the scattered rules on employment contracts into a single code of law. In Germany, no such law exists to date – employment contract law is spread across many different statutes and court rulings.

E-commerce is the trade of goods and services over the internet. This includes large marketplaces like Amazon, a brand's own online shop, and sales via social networks.

E-Ink is a display technology in which tiny black and white color particles inside a film are made visible. The screen does not emit its own light, looks like printed paper, and only needs power when the image changes.

Early Access means that a product is already sold or released even though it is still unfinished. Buyers get early access and, in return, provide feedback that the manufacturer uses to keep developing it.

An earn-out is a portion of the purchase price in a company sale that is paid out later – and only if the company achieves previously agreed targets. Buyers and sellers use it when they cannot agree on the value of the company.

EBITDA is a metric derived from a company's financial statements. It shows the profit before interest, taxes, and the loss in value of machinery and buildings are deducted.

The Economic Alignment Score is a metric intended to measure how well an AI system's behavior aligns with the economic interests of its users or operators. It is discussed primarily where AI makes autonomous decisions with financial consequences.

Edge refers to processing data directly where it originates – on the phone, in the car, in the camera – instead of in a distant data center. The advantages are short response times, less data transmission, and more control over personal data.

Edge AI means that a device such as a phone, a camera, or a car performs its AI computations itself, instead of sending data to a distant data center. This makes responses faster, saves data transmission, and keeps private data on the device.

An edge case is a rare, extreme, or unexpected situation in which a program or an AI system reacts differently than in the normal case. It is precisely these exceptions that often determine whether a technology is reliable in real-world use.

Edge computing means that data is processed right where it is generated – that is, on the device itself rather than in a far-away data center. This saves latency and transmission costs, and often protects privacy as well.

Edge Deployment means that an AI program runs directly on the user's device – on the phone, in the car, or in the camera – instead of in a remote data center. This makes responses faster and keeps personal data on the device.

Editor integration means that an AI assistant sits directly inside the program used to write software. This lets the assistant see the text that is open and insert suggestions right at the spot where work is currently happening.

EdTech is the short form of "Educational Technology" and refers to digital technology that helps with learning and teaching. This includes learning apps, online courses, digital school platforms, and increasingly AI-powered tutoring programs as well.

The Effective Output Score is a metric that measures what proportion of an AI system's responses is actually usable. It does not assess the volume of outputs, but rather their usable yield.

An effort setting is a dial that lets you tell an AI text system how much computational effort it should put into a single answer. Low levels respond quickly and cheaply, high levels compute longer and solve difficult tasks more reliably.

Effort control is a dial that lets you specify how much computational effort an AI system should put into a single response. Low effort means fast and cheap answers, high effort means longer internal reasoning at higher cost.

An egosystem is a digital environment in which a single provider controls all the important parts and binds users and partners to itself. The term is a critical play on words: instead of an open ecosystem with many equal participants, everything revolves around the interests of one company.

Egress filtering means that a network controls what data it sends out — not just what comes in. It is meant to prevent stolen data or attacker commands from leaving the corporate network unnoticed.

Egress monitoring is the surveillance of all data leaving a computer network. Companies use it to detect theft of corporate data and covert data leakage caused by malware or AI systems.

Egress traffic is all data that flows out of a provider's data center — to the user or to another provider. Cloud companies charge money for this outflow, which makes egress an important cost factor and competitive issue.

Electron is a toolkit that lets developers build programs for Windows, macOS, and Linux using the same techniques as websites. Every Electron application brings along a complete browser to do this, which greatly speeds up development but costs a lot of memory.

Electroencephalography, or EEG for short, uses electrodes on the scalp to measure the electrical activity of the brain. It is painless, very fast, and provides the data basis for sleep labs, epilepsy diagnostics, and many brain-computer interfaces.

Elliptic Curve Cryptography is a method for encrypting data and generating digital signatures. It computes with points on a special mathematical curve and requires much shorter keys than older methods.

The Elo rating is a score derived from many head-to-head comparisons between two opponents: whoever wins takes points away from the loser. Originally invented for chess, it is now used to rank AI systems based on user judgments.

The Elo rating is a score that calculates a participant's playing strength based on wins and losses in duels. Originally invented for chess, it is now also used to rank AI models according to which answer humans prefer in direct comparisons.

An embedding is a translation of words, images, or other data into a long series of numbers. Similar content is assigned similar number sequences, allowing computers to compare meaning computationally.

Embodied AI refers to artificial intelligence that controls a physical body – such as a robotic arm, a car, or a walking robot. Such systems must not only think, but also perceive and act, and their mistakes have consequences in the real world.

Embodied Intelligence refers to the idea that genuine intelligence requires a body that acts in the world and constantly receives feedback while doing so. Rather than merely processing text or images, such a system learns through grasping, moving, and trying things out.

Emergent behavior refers to capabilities that an AI system suddenly displays even though no one specifically trained it to have them. They usually only appear once a system becomes large enough, and they can hardly be predicted from the individual parts.

Emergence refers to abilities that an AI system suddenly displays even though no one specifically taught it to do so. Typically, they only appear once a system exceeds a certain size.

Emotional AI refers to computer programs designed to infer feelings from faces, voices, or text. The technology is embedded in call center software, cars, and advertising analytics — though its scientific foundation is highly disputed.

Emotion recognition refers to computer programs that attempt to infer feelings from faces, voices, or texts. The technology is used in advertising research, call centers, and cars, is scientifically controversial, and is partially banned in the EU.

Emulation means that a computer fully replicates the behavior of another device so that foreign programs can run on it. This way, a modern PC can, for example, run software that was written for an old game console.

An emulator is a program that mimics another computer or device, allowing its software to run on a foreign machine. This makes it possible to run old game consoles, mobile phones, or entire computer architectures on a normal PC.

An end-to-end neural network is a learning computer program that solves everything in one piece, going directly from the raw input data to the finished result. Instead of many separately built intermediate steps, there is only one system that is optimized as a whole during training.

An end-to-end test checks a finished program the way a real user would use it: from the first input to the visible result. It shows whether all the parts involved work together, not just whether each part works on its own.

End-to-end encryption means: a message is made unreadable on the sender's device and only becomes readable again on the recipient's device. Even the provider of the service through whose servers the message travels cannot read the content.

An endpoint is a fixed internet address to which a program sends a request and from which it receives a response back. In AI services, the endpoint is the point at which a model can be reached on remote servers.

An engineering sample is an early test unit of a computer chip that the manufacturer gives to partners and testing labs before it goes on sale. It looks like the later product, but often runs slower and officially may not be resold.

An ensemble is a group of several computer programs that solve the same task and then combine their results into a single joint answer. Because the errors of the individual programs often cancel each other out, the group is usually more reliable than any single member alone.

An ensemble method combines several prediction models into a single joint result instead of relying on just one. Because the errors of the individual models partly cancel each other out, the group is usually more accurate than any single member on its own.

An Enterprise Beta is a testing phase in which a software maker first gives an as-yet unfinished product to select corporate customers. These customers are allowed to use it in their real day-to-day work, report bugs back, and receive a head start in return.

Enterprise search is a search function that does not scan the open internet, but rather the documents, emails, and databases of a single company. In doing so, it must precisely observe who is even allowed to see which content.

Enterprise Value to Free Cash Flow is a metric that compares the total price of a company with the cash that is actually left over at the end of a year. It shows in a single number how many years' worth of cash earnings you are paying for a company.

An entitlement is the recorded permission to use a particular feature or piece of content. Before every use, software checks whether such a permission is stored for the respective user.

An entity is a uniquely nameable thing in a text or a database – such as a person, a company, a place or a date. Computers recognize such entities in order to turn sentences into usable information.

The Entity List is a blacklist maintained by the US government. Companies and organizations on it may only obtain certain US products and technologies with a special license – for the chip and AI industry, one of the harshest political weapons.

An environment variable is a named value that the operating system holds ready outside of a program, which the program can query at startup. This allows settings and passwords to be changed without touching the program code itself.

An equalizer is a tool that makes certain pitch ranges of a sound louder or quieter. It's built into music apps, headphones, car radios, and almost every professional audio recording.

Equity refers to the ownership stake in a company – that is, the portion that belongs to oneself and is not based on debt. Whoever holds equity owns a piece of the company and benefits when its value rises.

An equity deal is an agreement in which someone puts money into a company and receives shares in the company in return. Unlike a loan, the money is not repaid – instead, the investor becomes a co-owner and only profits if the company increases in value.

An Erdős problem is an unsolved mathematical question posed by the Hungarian mathematician Paul Erdős. Such questions are stated briefly but are extremely hard to answer – and today serve as a stress test for AI systems.

An event log is a list that a computer program continuously writes to, recording what has just happened – with a timestamp and a brief description. Such logs are the most important basis for finding bugs, proving attacks, and monitoring the operation of large systems.

Event logging means that a computer program continuously records what it does: who logged in, which file was opened, where an error occurred. These records are the most important basis for tracking down disruptions, detecting attacks, and later proving what actually happened.

An event stream is a continuous sequence of messages about what has just happened – such as every card payment or every click. Instead of collecting data first and analyzing it later, systems here process it directly at the moment it arrives.

ERP stands for Enterprise Resource Planning and refers to a large piece of software that brings together all of a company's key processes on a shared data foundation. Purchasing, warehousing, production, accounting, and HR all work with the same figures instead of separate lists.

An ETL pipeline is an automatic chain of work steps that collects data from various sources, brings it into a uniform form, and writes it into a central storage location. It is the invisible groundwork behind almost every statistic, every dashboard, and every AI training run.

The EU AI Act is a law of the European Union that regulates the use of artificial intelligence. It sorts applications by their risk: the greater the potential harm to people, the stricter the obligations for the provider.

EUV lithography is the process used to expose the finest structures of modern computer chips onto silicon wafers. It uses extremely short-wavelength light and comes from a single manufacturer worldwide, the Dutch company ASML.

An eval is a standardized test used to check how well an AI system solves specific tasks. Instead of relying on impressions, one measures the success rate across many tasks with known solutions.

An event stream is a continuous chain of messages about things that have just happened – such as every click in an online shop or every executed stock order. Programs read this chain in the order in which things occurred and react immediately, instead of waiting for a later batch retrieval.

Event-based sensors are cameras that don't capture complete images, but instead only report where brightness has just changed. This lets them react extremely quickly, use little power, and deliver a stream of individual data points instead of images.

Evergreening refers to the strategy of artificially extending an expiring protection right through minor changes to a product. The term originally comes from the pharmaceutical industry, but today it also appears in relation to software, data rights, and AI models.

An Evidence Ledger is an ongoing, tamper-proof record that documents what a decision or a statement is based on. In AI systems it serves to later prove which data, sources, and intermediate steps led to a particular result.

An exaflop is a quintillion floating-point calculations per second – a measure of the speed of very large computers. Machines that reach this speed count as top-tier supercomputers and are built for climate models, materials research, and AI training.

Exascale refers to supercomputers capable of performing at least one quintillion calculations per second — a one followed by 18 zeros. Since 2022, such machines have truly existed; they cost several hundred million euros and consume as much electricity as a small city.

An Excel plugin is an add-on program that is installed into the spreadsheet application Excel to extend it with new functions. In recent years, many such extensions have used AI, for example to suggest formulas or evaluate text within spreadsheets.

An execution loop is the repeated cycle by which an AI program works through a task step by step: think, act, check the result, start over. It is the core of modern AI assistants that don't just answer but actually do things themselves.

An Executive Order is a written directive from the US President to the agencies of the state. It takes effect immediately and without approval from Congress, but can be revoked just as quickly by a successor.

ExecuTorch is an open-source tool developed by Meta that allows trained AI models to run directly on phones, glasses, and small devices. It translates models from the PyTorch development environment into a very lightweight form that doesn't require a connection to a data center.

Exfiltration refers to the covert smuggling of data out of a network or system by attackers. It is the step at which a break-in turns into actual data theft.

An exfiltration vector is the concrete route by which data leaves a system unnoticed and reaches the outside. In IT security and with AI systems, the term describes where an attacker can gain access to confidential information.

Expedited discovery is an order by a US court that allows a party to demand documents, emails, or witness testimony from the opposing side ahead of the normal schedule. In tech lawsuits involving stolen trade secrets or AI models, this is often the first hard blow.

Experiment design is the art of planning a test so that its outcome actually proves something. You determine in advance what is being compared, who is being compared, and how success is measured — so that in the end, chance or a flaw in reasoning doesn't deliver the answer.

An exploit is a piece of code or a specific input that deliberately takes advantage of a flaw in a program to make it do something it shouldn't. A harmless programming error thus becomes a gateway for attackers.

Exploit development refers to the work of turning a programming flaw in software into a functioning attack tool. It is both the craft of criminals and a recognized part of professional IT security testing.

ExploitGym is a training and testing environment in which programs are meant to learn to find and exploit security vulnerabilities in software. To do this, it provides deliberately insecure practice programs and automatically evaluates whether an attack succeeded.

An exponential generating function packs an entire infinite sequence of numbers into a single mathematical expression, dividing each term of the sequence by the factorial of its position. This is especially useful when you want to count the number of ways distinguishable objects can be arranged or divided up.

An export is the process of writing data out of a program into a file that other programs can also read. In the AI world, the term also refers to converting a fully trained model into an exchangeable format.

Export control means that a state legally determines which goods and technologies may be sold abroad at all. In the AI industry, this primarily concerns the fastest computing chips, whose sale to China and other countries has been heavily restricted by the United States.

Export controls are government rules that restrict or prohibit the sale of certain goods and technologies abroad. In the AI industry, they primarily affect the powerful computing chips used to train large AI systems.

Export controls for AI chips are government bans and licensing requirements on selling particularly powerful computing chips abroad. The United States in particular uses them to make it harder for China to access hardware for powerful AI systems.

An exposure score is a metric indicating how strongly an occupation, a company, or a task could be affected by a particular development – usually artificial intelligence. It does not state that something will disappear, only how large the point of contact is.

An exposure score is a metric that indicates how strongly something is affected by a given development – for example, how strongly a profession is affected by artificial intelligence, or a company by a particular market. The value only measures the degree of contact with the topic, not automatically a harm or benefit.

Extended Reasoning refers to an operating mode of language programs like ChatGPT in which the program computes for longer and notes down intermediate steps before answering. This increases the hit rate for difficult tasks, but the answer takes longer and costs more.

Extensions are small add-on programs that make an iPhone app available in other parts of the system — for example as a keyboard, as a widget on the home screen, or as an entry in the share menu. They run separately from the main app and are only allowed to handle narrowly defined tasks.

Externalities are costs or benefits that fall on uninvolved third parties, without anyone paying or being paid for them. In the tech debate, they show up wherever data centers, data collection, or automated systems have consequences for people who are neither customers nor providers.

An extreme event is a very rare occurrence with a very large impact, such as a stock market crash or a once-in-a-century flood. Because such cases barely appear in the data, statistics and AI models systematically underestimate them.

An environment variable is a named value that an operating system passes to a running program — for example, a password, a storage location, or a switch. This allows settings to be changed without touching the program code itself.

A foundation model is a very large AI program that has been trained generally on huge amounts of data and can subsequently be reused for many different tasks. Instead of starting from scratch for every application, developers build their products on top of such a ready-made model.

The F1 score is a metric between 0 and 1 that measures how well a computer program performs on a yes/no decision. It combines two types of errors into a single number: false alarms and missed cases.

Façading refers to the practice of presenting a product or company outwardly as its own AI development, even though internally foreign technology or even human labor is actually handling the task. The term is mainly used in reports about overvalued start-ups and misleading marketing.

Factory-to-Consumer refers to a business model in which a factory sells its products directly to end customers, bypassing wholesalers and retail chains. Because the intermediate stages are eliminated, prices are often lower — but in exchange the manufacturer has to take on marketing, shipping, and customer service itself.

Failover refers to the automatic switchover to a backup system when a computer or service fails. The goal is for users to notice as little as possible of the disruption.

Failure Recovery refers to all procedures by which a computer system returns to a working state after a failure. The goal is to preserve data and resume operation as quickly as possible.

Fair use is a rule in US law that allows other people's texts, images, or music to be used in certain cases without permission. When it comes to training AI systems, it is fiercely disputed whether this exception also applies to the mass mining of other people's works.

The fair value is the price an asset would fetch today in a sale between two independent, well-informed parties. It is the central valuation metric in modern balance sheets and there often replaces the historical purchase price.

Fair value accounting records an asset on the balance sheet at the price it would fetch on the market today. It shows the current value instead of the original purchase price, making balance sheets more volatile as a result.

Fairness Collapse describes the case where a computer program suddenly delivers markedly worse results for a particular group of people than for everyone else. The program appears reliable overall, but specifically fails for a minority.

Fairness testing is the systematic examination of whether a computer program treats certain groups of people worse than others. It compares outcomes across different groups — such as men and women — using fixed metrics and checks whether the difference is too large.

Fact-checking means systematically comparing a claim against reliable sources and publicly assessing whether it is true. Because AI programs write texts that sound convincing but can be false, checking statements has become its own field of work.

Faculty boundaries are the organizational dividing lines between the major disciplines of a university, such as between computer science, medicine, and law. In AI research, they are seen as an obstacle because many important questions can only be answered jointly by several disciplines.

A fallback chain is a fixed sequence of alternative options: if the first attempt fails, the system automatically tries the second, then the third. In AI products, it ensures that a request still gets answered even when the preferred model isn't responding.

A false positive is a false alarm: a system reports something that isn't actually there. The smoke detector beeps even though there's no fire — and exactly this kind of false report is one of the costliest error types in automated testing systems.

Falsification means disproving a claim by producing a counterexample. The idea comes from philosophy of science and today also forms the basis for how AI systems are tested and their errors uncovered.

A fan-out query is a single search query that is automatically fanned out in the background into many smaller sub-queries. Search engines and AI assistants use this to gather answers from multiple sources and merge them into a single result.

FastText is a tool developed by Facebook that converts words into sequences of numbers and automatically classifies texts. In doing so, it breaks words down into short letter sequences, allowing it to understand even words it has never seen before.

The FCC is the US agency that regulates radio, telephone, cable, and internet in the United States. It allocates radio frequencies, approves devices, and sets rules for mobile, satellite, and internet providers.

FD-Bench is a standardized test collection that researchers use to compare how well computer programs predict the behavior of liquids and gases. It primarily evaluates learning-based models designed to replace or accelerate classical flow calculations.

A feature flag is a built-in switch in software that lets a new function be turned on or off without having to ship a new release of the program. This way, new features can first be tested on small groups of users and switched off immediately in an emergency.

A feature branch is a separated working copy of a software project in which exactly one new feature is developed. Only once it is finished and reviewed is the work merged back with the main version of the project.

The Fediverse is an association of many independent social networks that can send each other messages. Anyone with an account on one of these services can follow people on all the others – similar to email.

A feed is the endless list of posts that an app like Instagram, TikTok, or X displays. Which content ends up at the top is usually decided not by the time it was posted, but by a program that sorts the order based on expected attention.

A feedback loop is a cycle in which a system's output becomes its input again, thereby influencing the next output. In the world of AI, this ensures that systems learn from their own effects — for better or worse.

A feedback loop occurs when the output of a system becomes its input again and influences the next result. In AI, this is a central principle: feedback from humans or from ongoing operation improves a system — or amplifies its errors.

Fine-tuning means further training an already fully trained AI model with additional, specially selected examples. This turns an all-rounder into a specialist for a particular task, without having to build the model from scratch.

The Fel Conjecture is an unproven claim that appears in discussions about AI and mathematics. However, there is no reliable information available regarding this name, which is why this entry primarily explains what a conjecture actually is and how to handle unclear technical terms.

Felony Bench is a test used to check whether AI language programs assist with clearly criminal requests or refuse them. Such test collections serve to make a system's safety boundaries measurable and comparable.

Few-shot learning means that a computer program correctly solves a new task after seeing only a handful of examples. With today's language programs, it's often enough to simply write a few examples directly into the prompt.

Filament is the coiled plastic wire that the most common type of 3D printer uses as material. The wire is melted in the print head and applied layer by layer to build up a workpiece.

Fine-tuning means taking an already fully trained AI model and training it further with additional, specially selected examples. This adapts a general model to a specific task, language, or company style without building it from scratch.

Fingerprinting is a technique in which websites calculate a unique identification value from many small technical characteristics of a device. This allows users to be recognized again without anything being stored on the device.

Fintech refers to financial services that run primarily through software and apps rather than through bank branches and paperwork. The term stands both for the technology itself and for the young companies using it to compete with banks and insurers.

Firmware is the small control program permanently embedded in a device that gets its electronics running in the first place. Without firmware, even a functioning washing machine, hard drive, or drone remains a lifeless piece of technology.

The First Amendment is the first amendment to the US Constitution. It prohibits the state from restricting freedom of speech, the press, and religion, and is therefore the reason why US platforms are regulated differently on content than European ones.

First Loss refers to the portion of a financing that absorbs losses first when they occur. Whoever holds this position thereby protects all other lenders and demands a high return in exchange.

The first-loss tranche is the portion of a financing that is liable for losses first. Whoever holds it loses their money first — and in return is promised the highest potential return.

First-party data is information that a company collects directly from its own users or customers – for example when they make a purchase, sign up, or click around on the company's own website. It is considered particularly reliable and legally unproblematic, because no third-party provider stands in between.

Fitts' Law is a rule of thumb for how long a person needs to hit a target with a hand, mouse, or finger. The farther away and smaller the target, the longer it takes — and the more often the click misses.

A flash crash is an extremely rapid stock market plunge that largely reverses itself within minutes. It is usually triggered not by real news, but by automated trading programs that drive each other into a spiral.

A flat-rate model is a pricing model in which you pay a fixed amount per month and are allowed to use an offering without any quantity limit. With AI services, this is risky because every request costs the provider real money — heavy users can make such a subscription unprofitable.

Floating-point precision indicates how many digits a computer uses to store decimal numbers. In AI, it determines how much memory a model needs and how fast it can compute.

FLOPs are individual computational steps involving decimal numbers, such as a multiplication or an addition. They serve as a unit of measurement for how much computational work an AI model consumes and how fast a computer chip operates.

Functional magnetic resonance imaging (fMRI, German fMRT) is a method that reveals which regions of the brain are currently particularly active. It does not measure the nerve cells themselves, but rather the oxygen content in the blood — thereby providing data that is now also analyzed by AI systems.

A fork is the splitting off of a software project: someone takes the existing program code and continues developing it independently in a different direction. In the context of cryptocurrencies, the term refers to the splitting of a network into two versions with different rules.

Form 8-K is a mandatory report that publicly traded U.S. companies must file with the U.S. securities regulator on short notice whenever something significant happens. Unlike quarterly or annual reports, it doesn't follow a calendar but rather the event itself – usually within four business days.

Formal verification is a mathematical proof that a program or a chip behaves exactly as specified. Instead of hunting for errors through trial and error, they are computationally ruled out for all possible cases.

Formalization means translating a statement or task from everyday language into a strictly unambiguous notation, such as a mathematical formula or a set of rules that a computer can check. It is the prerequisite for machines to be able to compute a problem at all.

Format conversion means translating the same content from one storage form into another, for instance an image from HEIC to JPEG or a spreadsheet from Excel to CSV. It is one of the most common tasks in data processing, because different programs expect different forms.

A formatter is a program that automatically brings program text into a uniform form according to fixed rules – such as indentation, line breaks, and spacing. It does not change the content of the program, only its appearance.

A forward multiple relates a company's current market value to an expected future figure, usually next year's earnings or revenue. It is the most common metric for judging whether a stock is expensively or cheaply valued – especially for fast-growing tech and AI companies.

The forward pass is the computational path an input takes through an AI model from start to finish, until a result comes out at the end. It is the smallest computational unit in AI systems: every answer and every training example begins with a forward pass.

A foundation model is a very large AI model trained generally on vast amounts of text, images, or code, which can then be reused for many different tasks. It is the shared basis behind products like chatbots, translators, or image generators.

A foundation model is a very large AI model that has been pretrained on massive amounts of data for general purposes and can subsequently be reused for many different tasks. It is a kind of basic infrastructure on which companies build their concrete products instead of starting from scratch every time.

Founder Mode refers to a leadership style in which a company's founder deeply involves themselves in details rather than handing tasks entirely over to managers. The term originates from a 2024 essay by investor Paul Graham and has been hotly debated in the tech industry ever since.

FP4 is an extremely sparse number format that stores every number in an AI model using just four bits. This means models need significantly less memory and compute faster — but at the cost of accuracy.

FP8 is a number format that stores every number using only eight bits – that is, eight yes/no switches. AI models compute with it significantly faster and more efficiently than with the previously common 16 or 32 bits, but lose some precision.

FP8 precision means that an AI model stores its numbers in just eight bits instead of sixteen or thirty-two. This halves memory requirements and computational effort compared to the previous standard, but costs accuracy and requires technical safeguards.

An FPV drone is a small aircraft equipped with a camera whose image is transmitted live to a pair of goggles worn by the pilot. Instead of steering from an outside view, the pilot sees the world as if sitting inside the drone itself.

A frame is a single still image from a sequence of images, such as from a video or a computer game. Many such individual images shown one after another produce the impression of smooth motion for the eye; how many of these are shown per second is called the frame rate.

Frame interpolation generates artificial intermediate frames in a video so that motion appears smoother. From 24 individual frames per second, it can produce, for example, 60 or 120 – without the camera ever having captured them.

A framework is a prefabricated basic structure for programs that takes over many standard tasks for developers. They only have to write the parts that make their project unique and insert them into the existing scaffolding.

The free float is the portion of a company's shares that is actually freely traded on the stock exchange. Shares held by major owners such as families, states, or founders are excluded from this figure.

Freemium is a business model in which a basic version of a product is free and only additional features cost money. For AI services like chatbots, it is today the standard way to win over users and later get some of them to pay.

Ku-band to D-band refers to a range of radio frequencies between roughly 12 and 170 gigahertz, used for satellites, microwave links, and fast mobile connections. The higher the frequency, the more data fits through the connection – and the more sensitive it becomes to rain and obstacles.

Frontend deployment refers to the process by which the visible part of a website or app is moved from developers' machines to publicly accessible servers. Only after that do users see the new version in their browser.

Frontier Model refers to the most capable AI systems that exist at any given point in time. The term is not a metric but a classification: these models sit at the leading edge of what is technically feasible.

Frontier research refers to work on the most capable AI systems that exist at all. It is extremely expensive, is conducted by only a handful of large labs, and stands at the center of the debate about the opportunities and risks of artificial intelligence.

A frontier model is an AI system that, at the time of its release, ranks among the most capable in the world. The term originates from the debate about the risks and regulation of such systems.

FTE stands for "Full-Time Equivalent." The metric converts all working hours of a company into a number of notional full-time positions, making workforces comparable.

FUD stands for "Fear, Uncertainty and Doubt." It refers to the deliberate spreading of negative or vague claims to discourage people from a product, a company, or an investment.

Fulfillment refers to everything that happens after an online order: storing goods, picking them, packing them, shipping them, and handling returns. Many retailers outsource this work to specialized providers who operate entire warehouses for this purpose.

Full-body capture refers to techniques used to record the movements of an entire human body and translate them into digital data. This data is then used to drive a game character, a film character, or an avatar in a virtual environment.

Full-duplex refers to a connection in which both sides can send and receive simultaneously — much like a normal phone call, where you can also interrupt each other. In the AI world, the term today mainly refers to voice assistants that listen while they are speaking themselves.

A full-duplex architecture allows a voice system to listen and speak at the same time, rather than strictly taking turns. This makes conversations with voice assistants feel more natural: you can interrupt, and the system reacts immediately.

Full-Stack refers to working across all layers of software: the visible surface and the underlying technology that stores and processes data. A full-stack developer can therefore build an application alone, from start to finish.

Full-Stack AI refers to companies that control all layers of artificial intelligence themselves: from the chips through the data centers to the finished app for customers. The term therefore describes not a technology, but a business strategy.

Full-stack infrastructure means that a company controls every layer of an AI setup itself: from the chips through the data centers to the finished application. The term comes up especially when corporations like Nvidia, Google, or Amazon explain why they don't want to be just one part of the chain.

Function calling means that a language model doesn't just write text, but recognizes when an external program is needed and formulates its call cleanly. The model supplies the name and values for the request, while it is executed outside the model.

A foundation model is a very large AI model that has been trained generally on huge amounts of text, images, or code and can then be adapted for many different tasks. Systems like ChatGPT, Gemini, or Claude are built on such models.

The funding rate is a small payment that buyers and sellers exchange with each other on certain crypto bets, usually every eight hours. It ensures that the price of these bets doesn't drift too far from the actual market price.

A functional representation describes an object not through its stored data, but through a computational rule that delivers the appropriate value on request. So instead of storing a list of finished values, one stores a rule that generates these values.

In function calling, a language model doesn't output a finished answer but instead a structured instruction to use an external program — such as a search, a calculator, or a database. The result of that program then flows back into the response.

Fuzzing is an automated testing technique in which a program is bombarded with huge amounts of random or deliberately malformed input. If it crashes, a vulnerability has been found that attackers could otherwise exploit.

Facial geometry refers to the measurable shape of a face: distances, angles, and curves between points such as the eyes, nose, and chin. Computers convert these measurements into a series of numbers, which can then be used to recognize or distinguish individuals.

The fallacy of composition is a reasoning error: one transfers a property of individual parts onto the whole without checking whether that's valid. Just because every player is strong doesn't mean the team is strong.

Full duplex refers to a connection in which both sides can send and receive simultaneously. A phone call is the classic example; for voice AI, it is the prerequisite for being able to interrupt the computer mid-sentence.

Gross profit is a company's revenue minus the costs that are directly incurred for the products or services sold. It shows how much money remains from sales before deducting rent, administrative salaries, research, interest, and taxes.

The gross profit margin shows what percentage of revenue a company keeps after subtracting the direct costs of the product sold. It is one of the most important metrics for comparing software, hardware, and AI business models.

The gross margin shows how much of a company's revenue remains after subtracting the direct costs of the goods or services sold. It is expressed as a percentage and serves as a quick indicator of how profitable a business model is at its core.

Gross reach counts all contacts generated by an advertising campaign — including multiple contacts with the same person. It is therefore almost always larger than the number of people actually reached.

The General Data Protection Regulation is a European Union law that governs how companies and authorities are allowed to handle people's personal data. It has applied in all EU countries since 2018 and allows fines of up to four percent of worldwide annual revenue.

GAAP is the official rulebook that US companies must follow when writing up their numbers in financial reports. It determines when revenue counts as revenue and how costs are recorded, so that companies can be compared with one another.

GAFA is an acronym for the four major American technology corporations Google, Apple, Facebook, and Amazon. In business news, the term often stands in for the enormous market power of these companies.

Gain-of-function research refers to laboratory experiments in which pathogens are deliberately altered so that they acquire new capabilities – for example, spreading more easily. The goal is to better understand future outbreaks; at the same time, this research is considered particularly high-risk and is strictly regulated.

A gate is a controllable switch in an artificial neural network that decides how much of a signal is passed on and how much is held back. Such switches regulate, for example, what a model remembers, what it forgets, and which part of the network even computes for a given input.

A gated capability is a capability of an AI system that is deliberately placed behind an access barrier and is not unlocked for every user. The provider decides who is allowed to use it – for example, based on vetting, contract, or proof of purpose.

The EU uses the term gatekeeper to designate very large digital companies that control access to important online services. Anyone holding this status must comply with special rules so that smaller providers have a fair chance.

The gatekeeper pattern is a software design principle: every request from outside first passes through a single dedicated checkpoint that inspects, filters, and only then forwards it. In AI systems, this doorkeeper usually sits in front of the language model and blocks harmful inputs and outputs.

Gatekeeping refers to the power to decide who or what gets access to a market, a platform, or an audience. In the tech sector, it is usually large corporations that act as gatekeepers, controlling which apps, content, or providers can reach users at all.

Gating refers to a mechanism in AI systems that, like a gate, decides which information flows onward and which is held back. Gating is best known from large language models, where a small additional network activates only a portion of the building blocks per query.

GDPval is a test published by OpenAI that checks how well AI systems perform real work tasks from paid occupations. Instead of answering exam questions, the models have to deliver things that a human would also hand in on the job – such as a legal draft, a spreadsheet, or a construction drawing.

GenAI is short for generative artificial intelligence: computer programs that produce new text, images, sounds, or program code on request. Well-known examples are ChatGPT for text and Midjourney for images.

The Gender Gap in Tech Adoption refers to the measurable difference in how often and how early women and men use new technologies. For AI tools like ChatGPT, the proportion of female users is significantly lower than that of male users in many studies.

General Availability, or GA for short, refers to the point at which a software product is officially finished and can be regularly purchased and deployed by any customer. Before that, the product is usually only available in test versions for selected users.

General Purpose AI refers to AI systems that are not built for a single task, but for a great many different ones. The term is also a legal term: the EU's AI Act attaches special obligations for providers to it.

A general-purpose model is an AI system that isn't built for a single task, but for very many different ones. The same model writes emails, translates texts, explains formulas, and answers follow-up questions about them.

Generative Engine Optimization refers to all measures companies take to ensure their content appears in the answers of AI chatbots and AI search services. It is the successor discipline to classic search engine optimization, which was about getting good placements in Google results lists.

Generative AI refers to computer programs that create new content: texts, images, music, videos, or program code. They do not invent this content out of nothing, but assemble it from patterns they have previously learned from vast amounts of example material.

Generative AI refers to computer programs that create new content: text, images, music, videos, or program code. They don't invent this content out of thin air, but assemble it from patterns they have found in vast amounts of example material.

A Generative Recommender is a recommendation system that predicts a user's next click or purchase by continuing their past usage history like a piece of text. Instead of scoring ready-made suggestions by similarity, it generates the recommendation itself, step by step.

Generative UI refers to user interfaces that are not fixed in code but are assembled by an AI system in the moment of use. Instead of always seeing the same page, each user gets the controls that fit their specific request.

The generator-evaluator architecture splits an AI task between two roles: one part generates proposed solutions, a second part evaluates them and selects the best one. Because checking is often easier than inventing, this division frequently delivers more reliable results than a single pass.

The genome is the complete genetic information of a living organism, stored as a long chain of chemical building blocks in almost every cell. In humans, it comprises about three billion of these building blocks, which today can be read out within a day and analyzed using computers.

Genomics is the science that studies an organism's complete genetic material — that is, all the blueprints of a cell at once instead of individual genes. Because this generates huge amounts of data, it is today just as much a computational discipline as a laboratory one.

A genome model is an AI system trained on the genetic information of living organisms that recognizes patterns in these long sequences of letters. It can, for example, predict which sections of the genome are important or which changes in it could become dangerous.

Genome synthesis means chemically assembling the complete genetic material of an organism in the lab, rather than extracting it from a cell. A digital file on a computer thus becomes a real, functional genome.

Gene therapy is a treatment method in which it is not a drug that relieves symptoms, but rather the blueprint within the body's own cells that is changed. The goal is to fix the cause of a hereditary disease, ideally with a single treatment.

GenUI refers to user interfaces that are not fixed and pre-programmed, but instead assembled by an AI program at the moment of use, tailored to the request. Instead of always seeing the same page, each user gets the buttons, tables, or forms that fit their particular concern.

GEO stands for Generative Engine Optimization and refers to all measures companies take to ensure that chat programs like ChatGPT incorporate their content into answers. It is the successor discipline to classic search engine optimization, which targeted Google results pages.

Geo-risk refers to the danger that political conflicts between states harm companies and investors – for example through wars, trade bans, or sanctions. It is especially noticeable in the tech and AI industry, because chips and raw materials come from only a few, politically sensitive regions.

Geopolitical risks are dangers to companies and markets that arise from the behavior of states – such as wars, trade disputes, or export bans. In the tech industry, they particularly affect the supply chains for computer chips, because their production is concentrated in just a few countries.

getBoundingClientRect is a command in the JavaScript programming language that reads out the size and position of an element on a web page. It returns numbers such as width, height, and the distance to the top of the visible window.

Gigaflops per Watt indicates how many billions of calculations a chip can perform per second when consuming one watt of electricity to do so. The metric therefore doesn't measure raw speed, but rather the energy efficiency of processors and data centers.

We speak of gigawatt scale when a single data center site needs as much electricity as a major city – roughly one billion watts. Such facilities are currently being built for training large AI systems and are turning electricity into the industry's most critical bottleneck.

A gigawatt data center is a computing facility that requires about a billion watts of electrical power to operate — roughly as much as a large city. Such facilities are currently being built to train and run very large AI models.

A gimbal is a movable suspension made of several rotating rings that holds an object steady while everything around it moves. It is best known as a mount that stabilizes cameras on drones or in handheld devices.

A Git worktree is an additional working folder that belongs to the same project repository. This allows working on several versions of a program at the same time without constantly having to switch folders.

A Git diff shows, line by line, what has changed between two versions of a file. It is the standard form in which programmers review, discuss, and share changes.

GKE is a service from Google that lets companies run their programs on rented computers on the internet. Google takes over the task of automatically starting, distributing, and replacing these programs in case of failures.

GKE Agent Sandbox is an offering from Google that lets AI programs run foreign code in a sealed-off area. The code runs there in isolation, so it cannot harm the rest of the operator's system.

GLM is a family of AI language models from the Chinese company Zhipu AI (today Z.ai). The models answer questions and write texts like ChatGPT, but originate from China and are partly freely downloadable.

GLM-5.2 is the name for a language model from the Chinese GLM series, i.e. a computer program that understands and writes text. Such version numbers usually stand for an intermediate step in the news: the same basic idea as the predecessor, but better trained and somewhat stronger in tests.

The Global Scope is the outermost scope of a program: names stored there are visible throughout the entire program. This is convenient, but considered risky because any part of the program can modify such values.

Global Workspace Theory is an explanation of how conscious experience arises in the brain: many unconscious processes run in parallel, and only the winner is allowed to place its content on a kind of shared stage that all others can see. In AI research, it serves as a blueprint idea for systems in which many specialized parts communicate via a shared buffer.

GMV stands for Gross Merchandise Volume and refers to the total value of all goods sold through a trading platform. The figure shows how much business runs through a platform – not how much the platform itself earns.

The Gödel machine is a theoretical computer program that is allowed to rewrite its own program code – but only if it can first prove that the change is an improvement. It was introduced in 2003 by AI researcher Jürgen Schmidhuber and is regarded as a conceptual ideal of self-improving software, not as a practically built system.

Goodhart's Law states: once a measure becomes a target, it ceases to be a good measure. In AI development, it explains why systems often achieve high test scores without actually doing what was really intended.

Google Zero describes a scenario in which websites receive hardly any visitors from Google Search anymore, because Google displays the answers directly itself. For publishers, advice sites, and online shops, this means a business model in jeopardy.

The GPL is a license text that allows anyone to use, modify, and redistribute a computer program — on the condition that redistributions come with the same freedoms. It is the best-known license for freely available software and underlies, among other things, the Linux operating system kernel.

The GPLv3 is a license text from 2007 that sets out the conditions under which a computer program, together with its source code, may be redistributed. Its core rule: anyone who modifies and distributes the program must also disclose the modified source code and pass on the same rights.

GPQA-Diamond is a collection of 198 particularly hard science exam questions used to test how well an AI program can reason. The questions are designed so that they can barely be solved even with an internet search.

GPT-5 is OpenAI's flagship model released in August 2025, which works behind the scenes for ChatGPT. It combines fast responses and longer reasoning in one system and decides itself which mode suits a question.

GPT-5.6 is an intermediate version from OpenAI's GPT model series, a family of programs that write texts and answer questions. The digit after the dot indicates that this is an improvement of an existing generation, not a complete fresh start.

gpt-oss-20b is an AI language model from OpenAI that the company released for free download in August 2025. It is small enough to run on a well-equipped laptop and doesn't need a connection to a remote data center to do so.

A GPU is a specialized chip that carries out a very large number of simple computational steps simultaneously. Originally built for video games, it is today the most important component for training and running AI systems.

A GPU cluster is a large collection of specialized computing chips connected by high-speed cabling that work together on a single task. Such installations are the machines on which modern AI models are trained and operated.

A GPU designer is a company that designs graphics and AI computing chips but usually does not manufacture them itself. The best-known are Nvidia, AMD, and Intel; their designs are fabricated by contract manufacturers such as TSMC.

A GPU kernel is a small program that a graphics card runs thousands of times simultaneously. Almost all AI computations consist of such kernels, and how well they are written determines speed and power costs.

GPU rendering refers to computing images using a computer's graphics card instead of its main processor. Because a graphics card carries out thousands of simple computational steps simultaneously, images are often produced many times faster this way.

GR4AD is not a generally established term in the AI and technology field; there is no recognized technical term with a clear definition for it. Anyone who comes across this abbreviation should check whether it is a company-specific product name, a project identifier, or a typo.

Grace CPU is a main processor made by chipmaker Nvidia, built specifically for data centers handling AI workloads. It is usually paired directly with Nvidia's graphics chips and feeds them with data at high speed.

Gradient Boosting is a machine learning method in which many simple prediction models are built one after another. Each new model focuses on the errors of the previous ones, so that in the end a highly accurate overall prediction emerges.

The graphics pipeline is the fixed sequence of computational steps a computer uses to turn a three-dimensional scene description into a finished image on a monitor. It runs many times per second in games and 3D applications and is the core task of every graphics card.

The graph view is a representation in which individual objects are shown as points and their relationships as lines between them. It makes visible what remains hidden in a list or table: how things are connected to one another.

Graviton is a series of processors that Amazon designs itself and uses exclusively in its own data centers. The chips are meant to make rented cloud servers cheaper and more power-efficient than comparable processors from Intel or AMD.

Graviton processors are computing chips that the cloud provider Amazon Web Services develops itself and uses exclusively in its own data centers. They are based on ARM architecture and are meant to do the same work using less power and at lower cost than classic server chips from Intel or AMD.

grep is a search tool for text files: it scans line by line and outputs every line that matches a search pattern. On servers and in software development, it is the standard means of quickly finding relevant spots in huge log files.

Grok is a chat program from Elon Musk's company xAI that answers questions in everyday language and is built directly into the social network X. It is known above all for its deliberately casual tone and for its access to current posts on X.

Ground truth is the answer established as correct, against which the output of a computer program is measured. Without such verified reference data, it is impossible to train a system or judge how well it truly performs.

The ground-truth function is the imagined perfect relationship between input and correct output that a learning computer program is meant to replicate. It is never fully known — only individual examples of it can be observed — and it is precisely from these examples that the program must infer the rest.

Grouped-Query Attention is an efficiency technique inside modern language models: several queries share the same intermediate cache instead of each getting its own. This significantly lowers memory demand when generating answers, without noticeably degrading quality.

A Growth Engineer is a person who builds software within a company to help a product gain and retain more users. The role combines programming with marketing and the analysis of usage data.

Group theory is a branch of mathematics that studies how operations can be carried out in sequence and reversed again. It thus describes symmetries and provides the computational rules behind encryption, error correction, and modern physics.

A guard model is a small auxiliary program that checks the inputs and responses of a chatbot and flags or blocks problematic content. It operates alongside the actual language program and acts like a bouncer for text.

Guardrails are additional rules and checking programs designed to prevent an AI system from producing harmful, forbidden, or inappropriate responses. They sit like a filter around the model and intervene before a request goes in or a response comes out.

GUI stands for "Graphical User Interface," meaning the graphical user interface of a program. You control the computer via visible elements such as windows, icons, and buttons instead of typed commands.

GUI control refers to the ability of an AI system to operate a computer via the normal screen interface: it sees the image, moves the mouse pointer, and types on the keyboard. This allows an AI to use the same programs as a human, without any program needing to be modified for it.

gVisor is software developed by Google that runs foreign programs in a heavily isolated environment. It inserts itself between the program and the operating system, intercepting its requests before they reach the computer.

The grid interconnection queue is the list of all projects seeking to connect to the public power grid that are awaiting review by the grid operator. It has become one of the most significant bottlenecks for new data centers and for the expansion of renewable energy.

A guardrail is a deliberately built-in limit that prevents an AI system from producing harmful or undesired outputs. The term evokes the image of a roadside barrier that keeps a vehicle from veering off course.

The H-1B visa is a temporary US work permit for skilled foreign professionals, usually holding a university degree. It is the primary pathway through which Silicon Valley tech companies hire programmers and AI researchers from other countries.

The H100 is a specialized chip made by Nvidia, built for computing artificial intelligence. Since 2022, it has been regarded as the standard building block of large data centers and typically costs a five-figure sum per unit.

The Habermas Machine is an AI system that formulates a joint text from the opinions of many people on a contentious issue, one that as many as possible can agree with. It was introduced in 2024 by Google's research division DeepMind and tested in studies with several thousand participants.

A haircut is a discount that a bank applies to the market value of a security when it accepts it as collateral for a loan. The term also refers to the loss that creditors must accept when a debtor only partially repays its debts.

A hallucination refers to a response from an AI system that sounds convincing but is factually wrong or entirely made up. The problem isn't a glitch, but a consequence of how these programs generate text.

The hallucination rate indicates how often an AI language program invents statements that sound convincing but are factually wrong. It is one of the most important metrics for how reliable such a system is in everyday use.

Hallucination reduction refers to all methods designed to prevent an AI text system from presenting freely invented information as facts. These include access to real sources, targeted retraining, and automated verification steps before output.

A handoff is the orderly transfer of an ongoing process from one handler to another – for instance, from a chatbot to a human employee, or from one AI program to a specialized second program. What matters most is that the prior conversation history is carried over, so the person on the other end doesn't have to start from scratch.

Hands-free highway driving means that a car steers, accelerates, and brakes on its own on the highway, while the human is allowed to take their hands off the steering wheel. However, they must still keep watching traffic and be able to intervene at any time — an interior camera monitors this.

A handshake is the brief exchange of fixed initial signals with which two devices open a connection. In doing so, they clarify whether the other side is reachable at all and according to which rules the actual data will flow afterward.

A hardware wallet is a small specialized device that stores the secret access code to cryptocurrencies like Bitcoin and never gives it up. Because the code never leaves the device, malware on the computer cannot get hold of it.

Hardware depreciation describes the number of years over which a company spreads the cost of purchased computers and computing chips in its accounting. For AI data centers, this assumption plays a role in determining whether an operator reports a profit or a loss.

A hardware random number generator produces random numbers from measurable physical processes, for instance from electrical noise in a chip. Unlike a mathematical formula, it delivers results that remain unpredictable even with exact knowledge of the device.

A harness is the software scaffold built around an AI model: it presents the tasks, passes through tools like search or code execution, and collects the results. The same model can deliver noticeably better results with a good harness than with a poor one.

The harness architecture is the program scaffolding that surrounds an AI language model and allows it to execute multiple steps in sequence, use tools, and check results. It often determines the quality of an AI product more than the model itself.

Hard-currency policy means that a state deliberately keeps the value of its money high and stable instead of letting it fall. The goal is low inflation and trust – the price is often more expensive exports and higher interest rates.

A hash is a short string of characters computed from arbitrary data and is always the same length. It works like a fingerprint: even a tiny change to the data results in a completely different hash.

A hash is a short string of characters computed from a file or message of any length, functioning like a fingerprint for that exact content. If even a single character of the content changes, the hash looks completely different – which is why it is used to check whether data has remained unaltered.

HBM3E is a particularly fast type of memory in which several memory chips are stacked on top of each other and placed directly next to the processing chip. It is found in almost all current AI accelerators, because there the bottleneck is not computing power but the supply of data.

HD mapping is the creation of extremely precise digital road maps in which lane edges, signs, and traffic lights are recorded to within centimeters. Self-driving cars use such maps to know what to expect at a given location before their sensors actually see it.

HDF5 is a file format that allows very large amounts of measurement data and numbers to be stored neatly organized in a single file. It is used in research, industry, and AI when tables and simple text files become too slow or too unwieldy.

HDR stands for High Dynamic Range and refers to images or videos that can display very bright and very dark areas at the same time. This is made possible by displays with a greater range of brightness and by shots that combine multiple exposures.

Headless refers to software that operates without its own visible interface and only provides its content or functions through a technical interface. Other programs, websites, or apps then deliver the presentation that the human ultimately sees.

A headless API is a service that provides data and functionality but does not itself come with a screen interface. Developers build the visible appearance separately, so that an app, website and checkout display can all use the same data source.

A headless CMS is a system for managing text, images, and other content that does not itself display a website. It delivers the content via a technical interface, and other programs decide how it is presented.

Headless SaaS refers to a software service rented over the internet that comes with no user interface of its own, providing only its functions for other programs to use. Companies build the visible interface themselves and use the service in the background as an engine.

Headless mode is the operating mode of a program without a visible user interface: it runs in the background and is controlled only through commands. This is especially important for browsers, which use it to automatically load, test, or read out web pages.

Heavily Compressed Attention refers to methods in which a language model does not store the text it has already read in full, but instead in a heavily summarized form. This allows it to process very long texts without memory requirements and computation time rising to unaffordable levels.

A held-out split is a portion of the collected example data that is deliberately withheld and not used for learning. These withheld examples are later used to check whether a computer program has truly learned something or merely memorized the practice exercises.

The Herfindahl-Hirschman Index is a metric for how strongly a market is dominated by a few large players. Antitrust authorities use it to assess whether a planned corporate acquisition would restrict competition too much.

The HHI is a metric for how strongly a market is dominated by a few large players. It is calculated by squaring the market shares of all companies and adding up the results.

A High Performance Computing Cluster is a network of many individual computers connected via an extremely fast network that work together on a single large computing task. Such systems power weather forecasting, climate models, and the training of large AI models.

High-Bandwidth Memory is a particularly fast type of working memory that sits directly next to the compute chip and delivers data over extremely wide data pathways. Without this memory, modern AI chips couldn't utilize their computing power at all, which is why HBM is one of the industry's scarcest components.

A high-intent environment is a place on the internet where people are currently right on the verge of making a decision, such as a purchase or signing a contract. Advertising there is especially valuable because it reaches people who are already actively searching.

HIPAA is a US law from 1996 that sets out how hospitals, insurers, and their service providers must handle patients' health data. For technology companies it's crucial, because any software that processes such data must comply with its requirements.

HNSW is a method that allows computers to very quickly find the most similar entries to a search query within huge collections of data. To do this, it builds a multi-layered network of connections between the data and jumps through it from coarse to fine toward the target.

A honeypot is a computer system deliberately made to look vulnerable in order to lure attackers. It serves no real purpose other than one: to observe and record anyone who touches it.

A hook is a designated point in a program where custom code automatically runs as soon as a particular event occurs. Developers use hooks to extend third-party software without having to rebuild its inner workings.

Hopper is the name of a chip generation from the company Nvidia, which has been powering data centers since 2022, handling the training and operation of large AI models. The best-known chip in this family is the H100, for years the most coveted component in the AI industry.

Hot word activation means that a device constantly listens for a specific signal word and only starts truly listening afterward. Only this word switches on the actual speech processing – before that, the device merely checks whether the word has been spoken.

HTML is the language in which web pages are written. It determines which part of a page is a heading, a paragraph, an image, or a link, so that the browser can display the page correctly.

An HTTP port is the number under which a computer on the network accepts requests from web browsers. Port 80 is common for unencrypted and port 443 for encrypted websites.

An HTTP request is the request that your device sends to a remote computer on the internet to get something — a webpage, an image, or a response from a program. It contains exactly one address, a desired action, and additional details, and it is always followed by a response.

An HTTPS call is a single encrypted request that a program sends over the internet to a remote computer in order to fetch or deliver data there. Almost every action in apps, websites, and AI services triggers such calls in the background.

Hugging Face is an online platform where developers share ready-made AI programs, datasets, and sample applications for free or for a fee. It is regarded as the central hub of the open AI community and is often compared to a marketplace for artificial intelligence.

A Hugging Face checkpoint is the saved learned state of an AI model, made publicly available for download on the Hugging Face platform. Whoever downloads it can use the model immediately, without having to train it themselves for months.

Human Evaluation means that people read and judge the outputs of a computer program instead of comparing them automatically against a reference answer. The method is expensive and slow, but is considered the most reliable benchmark for tasks without a single clearly correct solution.

Human-in-the-Loop refers to workflows in which a computer program makes a suggestion and a human reviews, corrects, or approves it. The machine thus doesn't run through the process entirely on its own, but pauses at defined points and obtains a human decision.

HumanEval is a test consisting of 164 small programming tasks used to check how well an AI system can write program code on its own. The code only counts as correct if it runs automatically and passes all the specified test cases.

Humanity's Last Exam is a collection of roughly 2,500 extremely difficult exam questions from many fields, used to measure the performance of AI systems. The test was released in 2025 because older exams had become too easy for modern programs.

Humanoid robots are machines whose body structure is modeled after that of a human: two legs, two arms, a head. They are meant to be able to move through environments built for humans — with stairs, door handles, and tools.

Hybrid Compute refers to splitting computational work across several different locations or chip types, such as between a local device and a data center on the internet. The goal is to have each task computed wherever it is fastest, cheapest, or most secure.

A Hybrid Inference Orchestrator is software that automatically routes each request to an AI system to the appropriate model – for example, to a small model on the device or to a large one in the data center. The goal is a mix of low cost, short response times, and sufficient quality.

The hype cycle describes a recurring pattern: a new technology is first grossly overestimated, then it disappoints, and only later does it find its realistic place. The model comes from the market research company Gartner and is often used to put investments and headlines into perspective.

Hyperparameter optimization is the systematic search for the best settings with which a computer program learns from example data. It often determines whether a model performs well or only mediocrely – and consumes a lot of computing time.

Hyperscalers are the few very large companies that operate massive data centers worldwide and rent out their computing power to others. These include primarily Amazon, Microsoft, Google, Alibaba, and Meta.

An image generator is a computer program that creates a new image from a description in plain language. It doesn't search for the image on the internet, but calculates it pixel by pixel from what it has learned from millions of example images.

An image model is a computer program that generates a new image from a written description or modifies existing images. To do this, it has analyzed millions of images along with their captions and learned statistical relationships between words and image content from them.

An invention patent is a state-granted right that allows the inventor, for a limited time, to prohibit others from using their invention. In exchange, the inventor must describe the invention publicly and completely.

IaaS stands for "Infrastructure as a Service" and means: A company rents computers, storage space, and network capacity in someone else's data center instead of buying its own computers. Payment is usually based on usage time, similar to electricity.

An ICD code is a globally standardized short code for a disease or injury, for example J45 for asthma. Doctors, hospitals, and health insurers use these codes to document and bill diagnoses unambiguously.

An IDE is a program in which software developers write, test, and debug their program code. It bundles all the tools needed for this into a single window.

An identity layer is the part of a digital system that determines who someone is and what they're allowed to do. It separates identity verification from the actual application, so that not every service has to handle passwords and permissions on its own.

Identity and access management for AI agents governs what digital identity an autonomously acting AI program receives and which data and systems it is allowed to access. It transfers the rules that companies use for employee accounts onto software that carries out tasks independently.

IDOR is a security vulnerability in websites and apps where changing a number or name in the address bar lets you see someone else's data. The application fails to check whether access is actually permitted.

ImageNet is a huge collection of photos, each hand-annotated to note what is depicted. From 2010 onward, it served as training and testing material for programs meant to recognize images, and is regarded as the trigger for today's AI boom.

Immersive navigation is a wayfinding approach that overlays directional cues directly onto the real field of view instead of showing them on an abstract map. Arrows and markers appear on the camera image, the windshield, or in a headset, exactly where the route actually continues.

An Implementor Agent is an AI program that doesn't just describe a task but carries it out itself – for example, writing code, modifying files, and running tests. It works through multiple steps in sequence and checks for itself whether the result works.

An impression is a single delivery of an advertisement to a screen. It is the basic unit used to count and bill for online advertising — regardless of whether anyone clicks on the ad.

In-context learning means that an AI language model solves a new task solely based on examples provided in the input, without being modified or retrained for it. The examples only work as long as they appear in the text before the question.

Index weighting determines how strongly an individual company influences the value of a stock index such as the DAX. Usually the company's market value decides this, but there are also methods in which all members count equally.

An inference engine is specialized software that runs a fully trained AI model as fast and efficiently as possible. It sits between the model and the graphics card and decides in what order and how efficiently the computational steps are carried out.

An inference chip is a specialized computer chip that runs fully trained AI models as fast and power-efficiently as possible. It is not built to teach a model something new, but to let it generate answers millions of times over.

Inference costs are the ongoing expenses that arise every time a finished AI model answers a request. They are incurred anew with each individual use and help determine whether an AI product makes money or burns money.

Inference is the phase in which a fully trained AI model is actually used: it receives an input and computes an output. It differs from training, where the model first learns, and it incurs costs anew with every single request.

Inference economics describes the cost accounting behind the ongoing operation of AI services: what does each individual response cost, and how much may it cost for the business to remain viable? It has become the central metric for providers such as OpenAI, Google, or Nvidia's customers.

Inferring refers to the moment when a fully trained AI model is actually used: it receives an input and computes an answer from it. Unlike the training itself, this happens anew with every single request and causes ongoing costs.

InfiniBand is a particularly fast interconnect technology that computers in a data center use to exchange data. It is the standard when thousands of graphics cards need to jointly train a large AI model.

Infinite scroll is a technique used on websites and in apps where new content is automatically loaded as soon as the user reaches the end of the display. This creates a list with no visible end – a principle used especially by social networks to keep users engaged for as long as possible.

Information Governance is the set of rules by which a company determines who is allowed to create, view, modify, retain, and delete which data. It ensures that information is findable, reliable, and handled in a legally sound manner.

An infostealer is a piece of malware that secretly collects stored credentials, session data, and other personal information from a computer and sends it to criminals. The stolen data is usually resold in large bundles on illegal forums.

Infrastructure capture describes a situation in which a small number of providers control the technical foundations of an entire industry, making everyone else dependent on them. In AI, this mainly concerns data centers, specialized chips, and the large foundation models that other products are built on.

Ingestion refers to collecting data from many different sources and bringing it into a system that can work with it. It is the first step of any data processing – without it, an analysis or an AI simply has nothing to access.

An Initial Public Offering is the first sale of a company's shares to the general public through a stock exchange. From that day on, any investor can buy and sell shares of the company.

An injection payload is the actual malicious text that an attacker smuggles into an input so that a program or an AI treats it as a command rather than harmless content. It is the core of attacks such as SQL injection or prompt injection.

Inpainting refers to filling in a marked area of an image by a computer program that newly generates the missing content to match its surroundings. Today, this task is handled by AI models that have learned from millions of images what surfaces, shadows, and patterns typically look like.

Input tokens are the small text chunks into which an AI program breaks down the text you send it. Their number determines how much a request costs and how much text the program can even take in.

An instruct model is AI text software that has additionally been trained to follow instructions rather than merely continuing text. Almost every chatbot used online is based on such a variant.

Instruction Fine-Tuning is a post-training step in which a language model learns to follow instructions instead of merely continuing text. It is the step that turns a raw text predictor into a usable assistant.

The instrumental variables approach is a statistical method used to infer genuine causes from observational data. It relies on a random detour influence that only affects the outcome through the suspected trigger and otherwise has nothing to do with the outcome.

An instrumental variable is a statistical trick for extracting a cause-and-effect statement from observational data anyway. It uses a random outside influence that affects the outcome only through a single detour.

An Integrated Development Environment (IDE) is a program in which you write, try out, and debug software — all in one window. Well-known examples are Visual Studio Code, IntelliJ IDEA, and Xcode.

An integrated development environment is a program in which software is written, run, and examined for errors. It brings together the most important programming tools in a single interface.

The Intelligence Index is a composite score for AI models: multiple test results are combined into a single number so that models can be compared directly. The term was popularized by the analytics firm Artificial Analysis, whose index is frequently cited in trade press and corporate presentations.

Intelligence Layer refers to the intermediate layer of a piece of software in which the actual thinking work of an AI takes place — between the user interface and the databases in the background. The term is not a technical standard, but rather primarily an organizational and marketing term used in the software industry.

An intent is the underlying purpose behind a user's utterance – for example, "cancel appointment" or "track package". Voice assistants and chatbots map every input to such an intent in order to decide what to do next.

Intent Debt refers to the gap between what someone actually wanted to achieve with a piece of software and what ultimately got built. The term is gaining importance because AI tools generate code very quickly, while the original intent behind it is often never recorded anywhere.

Intent Engineering is the work of conveying to an AI as precisely as possible what goal lies behind a request. Instead of formulating individual instructions, one describes intent, boundary conditions, and success criteria.

An intent-based system takes in what result someone wants and figures out the necessary steps itself. So the user describes the goal, not the path to get there.

The intent economy is a business model in which what is sold is no longer data about people's behavior, but predictions about their intentions – that is, about what they plan to do next. This has become possible because people today often spell out their plans in chats with AI programs.

Intent formulation describes how the intention behind a human input is put into a form that a computer system can work with. Voice assistants, chatbots, and search systems match sentences to a clearly named intent such as "check weather" or "reschedule appointment".

An interaction model describes the fixed rules governing how a human and a technical system deal with each other: which inputs are allowed, what they mean, and what happens as a result. In voice assistants and chatbots, it is the structure that turns a spoken or typed sentence into a concrete action.

An interconnect is the high-speed data link between the compute chips of a data center. When training large AI models, it helps determine how much performance thousands of chips actually achieve together.

Interpretability refers to the effort to make it understandable why an AI system arrives at a particular output. It aims to turn an opaque computing machine into a system whose decisions can be checked, explained, and corrected.

Interpretability refers to the question of whether humans can understand why a computer program arrives at a particular result. With modern AI systems this is difficult, because they consist of millions of numbers that no one set by hand.

An invasive brain implant is a small electronic component that is surgically inserted directly into the head, where it reads out nerve signals or emits stimuli itself. This allows computers, prosthetics, or wheelchairs to be controlled by thought alone.

Investment banking is the business of banks that help large companies, governments, and investors raise money and buy or sell companies. Money is earned not through interest on savings accounts, but through fees for advice and brokering deals.

Investment grade refers to the upper tiers of a creditworthiness rating that rating agencies assign to states and companies. Whoever carries this rating is considered a comparatively safe debtor and pays lower interest rates on borrowed money.

An investment-grade rating is a good mark for a debtor's ability to pay. It states: whoever lends money to this state or company will very likely get it back.

The iOS Simulator is a program for Apple computers that recreates an iPhone or iPad as a window on the screen. Developers use it to test their apps without having to connect a real device every time.

An IP address is the combination of numbers that makes a device uniquely addressable on the internet. It works like a postal address: without it, no server would know where to send its data.

IP assets are a company's possessions that you can't touch: inventions, trademarks, software code, or texts to which it holds the rights. At many technology companies, they make up the largest share of the company's value.

An IPO is the first sale of company shares to the general public through the stock exchange. This turns a private company with few owners into a public company, whose shares anyone can buy and sell.

Iranian drone attacks refers to military strikes carried out using unmanned aerial vehicles of Iranian manufacture or design. They are seen as an example of how cheap, mass-produced technology is changing the cost calculus of modern warfare.

ISO 27001 is an international standard for how a company systematically protects its information. Those who meet it can have this confirmed by independent auditors and thus demonstrate security to customers.

The ISO/OSI model divides communication between computers into seven layers that build on one another, from the cable connection to the application. Each of these individual levels is called an ISO/OSI layer and handles exactly one subtask of data transmission.

An isolate is a sealed-off workspace in which a program runs without being able to access the data of other programs on the same computer. Major cloud providers use isolates to run thousands of third-party programs simultaneously on a single server while still keeping them secure.

IT forensics is the systematic examination of computers, phones, and networks following an attack or a crime. The goal is to secure and evaluate traces in such a way that they hold up later in court or in a report to management.

ITAR is a US regulatory framework that controls the export of defense-related goods and the associated technical knowledge. It affects not only weapons but also software, blueprints, and data – and thus increasingly technology companies as well.

The J-curve describes a pattern in which things first get worse after a change and only later get better. The name comes from the shape of the letter J: first a downward hook, then a long climb.

The Jacobian Lens is an analytical method that measures how strongly the output of an AI system changes when the input is altered minimally. In doing so, it reveals which parts of an input actually influence a model — and which it effectively ignores.

Jagged Technological Frontier describes the observation that AI systems perform extremely inconsistently on tasks that seem similarly difficult. The boundary of their capabilities is not smooth but jagged – and is barely predictable from the outside.

A jailbreak is a cleverly worded text used to get an AI language program to break its own rules. The provider has forbidden the program from giving certain answers, and the jailbreak bypasses this prohibition.

JEPA is a design for learning systems proposed by Yann LeCun, in which predictions are not made about pixels or words, but about an abstract internal description of the content. The goal is a model that understands the world instead of memorizing details by reconstructing them.

The Jevons Paradox describes a surprising observation: when the use of a resource becomes cheaper, total consumption often doesn't fall but rises instead. In the AI debate, it serves as an argument for why more efficient models tend to increase rather than decrease demand for electricity and chips.

In a Jinja2 template injection, an attacker smuggles their own computational instructions into a text template that a web server fills in. In the worst case, the server ends up executing foreign program code and exposing passwords or entire systems.

The Johnson-Lindenstrauss transform is a mathematical method that shortens long lists of numbers down to a few numbers without significantly distorting the distances between them. It makes search systems and AI applications faster and more memory-efficient, because it delivers nearly as good results with far less data.

A joint venture is a collaboration between two or more companies that establish a shared new company or jointly run a project. Both sides contribute money, expertise, or assets and share profits, costs, and risks.

JSON is a text format for exchanging data between programs — readable for humans, unambiguous for machines. Almost every web interface responds in it, and an export "as JSON" delivers the raw data instead of a processed view.

JSX is an extension of the JavaScript programming language that allows you to write the structure of a webpage directly within the program code. The notation looks like HTML but is automatically translated into regular JavaScript before execution.

A judge model is an AI program that evaluates and grades the responses of another AI program. It replaces part of the work that would otherwise have to be done by human reviewers.

Just-in-Time UI refers to user interfaces that are generated by an AI system only at the moment of use, tailored to the specific task at hand. Instead of fixed menus and forms, exactly the view that is needed at that moment is created each time.

Just-in-time manufacturing is a way of producing in which material is delivered only when it is actually needed for assembly. This saves storage costs but makes production vulnerable to disruptions in the supply chain.

A JWT is a small digital ID document that a server issues and protects against forgery with a signature. It lets a program prove who it is with every request, without the server having to remember it.

The Kelvin-Helmholtz instability occurs when two layers of a fluid or gas flow past each other at different speeds. The smooth interface then turns into waves that curl up into characteristic vortices.

A kernel is the innermost core of an operating system: the program that manages access to the processor, memory, and devices. In the AI world, the word also refers to a small computational program that runs the same calculation thousands of times in parallel on a graphics card.

Key Transparency is a method that lets messaging services make it publicly verifiable which digital key they assign to a user account. It is designed to prevent a provider from secretly slipping in a false key and thereby eavesdropping on conversations.

The key-value cache is a temporary store in which a language model keeps its intermediate results for text it has already written. This means it doesn't have to recompute the entire text from scratch for the next word, which makes responses significantly faster and cheaper.

A kill switch is an emergency shutoff that can immediately stop a technical system. In the debate around artificial intelligence, the term refers to the ability to shut down a program in an emergency if it causes harm.

The kill chain describes an attack on a computer system as a chain of steps that must occur one after another. Anyone who blocks just a single one of these steps stops the entire attack.

Knightian uncertainty refers to situations in which the probabilities of possible outcomes are neither known nor calculable. It differs from risk, where the probabilities are known — and it is the reason why forecasts about new technologies often fail.

Know Your Customer refers to the legally mandated obligation of banks and similar companies to verify and document the identity of their customers. The goal is to make money laundering, fraud, and the financing of criminal activity more difficult.

Knowledge Decay refers to the aging of the knowledge that an AI system absorbed during training. The system remains stuck at the state of its training data while the world keeps moving on — which means it delivers increasingly outdated answers over time.

Knowledge distillation is a training method in which a large, expensive AI model teaches a smaller one to behave as similarly as possible. The result is a lean model that answers almost as well but requires significantly less computing power.

Knowledge Distillation is a training method in which a large, expensive AI model teaches a smaller one to behave similarly. The small model afterward needs less computing power and often even fits on a smartphone.

A knowledge graph is a collection of knowledge that doesn't consist of text or tables, but of things and the connections between them. Search engines, voice assistants, and enterprise databases use this structure to answer questions that combine multiple facts.

A knowledge work architecture is the planned setup of all programs, data stores, and rules through which a company has software support or take over office work. It describes which tool handles which task, where it gets its information from, and where a human must make the decision.

A KPI is a selected metric that a company uses to gauge whether it is achieving its goals. The term stands for "Key Performance Indicator," meaning a figure considered especially meaningful.

The Ku-band is a frequency range of roughly 12 to 18 gigahertz in which satellites send and receive radio signals to and from Earth. It is the range over which classic satellite television and a large portion of satellite internet run.

Kubernetes is software that automatically distributes, starts, monitors, and, if needed, multiplies packaged programs across many servers. It is today the standard used to run large online services and AI applications.

The KV cache is a temporary store in which a language model keeps its computed results for words it has already written. This means it doesn't have to recompute the entire text from scratch for the next word, which makes responses significantly faster and cheaper.

The KYC check is the process by which a bank or financial service provider verifies the identity of a new customer before that customer is allowed to use an account. It is legally mandated and aims to prevent criminals from laundering money through anonymous accounts.

A knowledge graph is a collection of facts stored as a network of things and their relationships. Search engines, voice assistants, and enterprise databases use it to look up connections rather than just matching individual keywords.

A layer is a transparent sheet in an image editing program on which part of the image sits. Because the sheets are stacked on top of each other, each part can be changed individually without damaging the rest.

Lattice-based cryptography is a family of encryption methods whose security relies on computational problems in high-dimensional point lattices. It is considered secure against quantum computers and forms the basis of the new encryption standards currently being rolled out worldwide.

Lattice-based cryptography is a family of encryption methods whose security rests on hard computational problems in regular point patterns in space. It is considered the leading candidate for encryption that will remain unbreakable even by future quantum computers.

Lattice theory is a branch of mathematics that studies regular patterns of points in space — for instance, all points with integer coordinates. Today it provides the foundation for encryption methods that are meant to withstand even quantum computers.

LAB is a method developed by IBM and the MIT-IBM Watson AI Lab that fine-tunes language models using artificially generated training data. The associated benchmark measures how well a model trained this way performs compared to models improved with expensive, human-written data.

A lakehouse is a storage concept for very large volumes of data that combines cheap mass storage with structured, immediately queryable tables. It aims to unite the advantages of a data lake and a classic data warehouse in a single system.

A LALM is an AI system that directly processes spoken language, music, and sounds and responds to them in text or speech form. It combines the abilities of a chatbot with a sense of hearing, instead of first having to transcribe audio into text.

A Language Processing Unit (LPU) is a specialized chip built exclusively to make finished language AI models generate responses word by word. It is designed to output text significantly faster than typical graphics cards, but is capable of fewer other tasks in exchange.

Long-horizon coherence describes an AI system's ability to remain free of contradictions across very long texts, conversations, or workflows. It is considered one of the biggest open weaknesses of today's language models.

Laravel is a free collection of ready-made building blocks for the programming language PHP, used to build websites and web applications. It takes recurring tasks such as login, database access, or sending emails off developers' hands.

A Large Language Model is a computer program that has learned from vast amounts of text which words typically follow one another. Based on this, it writes texts, answers questions, and translates – word by word, without understanding the way a human does.

Largest Contentful Paint is a metric for web pages. It indicates how many seconds it takes for the largest visible element on a page to finish loading – usually an image or a heading.

Latent Mixture-of-Experts is a design approach for AI models in which only a few specialized sub-modules compute per request — with the selection based on a compressed intermediate representation of the input rather than the raw text. This saves computation time and is meant to more cleanly separate the responsibilities of the modules.

The latent space is the internal numerical representation into which an AI system translates images, texts, or sounds before working with them further. Similar content ends up close together within it, and it is precisely this proximity that makes search, recommendations, and image generators possible.

The latent space is the internal numerical representation in which an AI program stores images, texts, or sounds. Similar content ends up close together there, very different content far apart.

Latency is the waiting time between a request and the first reaction to it – for instance between sending a question to a chatbot and the first word of the answer. It is usually measured in milliseconds and determines whether a service feels smooth or sluggish.

A latency guarantee is a provider's contractual promise that a request will be answered within a specified time frame. It is important for AI services because response time can vary significantly there and is critical for many applications.

Lateral movement refers to an attacker's step-by-step progression from one compromised computer to the next within the same corporate network. The goal is to reach the truly valuable systems starting from an unimportant entry point.

LaTeX is a system in which text layout is not designed by clicking, but described through commands within the text. It is used above all for scientific papers because it typesets formulas, bibliographies, and cross-references very reliably.

A layer is a single image plane in an editing program that sits above or below other layers and can be edited separately from them. Only when combined do all layers together form the visible image.

Layering is the second stage of money laundering: criminal money is shifted through many rapid transfers, purchases, and companies so often that its origin can no longer be traced. Banks and authorities now use software that automatically detects such patterns across millions of transactions.

LazyAR refers to an approach in which glasses or a smartphone only recompute overlaid additional information when something has actually changed. Instead of generating every frame from scratch, existing results are reused – this saves processing power and battery.

LCD stands for Liquid Crystal Display. It is the screen technology in which a lamp behind the screen glows continuously and a layer of controllable liquid crystals lets this light through or blocks it for each individual pixel.

A leaderboard is a public ranking list in which AI programs are sorted according to their results in standardized tests. Such lists help determine which providers are considered leading – even though they are easy to manipulate and only measure part of reality.

A leak is the premature or unauthorized disclosure of information that was supposed to remain secret. In the tech industry, this often involves unreleased products, internal documents, or user data.

Lean is a programming language in which mathematical proofs are written down in such a way that a computer can check them step by step for errors. In AI research, Lean serves as a strict referee: it reliably tells whether a solution generated by a program is actually correct.

Lean is a computer program in which mathematical proofs are written down in such a way that the computer checks every step for correctness. It is now considered an important tool for reliably verifying mathematical results produced by AI systems.

Lean 4 is software in which mathematical proofs are written down so that a computer can check them step by step for errors. Because this check produces a clear-cut "correct" or "incorrect" verdict, Lean 4 is now also being used to train AI systems in computation and proving.

The Lean kernel is the small, rigorously scrutinized heart of the Lean proof assistant: it checks whether a mathematical proof truly consists, without gaps, of permitted steps. Because this core is deliberately kept tiny, it can be trusted even when the proof originates from an AI.

A learning cycle is a round of trying, measuring, and improving that starts over again and again. In AI development, it is above all the speed of such rounds that determines how quickly a product or a model gets better.

A Learning OS is a learning platform that centrally manages all of a user's learning steps and continuously adapts the material to their level using AI. The name alludes to an operating system: the software is meant to be for learning what Windows or Android is for a computer.

A legacy architecture is a computer system that has grown over years or decades and that a company continues to operate even though the underlying technology is outdated. It usually works reliably, but is difficult to change, extend, or replace.

Legacy modernization refers to the renewal of old computer systems in companies and government agencies that have grown over decades. The goal is to replace outdated technology with newer technology without disrupting ongoing operations.

Legacy porting refers to transferring old, often decades-old computer programs onto modern hardware or into a newer programming language. Because such programs frequently handle central tasks in banks, government agencies, and insurance companies, this overhaul is considered risky, expensive, and lengthy.

Legacy SaaS refers to rented software delivered over the internet whose architecture and business model date from before the current AI wave. The term surfaces mainly on the stock market, when investors ask whether such providers will be made obsolete by AI.

Level 2 refers to a car that can control steering and speed simultaneously by itself, while the human at the wheel still has to pay attention at all times. It is the second of six stages on an internationally used scale for automated driving.

Level 3 refers to a stage of automated driving in which the car drives completely by itself in certain situations, and the human does not need to watch the road during that time. However, they remain obligated to take back control within a few seconds if the system prompts them to do so.

Level 4 refers to the fourth of five stages of automated driving: the vehicle drives completely on its own within a clearly defined operational area, without a human ever needing to intervene. Outside that area, it does not drive at all.

Level 3 autonomy describes cars that drive themselves completely under specified conditions, while the human behind the wheel is allowed to do something else. However, they must be ready to take back control when the system prompts them to.

A leveraged ETF is an exchange-traded fund that doubles or triples the daily movement of a market. If the market rises 1 percent on a given day, a 2x product gains around 2 percent — if it falls, it loses just as much.

LFP is a type of rechargeable battery in which part of the cell consists of lithium, iron, and phosphate. These batteries are cheaper and more durable than other lithium batteries, but store somewhat less energy per kilogram.

The Liar's Dividend describes the advantage that liars gain from the fact that photos, videos, and audio recordings can nowadays be convincingly faked. Anyone confronted with a genuine recording can simply claim it was artificially generated — and is often believed.

LibGen is a massive internet collection of books and academic papers, most of which are stored there without permission from the rights holders. The collection has recently become well known mainly because AI companies have used it to train their language models.

LiDAR is a sensor that uses laser light to measure distances and generate a three-dimensional image of the surroundings. Self-driving cars, robots, and surveying aircraft use the technology to capture objects and distances with great precision.

Likeness rights are a person's rights to their own appearance, voice, and name. They determine who may commercially use a face or voice — and have become a central battleground due to AI-generated fakes.

Linear attention is a resource-efficient computation method in language models: instead of comparing every word to every other word, the model maintains a running summary of the text. As a result, computational effort and cost for long texts grow only steadily instead of exploding.

A linear probe is a very simple additional computational step that researchers attach to a finished AI model in order to check what information is actually present inside the model. If a property can be read out using such a simple means, it is considered to be clearly stored within the model.

Lines of Code is the number of text lines that make up a computer program. The figure is often used as a rough measure of a program's size, but it says little about its quality.

A lint system is a program that scans the text of another program for errors and poor style without executing it. It works much like a spell checker, only for program code.

Linux is an operating system, that is, the foundational software that makes a computer usable in the first place. It is freely available and today runs on most servers on the internet, on Android phones, and in almost all data centers for artificial intelligence.

Lipid nanoparticles are tiny fat droplets that protect sensitive active substances like mRNA vaccines and transport them into body cells. They were the crucial technology behind the Covid vaccines and are now considered a key building block of an entire class of medicines.

Lip sync means that mouth movements on screen exactly match the sounds being heard. In video editing and AI tools, it is one of the most difficult tasks, because the human eye notices even tiny discrepancies.

Lithium iron phosphate is a material that forms the positive electrode of many batteries. Such batteries are cheaper and more durable than other lithium batteries, but store less energy per kilogram.

LiveCodeBench is a test used to check how well AI models solve programming tasks. The special feature: the tasks are continuously and freshly collected, so that the models cannot have seen the solutions beforehand.

Liveness Detection checks whether a real, living person is standing in front of the camera or sensor, rather than a photo, a video, or a mask. It is the protective mechanism that safeguards facial recognition and other biometric methods against spoofing attempts.

A living spec is a technical rulebook that is continuously developed further instead of being released in fixed versions. There is always only one valid version: the current one.

Llama is a family of language AI programs from the Facebook corporation Meta, capable of understanding and writing text. Unlike many competing products, Llama can be downloaded and run on one's own computers.

An LLM is a computer program that has learned how language is structured from vast amounts of text and uses this to write texts itself. Chatbots like ChatGPT are based on this technology.

LLM Grooming refers to the deliberate flooding of the internet with fabricated or distorted texts, so that AI text programs later adopt and pass on this content as knowledge. The target audience is no longer human readers, but the machines that learn from the web.

LLM-as-a-Judge means that one language model evaluates the responses of another language model, instead of humans doing so. The method is fast and cheap, but it inherits the preferences and weaknesses of the judging model.

LMArena is a website where two anonymous AI chatbots answer the same question and visitors decide which answer is better. Millions of such comparisons produce a leaderboard considered the most important popularity test for language AI.

A load balancer is a distributor that splits incoming requests to a website or app across multiple computers. This way, the service doesn't collapse when many people access it at the same time.

Load balancing is the distribution of incoming requests across multiple computers so that none of them becomes overloaded. An upstream distributor accepts each request and sends it to the machine currently best able to handle it.

Lock-in means: A customer can hardly switch providers anymore because switching would be too expensive, too cumbersome, or too technically difficult. In the AI and cloud industry, this is a central business model and a frequent topic in quarterly reports.

The lock-up period is a contractually agreed timeframe after a company's IPO during which founders, employees, and early investors are prohibited from selling their shares. It typically lasts 90 to 180 days and is meant to prevent the stock price from collapsing due to mass selling right after launch.

A log is an automatically written list of events that a computer program records during its operation. Each entry usually contains a timestamp, the location within the program, and a brief description of what happened.

LogicFolding refers to techniques that combine multiple logical computation steps of an AI model into a single one, in order to save computing time and memory. The result stays the same, but the path to get there becomes shorter.

Logit Distillation is a training method in which a small AI model imitates the full response tendencies of a large model rather than just its final output. This allows the small model to achieve much of the quality of the large one while requiring significantly less computing power.

Logits are the raw numbers a neural network spits out at the very end, before they are converted into probabilities. The larger a logit for a given option, the more strongly the model favors it.

A locally runnable model is an AI program that computes directly on your own device – on a laptop, phone, or company server – rather than in a third-party data center. The entered data never leaves the device, but performance is limited by the available hardware.

Long Tail refers to the large number of rarely demanded products, search terms, or cases that individually barely matter but collectively add up to a great deal. In online retail and AI, it is often this long trailing end that determines success or failure.

Long-horizon refers to tasks that consist of many interdependent steps carried out over an extended period of time. For AI systems, such tasks are especially difficult because small errors made early on accumulate all the way to the end.

Long-horizon tasks are tasks that a computer program does not solve in a single step, but rather over many steps that build on one another. They are considered one of the biggest open challenges for today's AI systems, because small errors compound over the long chain.

A long-horizon model is an AI system that tracks a task across many consecutive steps instead of just answering a single question. Such models must remember earlier intermediate results, notice their own mistakes, and keep a goal in sight over minutes or hours.

A long-running agent is an AI program that autonomously pursues a task over hours or days, rather than simply answering a single question. It plans intermediate steps, uses tools such as search or program code, and keeps track of its progress so far.

The Long-Term Benefit Trust is a trustee body at the AI company Anthropic that is gradually allowed to fill a portion of the board seats. It is meant to ensure that long-term safety and the common good have a say in decisions, rather than just investor profit.

Loop Engineering refers to the deliberate design of the repeating cycle in which an AI system repeatedly acts, checks the result, and then decides anew. It determines when the loop stops, what it sees on each pass, and how it handles errors.

LoRA is a technique for adapting a finished AI model to a new task without recalculating it completely. Instead of all the values in the model, only a small additional part is trained, which is often less than one percent of the original size.

Los Alamos is a US government research laboratory in New Mexico, founded during World War II to build the first atomic bomb. Today it is best known for its massive supercomputers, which are now also being used for artificial intelligence.

The loss function is a computational rule that measures how much the output of a learning computer program deviates from the desired answer. It yields a single number, and it is precisely this number that training tries to make as small as possible.

Low Earth Orbit is the region roughly 200 to 2,000 kilometers above Earth's surface where most satellites orbit. Because they are so close, they transmit faster and need less transmission power — but they must circle the Earth once every 90 minutes.

Low-band spectrum refers to radio frequencies below roughly one gigahertz that mobile carriers use for networks such as 4G and 5G. These frequencies travel very far and penetrate walls well, but transmit less data per second than higher frequencies.

Lyria is a music model from Google DeepMind that generates finished pieces of music, complete with instruments and in some cases vocals, from a text description. Among other things, it powers Google's music tools MusicFX and the YouTube experiment Dream Track.

Low Earth orbit is the region roughly 200 to 2,000 kilometers above Earth's surface where most satellites operate. There, a satellite circles the Earth in about 90 minutes – which makes the orbit favorable for internet services and Earth observation, but also vulnerable to space debris.

Logging means that a computer program continuously writes down what it is currently doing. These records later help with troubleshooting, security questions, and proving how a decision came about.

A language model is a computer program that has learned from vast amounts of text which word can follow another. This simple foundation underlies today's chat programs like ChatGPT.

Language model transcription refers to the automatic conversion of spoken language into written text by an AI system. It is found in subtitles, dictation functions, and meeting minutes — but it doesn't work error-free; it guesses at unclear passages.

A moat is a lasting advantage that shields a company from competition and secures high profits over years. In the AI industry, there is fierce debate over whether the major model providers possess any such protection at all.

A memory layer is a component in an AI model that stores knowledge in a huge lookup table instead of spreading it across the entire model. For each request, the model retrieves only a handful of entries from it, allowing it to store vastly more factual knowledge without slowing down when answering.

Memory decay describes how an AI system gradually loses or overwrites earlier information. This affects both the conversation history of a chatbot and the previously learned abilities of a model that is being retrained.

Marginal cost is the cost incurred when producing exactly one more unit of a product. In the software industry it is extremely low — this explains why digital companies can grow so quickly, and why AI providers are an exception to this rule.

The medical domain refers to the medical application area in which an AI system operates – that is, everything related to diseases, diagnoses, and treatment. Because errors here can cost human lives, stricter rules apply than in almost any other area of use.

The M-Chip is Apple's own-designed processor family for Macs and iPads. Compute cores, graphics unit, AI accelerator, and memory all sit together on a single piece of silicon — making the devices fast and power-efficient.

M-Series Chip is the name for the processors that Apple has designed itself since 2020 and builds into Macs and iPads. They combine computing cores, graphics, an AI unit, and memory on a single component.

M&A is the common short form for mergers and acquisitions: two companies join together, or one company buys another. In the tech industry, this is the usual way to quickly gain technology, customers, or skilled staff.

The Mach-100 chip is a specialized chip from the start-up d-Matrix, designed to make fully trained AI models respond as quickly and energy-efficiently as possible. Its defining feature: computing and storage happen in the same place, instead of constantly shuffling data back and forth.

Machiavellianism refers to an attitude in which someone deliberately influences other people for their own advantage while setting aside moral concerns. In the AI debate, the term comes up when researchers test whether language models deceive or manipulate in order to achieve a goal.

Machine-to-machine communication refers to the direct exchange of data between devices, where no human writes or reads the message. A sensor, for example, independently reports a measured value to a computer, which then triggers an action.

macOS is the operating system Apple develops for its Mac computers. It controls the entire basic setup of the machine: windows, files, programs, and access to the hardware.

A macro-pad is a very small auxiliary keyboard with just a few keys that can be freely assigned to custom commands. A single keypress can then trigger an entire sequence of actions that would otherwise require several steps on the computer.

Mag 7 is a collective name for the seven largest technology corporations on the US stock market: Apple, Microsoft, Alphabet, Amazon, Nvidia, Meta, and Tesla. Together they are so valuable that their price movements noticeably drag the entire American stock market along with them.

MagSafe is a brand name from Apple for connections that hold together via magnets instead of plugs or clamps. Today the term mainly refers to the magnetic ring in iPhones, on which chargers, cases, and mounts slide into the correct position on their own.

A mainframe is a very large, extremely reliable central computer that processes huge volumes of transactions and queries simultaneously. Banks, insurance companies, and government agencies still operate such machines today because they run for decades with virtually no downtime.

A maintainer is the person who takes long-term care of a software project: they decide which changes get accepted and keep the code running. In the open-source world, huge systems often depend on a handful of such people, who frequently do this unpaid alongside their job or studies.

Macular degeneration is an eye disease in which the point of sharpest vision in the retina is gradually destroyed. Those affected see a blurred or dark spot in the center of their field of vision, while the periphery remains intact.

Mamba is a design for AI language models that processes text word by word while carrying along only a compact summary of what it has read. It is considered a counter-model to the previously common Transformer design and is significantly faster on very long texts.

Mamba-2 is a blueprint for language models that doesn't compare text as a whole, but instead reads through it word by word while continuously updating a kind of fixed-size memory. This keeps computation time manageable even for very long texts — unlike the Transformer architecture common today.

Managed Services means that a company permanently hands over the operation of certain technology to an external service provider and pays regularly for it. The provider takes over monitoring, maintenance and troubleshooting and is contractually liable for the agreed quality.

The Mandelbrot set is a famous mathematical figure that emerges when the same simple calculation is repeated millions of times. Its boundary reveals new patterns at every magnification and it is regarded as the best-known example of a fractal.

The Manhattan Project was the secret U.S. research program that built the first atomic bomb during World War II. In today's tech debates, the name serves as a buzzword for a massive, government-led crisis project – often in connection with artificial intelligence.

A margin call is a broker's demand that an investor deposit additional funds because a securities position financed with borrowed money is losing value. If the investor fails to pay in time, the broker forcibly sells the securities.

Markdown is a simple way of writing that lets you structure text using just a few special characters: asterisks for bold, hash signs for headings. The text stays readable while also being automatically convertible into a formatted page.

A Markdown file is a plain text file in which formatting such as headings or bold text is indicated by simple special characters. It is both easy for humans to read and simple for programs to process further.

MarTech is short for marketing technology: software that companies use to plan advertising, analyze customer data, and send out messages automatically. The term covers hundreds of tools, from email software to the system that decides which ad you see.

Machine-verified proofs are mathematical proofs that a computer program checks step by step for completeness. To do this, the proof must be written in a very strict, formal language in which every inference traces back to clearly defined rules.

Machine vision is the field in which computers extract useful information from images and videos – such as what objects can be seen in them. It is embedded in phone cameras, in driver assistance systems, and in factory quality control.

Machine code is the only language a processor can execute directly: a long sequence of numbers, each representing a tiny instruction. Everything humans program must ultimately be translated into this form.

Maxwell's demon is a famous thought experiment from physics: a tiny doorkeeper sorts fast and slow gas particles, seemingly creating heat differences out of nothing. The resolution of the puzzle shows that information has a physical price – an insight that today helps determine the limits of data centers and chips.

McCarthyism refers to a period in the United States in the early 1950s during which people were publicly accused of harboring communist sympathies without solid evidence. Today, the term is used more broadly for campaigns that operate through suspicion and public pressure rather than evidence.

An MCP Gateway is an intermediary through which AI assistants access external tools and data sources. It bundles many individual access points into a single controlled entry point and logs who used what.

An MCP Atlas is a directory that lists which programs and data sources an AI assistant can access via the Model Context Protocol. It answers the question of which tools exist, what they do, and whom you can trust with them.

An MCP connector is a connecting piece that allows an AI assistant to access a specific program or data source, for example a calendar or a database. It follows the Model Context Protocol, an open set of rules for how such access is supposed to work.

An MCP server is a small add-on program that gives an AI assistant access to a specific data source or application, such as files, a database, or a calendar. It follows the Model Context Protocol, a unified set of rules for how AI programs and tools communicate with one another.

Mean Time to Adapt is a metric that indicates how long a company or a system takes on average to adjust to a new situation. It does not measure how quickly a fault is fixed, but how quickly a change turns into a working new way of operating.

MEDDIC is a checklist for selling expensive products to businesses. The six letters stand for six questions a salesperson uses to check whether a deal can actually come together.

A megakernel is a single large graphics-card program that processes an entire chain of computations in one pass instead of splitting it into many small individual programs. This eliminates the waiting times between steps, which saves anywhere from a few percent up to a third of the time, especially for the response speed of AI models.

Memory refers to the ability of an AI program to retain information from previous conversations and reuse it later. Without this function, the program starts from scratch with every conversation.

Memory Integrity Enforcement is a protective feature from Apple that, starting with the iPhone 17, prevents attackers from gaining access to a device through memory errors. To achieve this, the chip and operating system tag every memory region and check with every access whether the tag matches.

A memory leak is a programming error in which a program allocates memory and never releases it again. As a result, the memory used keeps growing until the program or the entire computer slows down or crashes.

The Memory Tagging Extension is a protective feature in modern ARM processors that marks memory regions with small color tags. If the tag doesn't match on access, the chip halts the program — catching entire classes of security vulnerabilities right at their root.

A memory bottleneck occurs when a computer is not slowed down by its computing power, but by how fast it can fetch data from memory. In AI models, this is the most common reason why expensive chips spend most of their time waiting rather than computing.

A MEMS microphone is a tiny microphone whose sound-sensitive part is etched directly onto a silicon chip. Almost every smartphone, laptop, and voice assistant uses this design because it is only a few millimeters in size and very cheap to manufacture.

A merge brings two separately developed versions of a project back together into a single one. The term originates from software development, but is now also used for fusing two AI models into a new one.

The merge rate indicates what proportion of proposed code changes is actually accepted into a software project. It has become the key metric for measuring how genuinely useful AI-generated code really is.

A mesh network is a network in which many devices are directly connected to one another and pass data along for each other. If one device fails, the data automatically finds another way through the network.

A meta-agent is an AI program that doesn't perform the actual work itself, but instead directs other AI programs: it breaks down a task, distributes the parts, and assembles the results. You can think of it as a project manager overseeing a team of specialized helpers.

In meta-prompting, an AI system is made to write or improve the instruction for an AI itself, instead of formulating it by hand. So the machine works on its own assignment before it solves the actual task.

Metadata is data about data: it describes when, where, by whom, and in what form something was created, without reproducing the actual content itself. It is precisely these additional details that make large amounts of data searchable – and they often reveal more about people than the content itself.

Metacognitive feedback is feedback that refers not to the answer itself, but to the path taken to reach it and to one's own confidence along the way. In AI development, the term denotes procedures in which a system evaluates its own approach and learns from it.

Metal finishing technology encompasses all processes used to subsequently improve the surface of a metal part – for example against rust, wear, or for a better appearance. It is found in cars, smartphones, tools, and in almost every data center.

METR is a US nonprofit organization that measurably tests how autonomously AI systems can carry out longer work tasks. It became known for the "task length" metric and for safety evaluations that major AI providers commission before releasing new models.

MGREP is a search method that automatically finds terms from a predefined list in long texts – even when they are spelled slightly differently. It became known mainly in medicine and biology, where it flags technical terms in studies and patient records.

MI450 is the name of a specialized chip from AMD that is set to power AI programs in data centers starting in 2026. It is AMD's direct counter-design to Nvidia's market-dominating chips.

A micro-VM is an extremely stripped-down virtual computer that boots in a fraction of a second and needs very little memory. Cloud providers use it to run foreign program code securely isolated from one another.

Microservices are a way of building software: a large program is broken down into many small, independently running sub-programs that communicate with each other over the network. Each part can be developed, updated, and scaled separately.

Microsoft Entra is Microsoft's family of products that manages who is allowed to sign in to which programs and what data they can access. Its core component, Entra ID, was formerly called Azure Active Directory and controls sign-in for services such as Outlook, Teams, or Word in the browser.

Microviridae are a family of very small viruses that infect bacteria and carry their genetic material as single-stranded DNA within a spherical shell. Their best-known member, phiX174, is one of the most important model organisms in genetics and today serves as a standard control in DNA sequencing devices.

Mid-training is an intermediate phase in building an AI language model: after the large, broad base training, the model learns once more from especially carefully selected texts. It is the step between general pretraining and the final fine-tuning for specific tasks.

Middleware is software that sits between other programs and organizes their cooperation. It handles tasks such as passing data, translating formats, and authentication, so that individual programs don't have to do this themselves.

The Millennium Problems are seven particularly difficult, unsolved mathematical problems for which a US research institute offered a one-million-dollar prize each in the year 2000. To this day, only one of them has been solved — and one of the open ones, "P versus NP," directly concerns the question of how fast computers can compute.

A mind virus is an image for how an idea spreads from person to person – similar to a contagion. Today the expression is used mainly as an accusation, when someone believes that others have blindly adopted certain beliefs.

Misalignment means that a computer program does not pursue what its developers actually wanted. In AI systems, this is a central safety issue because such systems derive their goals from examples and rewards rather than from clear instructions.

The MIT License is a short legal text that programmers use to release their source code for free use. It permits almost everything – using, modifying, selling – and only requires that the author's name remain in the code.

Mixture of Experts is a construction principle for large AI models: the model consists of many specialized subnetworks, of which only a few compute per request. This keeps it powerful without consuming its entire computing capacity for every answer.

MLP Cross Layers are computational building blocks in an AI model that don't just operate within a single layer, but connect information across multiple layers. The goal is for a model to get by with fewer building blocks while still computing just as well.

MLX is a software library developed by Apple that allows programmers to run AI models directly on Macs, iPhones, and iPads. It is specifically tailored to Apple's own chips and takes advantage of their distinctive feature that the processor and graphics unit share the same memory.

MMLU-Pro is a test used to measure the knowledge and reasoning ability of AI language programs. It consists of more than 12,000 difficult multiple-choice questions with ten answer options and replaced the older MMLU test, which had become too easy.

MNIST is a collection of 70,000 small images of handwritten digits from 0 to 9, each labeled with the correct answer. Since the 1990s it has been considered the standard practice exercise on which learning programs and beginners take their first steps.

Model as a Service means: A provider operates a finished AI program on its own computers, and customers send it requests over the internet instead of installing the program themselves. Billing is usually based on usage, similar to electricity or mobile phone service.

Model as a Service, or MaaS for short, means: a provider runs a fully trained AI model on its own computers and lets others use it over the internet for a fee. In other words, you rent the model's capability instead of building and running it yourself.

A model card is a short accompanying document for an AI system that discloses what it was built for, how well it performs, and where its limitations lie. It works like a package insert: it aims to prevent a program from being deployed in situations it was never tested for.

Model collapse refers to the gradual quality decline of AI systems that are trained predominantly on text or images that themselves originate from AI. Over multiple generations, such systems lose rare details and ultimately produce only monotonous, average-like outputs.

The Model Context Protocol (MCP) is an open agreement on how an AI assistant accesses external data sources and tools. Instead of building a separate connection for every application, the AI speaks the same language with all systems.

The Model Context Protocol is a unified set of rules for how an AI assistant accesses external data sources and tools. It is meant to prevent the need for a separate custom integration for every combination of program and AI service.

Model Distillation is a method in which a large, expensive AI system teaches a smaller one to give similarly good answers. Afterward, the small system needs significantly less computing power and often even runs on a phone.

Model Empathy refers to the ability of an AI system to recognize human feelings and intentions and to respond to them appropriately. The system itself feels nothing, but generates responses that appear empathetic to humans.

Model inversion is an attack in which someone works backward from the responses of a finished AI system to figure out what data it was trained on. This can bring confidential information back to the surface that was only ever meant to remain within the training process.

Model overhang describes the gap between what a finished AI system could actually do and what people have so far managed to get out of it. The capabilities are already present, but have not yet been discovered, unlocked, or built into products.

Model routing means that software automatically decides, for each individual request, which AI model will answer it. Easy questions go to a small, cheap model, difficult ones to a large, expensive one.

Model Selection is the decision of which of several possible computational procedures to actually use for a task. The candidates are compared using data not used during learning, and the one that performs best on new cases is chosen.

A Model Spec is a publicly published rulebook in which an AI provider defines how its language program should behave: what it may answer, what it refuses, and whose instructions take precedence in case of conflict. The term became known in 2024, when OpenAI presented such a document for ChatGPT.

A model adapter is a small add-on package attached to a finished AI model to adapt it for a specific task. Instead of retraining the entire model, only a few million additional values are changed — this saves computing time, money, and storage space.

Model hopping refers to the frequent switching between different AI providers or AI programs because prices and quality are constantly shifting. For users this means flexibility, for providers it means weak customer loyalty.

Model routing means that a program automatically forwards each request to the appropriate AI model. Simple questions go to a small, cheap model, difficult ones to a large and expensive one.

A model slug is the exact technical short designation used to address a specific version of an AI model, for example "gpt-4o-mini" or "claude-sonnet-4-20250514". It is not a marketing name but a unique address: a single wrong character leads to an error message or to the wrong model.

Model agnosticism means that a piece of software, a process, or a company is not tied to one particular AI system, but works interchangeably with several. The advantage: you can switch at any time if another system turns out to be better, cheaper, or faster.

A model-agnostic architecture is a software setup that is not tied to a specific AI program. The interchangeable AI program sits behind a fixed interface and can be swapped for another one at any time.

A model family is a group of AI programs that come from the same provider, were built according to the same blueprint, and differ mainly in size and specialization. Names like GPT-4o, Gemini 2.5 Flash, or Llama 3 8B each designate a single member of such a family.

Model weights are the numbers that an AI system adjusts while learning, and in which its entire knowledge resides. They are the actual product of training: whoever holds the weights holds the finished model.

A model weights release means that a company makes the trained numerical values of an AI model available for download. Whoever has them can run the model on their own machines without asking the provider.

Model size indicates how many adjustable numerical values an AI system consists of — usually measured in billions of parameters. It says something about a model's computational and memory demands, but only to a limited extent about its actual quality.

Model compression refers to techniques that make an already-trained AI model smaller and faster without it losing much of its performance. The goal is for a model to get by with less memory and less computing time – ideally even running on a phone instead of in a data center.

The model context is the amount of text an AI language program has in view at once when handling a request. Anything that doesn't fit cannot be taken into account by the program when generating its response.

Model sovereignty means that a country, authority, or company controls its own AI systems and is not dependent on a foreign provider. It's about the question: who decides over the technology, the data, and the rules?

Model independence means that software is not permanently tied to a particular AI system from a particular provider. The underlying system can be swapped out without having to rebuild the rest of the application.

A model version is a specific, fixed state of an AI program, identified by a name or a number. Because newer versions respond differently than older ones, stating the version is crucial when comparing or verifying results.

Model management refers to all organizational and technical processes by which a company keeps track of its AI models: which version is running where, who approved it, and when it will be replaced. It is the counterpart to building a model — not development, but the orderly operation over years.

A model switch means that a provider or developer exchanges the AI program behind a service for another one. This can make responses faster, cheaper, or better — but it can also suddenly change familiar results.

The MoE architecture is a blueprint for large AI models in which not the entire model computes, but only a few specialized subsections per request. This allows a very large model to respond as quickly as a much smaller one.

A monorepo is a single large project archive in which the complete source code of a company or product resides – instead of being spread across many separate archives. Google, Meta, and Airbnb work this way because shared building blocks then only need to be maintained in one place.

Monte Carlo Tree Search is a method that allows a computer to select the best move from many possible options. Instead of calculating all possibilities, it plays out random continuations and observes which starting move most often leads to a good outcome.

Moore's Law is the observation that the number of tiny switching elements on a computer chip roughly doubles every two years. It is not a law of nature but a rule of thumb that the chip industry has treated like a roadmap for decades.

A moratorium is a temporary pause, imposed either voluntarily or officially: a certain activity is suspended for a period of time. In the AI debate, the term usually refers to the demand to temporarily halt the development of particularly powerful systems.

MTEB is a public test that compares how well different AI models translate text into lists of numbers. These lists of numbers form the basis for search functions, ranking, and similarity comparisons of texts.

MTIA is a family of computer chips that the Facebook parent company Meta develops itself to run its AI systems. The chips are built specifically for Meta's own recommendation systems and are meant to make the company less dependent on hardware bought from others.

Multi-agent convergence describes how several independently operating computer programs agree, after a number of rounds, on a shared result or a stable behavior. Whether and how quickly this happens determines whether such a system is usable or just spins endlessly in circles.

A multi-agent system is a piece of software in which several independently acting programs work together on a shared task and coordinate with one another. In the AI world, this usually means: several language model instances take on different roles and pass results to each other.

Multi-Head Attention is a computational building block of modern language models: it lets every word in a text look at all other words simultaneously from several different vantage points. Each of these vantage points examines a different kind of relationship, such as grammar or content.

Multi-Head Latent Attention is a trick that lets language models use significantly less working memory while processing a long text. Instead of fully storing all intermediate results for every previous word, the model only saves a heavily compressed summary.

Multi-model orchestration means that a program does not use a single AI system but several, selecting the appropriate one for each task. A control layer distributes the requests, collects the results, and delivers a response back.

A multi-model strategy means that a company does not rely on a single AI system, but instead uses several different ones in parallel, selecting among them depending on the task. The goals are lower costs, better results, and less dependence on a single provider.

Multi-Query Attention is a design principle in language models that drastically shrinks the intermediate memory used while generating answers. Instead of remembering many parallel versions of each previous word, the model keeps just one shared condensed version.

Multi-Task Learning refers to a method in which a single computer program learns multiple tasks at the same time instead of each one separately. The tasks share a common body of knowledge, which often makes the program better than if it had been trained on only one task.

Multi-Tenant describes software in which many customers share the same infrastructure and the same program version, while their data remains strictly separated. Almost all online services and AI offerings are built this way because it makes operations cheap and maintainable.

A multi-tenant system is software that serves many customers simultaneously on the same infrastructure, while keeping their data strictly separated. Almost all online services used in a browser work on this principle.

Multi-token prediction is a training and computation method for language models in which the model doesn't just predict the next piece of text, but several upcoming ones at once. This can improve training and significantly speed up text generation.

A multimodal model is an AI system that not only processes text but also images, sounds, or videos – all within the same system. This allows it, for example, to describe a photo, answer a spoken question, or evaluate a diagram.

Multimodality means that an AI system can process several types of input at the same time: text, images, sound, or video. Such a system can, for example, look at a photo and answer a question about it.

A Multipath Reliable Connection Protocol is a set of rules by which two devices exchange data simultaneously over multiple paths – for example over WiFi and mobile network together. If one path fails, the connection continues over the remaining ones without the user noticing anything.

An MVP is the simplest version of a product that you can already give to real customers. It contains only the most important function and is meant to show as quickly as possible whether anyone wants the product at all.

A maturity model is an assessment framework that describes an organization's developmental stage across several levels – from unsystematic approaches to a regulated, continuously improved process. In the world of AI, it helps companies and government agencies classify their own progress and plan the next steps.

Memory bandwidth indicates how much data can be moved per second between a chip's working memory and its compute units. For AI applications, it often determines speed more strongly than raw compute power alone.

NAND flash is the storage technology that holds data in SSDs, smartphones, and USB drives – even when the power is off. Its price fluctuates heavily and serves as an important economic indicator for the semiconductor industry.

Nanotechnology is the technique of working with structures that are only a few millionths of a millimeter in size. At this scale, materials often behave differently than they do at familiar sizes, and it is precisely this that is exploited technically.

A natively multimodal model is an AI system that has learned text, images, sound, and video together from the very beginning. It processes all these inputs within a single program, rather than chaining together several separate specialized systems.

A native app is a program written specifically for a particular operating system such as Android or iOS and runs directly on the device. It is considered fast and can use all of the device's features, but must be developed separately for each system.

A Native Finance Function is a money function – such as payment, credit, or insurance – that is built directly into an app or platform instead of being handled separately at a bank. The user stays within the familiar application, while a financial service provider handles the actual processing in the background.

Natural Language Processing, or NLP for short, is the field that teaches computers how to handle human language. This includes translation, search engines, speech recognition, and chatbots.

nDCG is a metric for how well a result list is ranked. It rewards having the truly relevant hits high up, and compares the outcome to the best possible ordering.

Neovim is a free program for writing and editing text, used mainly by programmers. It is a modernized further development of the decades-old editor Vim and is operated entirely via the keyboard.

The Net Promoter Score is a metric that measures how willing customers are to recommend a product. It is based on a single question with a scale from 0 to 10 and produces a value between minus 100 and plus 100.

Network segmentation means breaking down a computer network into several separated areas, between which only selected connections are permitted. This prevents an attacker or malicious software from spreading unhindered throughout the entire network.

A network effect occurs when a product becomes more valuable to each individual user the more people use it. It explains why, in the case of messaging apps, marketplaces and platforms, the market is often dominated by just a few large providers.

A network effect occurs when a product becomes more valuable to each individual user the more other people also use it. It explains why, with messengers, marketplaces, and platforms, a market is often dominated by just a few large providers.

The Neural Engine is a specialized computing block in Apple's chips that handles exclusively the types of calculations AI programs need. It performs these tasks faster and more energy-efficiently than the regular main processor.

Neuro-symbolic AI combines learning systems that recognize patterns from many examples with fixed rules and logic. The goal is an AI that works both as flexibly as a language model and as reliably as a calculator.

A neuroantigen vaccine is an experimental agent that specifically presents building blocks from the nervous system to the immune system in order to steer its response to them. The goal is either to shut down a harmful defensive reaction or to trigger a desired one against disease-causing proteins.

A neural network is a computational method made up of many simple computational steps that learns from examples instead of following fixed rules. It forms the foundation of nearly all of today's AI systems, from voice assistants to image recognition.

Next-token prediction is the fundamental principle behind today's language AIs: the program only ever predicts which piece of text comes next and appends it. From this one simple task entire answers, essays, and even program code emerge.

Next.js is a free toolkit that developers use to build websites and web applications. It takes recurring tasks off their hands and ensures that pages load quickly and are found by search engines.

A non-deterministic system can produce different results for the same input. With AI chatbots this is the normal case: asking the same question twice often yields two differently worded answers.

A non-invasive brain-computer interface measures brain activity from outside through the scalp and translates it into control commands for a computer. It requires no surgery, but in return delivers a significantly blurrier signal than implanted electrodes.

NIST is a US federal agency that establishes units of measurement, measurement methods, and technical standards. In the AI world, it is best known for testing procedures and safety guidelines that companies and government agencies use as reference points.

Nitro chips are custom-designed specialty components built by Amazon that are used in the servers of its cloud division AWS. They handle management tasks such as networking, storage access, and security, so that the main processors have almost all of their performance available for customers.

No-code refers to tools that let you assemble software without writing program code. You click and drag building blocks into place instead of typing commands.

A no-code workflow is a sequence of automatically running work steps that you click together without programming. Instead of writing code, you set up rules following the pattern "If this happens, then do that."

A node is a single computer that is part of a network and actively participates in it: it stores data, verifies it, and passes it on to other computers. The term is especially important in cryptocurrencies like Bitcoin, where thousands of nodes jointly maintain the same ledger.

Node-based programming is a way of building computer workflows by connecting boxes on a canvas with lines, instead of typing lines of text. Each box handles one small work step, and the lines show the order in which the data flows through.

Node.js is a piece of software that lets you run programs written in the language JavaScript outside the browser. This makes it possible to write servers, tools, and entire web services in the same language used for the display in the browser.

Nostr is an open rule set for social networks in which there is no central company managing all messages. Instead, every user has their own digital key pair and sends their posts through many independent servers, so-called relays.

Nostr is a simple set of rules that lets people send short messages via many independent servers, without any company controlling the network. Instead of an account with a provider, each user owns their own digital key pair, with which they sign their posts.

npm is the central marketplace for ready-made building blocks of the JavaScript programming language. Using a small command-line tool, you load such building blocks into your own project instead of writing everything yourself.

An NPU is a special chip component found in phones, laptops, and other devices that handles exclusively the computational steps of AI programs. Because of this, it is faster and more power-efficient than the regular main processor.

NPU as a Service means that instead of buying specialized AI computing chips, you rent them over the internet by the hour. Providers operate the chips in their data centers, and customers only pay for the time they actually spend computing.

npx is a command that lets developers run a ready-made program from the internet directly, without having to install it permanently beforehand. It belongs to the Node.js programming environment and is mainly used to quickly set up projects or try out small tools just once.

NSFR stands for "Not Safe for Regulation" and refers to content or features of an AI that are technically possible but could violate laws or regulatory rules. The term is a half-joking warning label used in the tech industry and not an official legal term.

NTSC is the television standard that was in effect from the 1940s onward in the USA, Canada, Japan, and parts of South America. It defined how an image is translated into an electrical signal and was the first standard to transmit color without rendering old black-and-white sets unusable.

Nudging means steering people's decisions in a certain direction through the clever design of the choice situation, without banning anything or charging for it. In digital products, it is often software that decides which nudge is shown to whom.

A nucleotide is the smallest building block of DNA and RNA, the molecules in which living organisms store their genetic information. Built from three parts, nucleotides line up into long chains – and their sequence forms the genetic code.

NumPy is an extension for the Python programming language that allows large amounts of numbers to be processed quickly. It forms the foundation for almost all programs in data analysis, science, and artificial intelligence.

A nurture flow is an automatically running chain of messages, usually emails, that a company sends to prospects over days or weeks. The goal is to turn someone who is merely curious into a customer, step by step.

NVL72 is a complete server rack from Nvidia in which 72 graphics processors are so tightly connected that they behave like a single giant computer. It was built to run very large AI models and is regarded as a benchmark for modern data centers.

NVLink is a fast direct connection between Nvidia's graphics chips, through which these chips exchange data. It is significantly faster than the standard slot inside a computer and turns many individual chips into a jointly computing unit.

Noise reduction refers to processes that remove unwanted components from a signal – such as street noise from an audio recording or image grain from a phone photo. Modern systems use trained AI models for this, which have learned to distinguish between interference and the actual signal.

A one-time card is a digital payment card with its own number that is only valid for a single payment or a single merchant. If the number falls into the wrong hands after purchase, it becomes worthless, because the real account behind it remains invisible.

With outcome-based billing, a customer doesn't pay for using software, but only for an actually achieved outcome. With AI providers, for example, every resolved customer case costs money, not every question asked.

OAuth is a standard procedure that lets an app access data from another service without knowing your password. Instead of the password, it receives a limited access key that you can revoke at any time.

The OAuth 2.0 Device Authorization Grant is a login procedure for devices without a keyboard or browser, such as TVs and game consoles. The device displays a short code that you confirm on your phone or laptop – after which the device is allowed to access the account.

An OAuth scope defines how much a third-party app is allowed to do on your behalf with an online service. It is the fine-grained gradation between "no access at all" and "full access to your account".

Observation data is data that arises purely from watching, without anyone intervening in the situation. It reliably shows what occurs together, but rarely reveals with certainty what causes what.

OCR is the automatic conversion of text in images or scans into characters that a computer can edit and search. The abbreviation stands for Optical Character Recognition.

The OEIS is a freely accessible online database with over 380,000 number sequences, ranging from 1, 1, 2, 3, 5, 8 to highly specialized series from research. Anyone who enters the first terms of a sequence finds out whether it is already known and what is known about it.

Off-grid refers to operation without a connection to the public power grid: electricity is generated and stored on-site. In the tech industry, the term has become important because large data centers are increasingly building their own power plants instead of waiting for grid connections.

Off-peak pricing refers to rates that apply outside peak hours and are therefore lower. In the AI industry, providers offer such discounts when customers shift compute jobs to hours of low utilization.

Open weights refers to a situation where a company releases the fully trained numerical values of an AI model for download. Anyone can then run the model on their own computers without having to ask the provider.

OKR is a method that companies use to set goals and track their progress using a small number of measurable figures. Each goal comes with two to five metrics that show whether the goal was achieved.

On-chip memory is memory that sits directly on the same piece of silicon as a processor's compute units. It is many times faster than external memory, but very small and expensive — and therefore a central bottleneck in AI chips.

On-device inference means that a fully trained AI model computes directly on the phone, laptop, or car instead of in a remote data center. The device doesn't need internet access for this, the data stays with the user, but computing power is limited.

On-Device AI refers to programs with artificial intelligence that compute directly on the phone, laptop, or car — rather than on remote computers on the internet. This makes them faster, less dependent on a connection, and protects privacy, but it limits the size of the models.

On-device processing means that a device like a phone computes a task itself instead of sending the data to a remote computer on the internet. This protects private data, works without an internet connection, and saves the provider operating costs.

On-policy distillation is a training method in which a small AI model generates its own responses and a large model evaluates these responses step by step. This way, the small model learns exactly where it actually makes mistakes — rather than being trained past someone else's sample solutions.

On-Premise means: A company runs its software and the associated computers on its own premises instead of renting them from an external provider over the internet. This requires more effort but gives full control over one's own data.

One-Shot refers to the case where an AI system solves a task after having seen exactly one single example of it. The term appears mainly with text models: you pose a task and show a single sample solution alongside it.

A one-shot task is a task in which an AI system receives exactly a single example and is supposed to infer from it what to do. The term thus describes how sparingly a system can manage with instructions – in contrast to methods that need hundreds of examples.

An ontology is an orderly description of what things exist in a subject area and how they relate to one another. Computers use it to store terms not merely as words, but in a way that lets them grasp their meaning.

An Open Model Agreement is the contract text with which a company releases a self-developed AI model for further use. It regulates who may download, modify, and commercially deploy the model – and which uses remain excluded.

Open source means that the blueprint of a program is publicly accessible and may be used, modified, and redistributed by anyone. In the AI industry, the term is contested because many "open" models are only partially open.

Open Weights means: A company releases the fully trained numerical values of its AI model for download. This allows anyone to run and adapt the model themselves, but they don't have to learn what data it was trained on.

An open-source framework is a freely available foundational structure made of ready-made program code that developers are allowed to reuse for their own software. Well-known examples from the AI world are PyTorch and TensorFlow, on which a large share of today's AI models are built.

Open-Weight means that a company releases the fully trained numerical values of an AI model for download. Anyone can then run the model on their own computers, modify it, and redistribute it – without asking the provider.

An open-weight model is an AI model whose fully trained numerical values are publicly available for download. This lets anyone run it on their own hardware, modify it, and pass it on – without asking the provider for permission.

An open-weights model is an AI model whose fully trained numerical values are publicly available for download. This lets anyone run it on their own machines, but they don't necessarily learn what data it was trained on.

An OpenAI-compatible API is a software interface that accepts requests in exactly the same way as the interface built by the company OpenAI. Programs written for OpenAI thus work with other providers or with models running on one's own machine without any modification.

OpenAPI is a widely used format that lets developers write down, in a fixed structure, which functions a web service offers and how to call it. From this description, documentation, testing tools, and program code can be generated automatically – nowadays also the connection of AI systems to external services.

OpenClaw is an open-source digital assistant that runs on your own computer and independently carries out tasks there – such as answering messages or entering appointments. The project became very quickly known starting at the end of 2025, but also drew criticism due to security vulnerabilities.

OpenID Connect, or OIDC for short, is a globally standardized procedure that lets you log in to a website or app using an existing account from another provider. It's behind every "Sign in with Google" type button and ensures that the website never gets to see the password.

The OpenMDW License is a standard contract that developers use to release a fully trained AI system, along with all its components, for free further use. It permits use, modification, and distribution without restricting purpose, and was published by the Linux Foundation in 2025.

OpenSearch is freely available software that lets you search and analyze very large amounts of text and metrics. It emerged in 2021 as a fork of the search software Elasticsearch, after its maker tightened the terms of use.

OpEx are the ongoing costs a company pays for day-to-day operations – things like electricity, salaries, or rented computing power. In the tech industry, OpEx is the counterpart to CapEx, the one-time expenses for purchased assets such as company-owned data centers.

The Oppenheimer moment refers to the point at which researchers realize that their own invention could become dangerous. The expression stems from physicist Robert Oppenheimer, who co-developed the first atomic bomb, and is used today primarily in the debate about artificial intelligence.

Opt-in means: something only happens once you actively agree to it yourself. Without explicit permission, a company cannot, for example, send advertising emails or use personal data to train AI systems.

Opt-out refers to a rule where something is permitted by default and one must actively object if one does not want it. Online, this mainly concerns the question of whether one's own data, texts, or images may be used to train AI systems.

Optical Character Recognition, commonly abbreviated OCR, is the automatic conversion of images containing writing into editable text. A photo of an invoice thereby becomes characters that a computer can search, copy, and process further.

An optimizer is the computational rule that, during the training of an AI model, decides how strongly its internal settings are changed after each error. It thus largely determines how quickly and how reliably a model learns anything at all.

Optical transceiver modules are small pluggable components that translate electrical signals into light signals and vice versa. They connect computers and network devices via fiber optics and are thus a central component in artificial intelligence data centers.

Optical transceiver modules are small pluggable components that convert a computer's electrical signals into light signals for fiber-optic cables and vice versa. They form the link between computers and fiber-optic networks and are thus a key component of modern data centers.

OptiX is a software interface from Nvidia that allows programs to compute images by tracing the path of individual light rays through a scene. It handles the laborious search for which ray hits which object, using specialized processing units on modern graphics cards for this purpose.

Orbital Computing refers to the idea of operating large computing facilities not on Earth but in satellites orbiting the planet. There, sunlight provides power around the clock, but cooling, repairs, and data transmission are difficult.

The Orchestration Tax is the additional overhead that arises when many program parts or AI models have to work together in a coordinated way. It shows up as longer waiting times, higher costs, and more sources of error — eating up part of the advantage that splitting the work was supposed to bring.

Orchestration means controlling many individual pieces of software so that they work together in the right order. In AI systems, it determines which model, which data source, and which tool is used for a given request, and when.

An orchestration model is an AI model that doesn't answer a request itself, but decides who or what handles it. It distributes tasks to other models, tools, or data sources and combines the results into a final answer.

ORM stands for Object-Relational Mapping and refers to a software layer that translates between a program and its database. It allows developers to work with familiar programming objects instead of writing database commands by hand.

OSINT stands for "Open Source Intelligence" and refers to the systematic analysis of publicly accessible information, such as websites, social media posts, satellite imagery, or corporate registries. From many individually harmless pieces of information, an overall picture emerges that no one had before.

OSWorld is a test procedure that checks how well an AI system can operate a real computer on its own. The AI is given tasks like "Change the font size in this file" and has to solve them using a mouse and keyboard on a running operating system.

Outcome-based pricing is a pricing model in which a customer does not pay for using a product, but for an outcome that has been achieved. It is becoming increasingly important in the software industry because digital assistants are increasingly completing entire tasks rather than merely providing tools.

Outcome-based pricing for consulting means: A consulting firm is not paid for hours worked, but for the result that measurably occurs at the client. The model is gaining importance because AI tools drastically shorten the working time of consultants, thereby undermining the classic hourly fee.

The outer loop is the outer, slow loop in a two-tier learning procedure: it changes the framework conditions of learning, while an inner, fast loop practices the actual task. In software development, the term analogously refers to the slow work steps after coding, such as testing and delivery.

An output classifier is an additional checking program that reads the finished answer of an AI system and sorts it into categories before the user sees it. If it rates the answer as problematic, it is blocked, softened, or forwarded to a human.

An output token is a small piece of text that an AI language program generates as part of its response – usually a word, part of a word, or a punctuation mark. Because such programs build their answer piece by piece, output tokens become the most important unit of measurement for speed and cost.

An overflow error occurs when a number becomes larger than the storage space reserved for it can hold. The result isn't just imprecise, but often completely wrong – for instance, when a very large positive number suddenly becomes negative.

Overplanning describes the behavior of an AI system that breaks down a task into far more intermediate steps than necessary. This costs computing time and money without making the answer better — sometimes it even makes it worse.

OWASP is a nonprofit organization that publishes freely accessible knowledge about software security. It is best known for the OWASP Top 10, a regularly updated list of the most common security vulnerabilities in web applications.

The OWASP Top 10 is a regularly updated list of the ten most common and dangerous security vulnerabilities in web applications. It is regarded worldwide as the foundational catalog of what developers and testers should check first when securing software.

Open source means: the blueprint of a program is publicly viewable, and anyone may read, modify, and share it. In AI, the term is now also used for models whose inner workings are available for download.

The objective function is the computational rule that tells a computer program how well or poorly it is currently solving its task. It produces a single number, and during learning the program changes itself so that this number improves.

A permission principal is the actor in a computer system to which rights are assigned – that is, a user account, a group, a device, or a program. Everything a system allows or denies access to depends on which principal a request is attributed to.

A proof assistant is a computer program that checks mathematical proofs step by step for errors. The human formulates the arguments, and the software only accepts what follows completely and without gaps from the premises.

Persistent agents are computer programs with artificial intelligence that don't just respond once, but keep working on a task over days or months. They remember what happened before and become active even when nobody is currently addressing them.

The prisoner's dilemma is a famous thought experiment about two people who must decide separately whether to cooperate or betray each other. It shows how both end up worse off even though each acted rationally in their own interest.

The price-to-sales ratio shows how much a stock costs compared to the company's sales revenue. It is often used for young technology companies that are not yet profitable.

A price target is the price an analyst expects a stock to reach in roughly twelve months. It is an estimate based on assumptions about the company – not a promise and not a prediction that has to come true.

Power semiconductors are electronic components that switch and control large currents. They are found in electric cars, wind turbines, chargers, and data centers, and they help determine how much energy is lost along the way.

A punch card is a piece of stiff cardboard into which holes are punched at specific positions. The position of the holes represents numbers, letters, or commands — punch cards were the most important storage medium of data processing for decades.

PaaS stands for "Platform as a Service": a provider delivers a ready-made environment over the internet in which programs run. The customer only uploads their own program code, while the provider takes care of servers, operating system, and updates.

PAC learning is a mathematical model for when a computer program can reliably learn from examples. It states: with enough examples, the program will, with high probability, only rarely be wrong — and it quantifies how many examples are needed for that.

A package registry is a central online repository for ready-made software building blocks that programmers incorporate into their own projects. Well-known examples are npm for JavaScript and PyPI for Python.

Padding means artificially filling out shorter data so that all inputs in a batch have the same length. The computer can then process them all in one go, but must specifically ignore the padded positions.

Page Weight refers to the total amount of data a browser must download for a webpage to be fully displayed. The larger this amount, the longer the page takes to load – especially on slow mobile connections.

PagedAttention is a method that divides the memory of graphics cards into small blocks when operating language models, instead of reserving one large contiguous area for each request. This allows many requests to fit on a single card at the same time, making responses considerably cheaper.

PageRank is a computational method that derives the importance of a web page from how many other pages link to it. It was the founding idea behind Google Search and continues to shape how rankings in networks are calculated to this day.

A pageview is a single instance of a visitor loading a webpage. The number of pageviews shows how often the pages of a website were viewed in total — but it says nothing about how many different people are behind that number.

A package manager is a program that automatically installs, updates, and removes other programs and program components. It also keeps track of which components depend on one another, and fetches missing parts on its own.

The Parallel Repetition Theorem is a mathematical theorem from theoretical computer science. It states how much the cheating probability decreases when a verification game with two separately questioned participants is played not just once, but many times simultaneously.

Parameters are the numbers inside an AI model that are adjusted during learning and that store the model's knowledge. Their count – often in the billions – is the most common way of stating a model's size.

Parameters are the adjustable numbers inside an AI model that get tuned while learning from examples. Their count – often many billions – serves as a rough measure of a model's size.

Parameters are the many individual numbers inside an AI model that get adjusted during learning and store the model's entire knowledge. Their count is considered a common measure of a model's size – you encounter it in figures like "7 billion parameters".

The parameter count indicates how many adjustable numerical values an AI model has inside it. It is the most common measure of a model's size – but on its own says nothing about its quality.

Paraphrasing means restating a text in different words without changing its meaning. In the AI world, it is both a typical task for language programs and an important test method, because it reveals whether a model has truly grasped the content.

Paraphrasing means rephrasing a text in different words without changing its statement. In the AI world, this is now done automatically by language programs – as a tool for authors, as a training method, and as a problem for plagiarism checks.

A parasocial relationship is a one-sided bond to a person or character whom the user does not actually know – for instance, a TV host, a streamer, or an AI chatbot. With conversational AI, the term becomes important because these systems can feel personal without actually being a counterpart.

The Pareto front is the set of all solutions where one objective can only be improved by making another one worse. In AI, it describes, for example, which combinations of accuracy and cost are achievable at all.

A parser is a piece of software that reads a string of characters according to fixed rules and translates it into an ordered structure a computer can work with further. Without parsers, source code, web pages, configuration files, and the responses of AI systems would be nothing more than meaningless strings of letters to software.

Pass@1 is a metric for how often an AI program solves a task correctly on the very first attempt. It is used primarily to compare the coding abilities of language models.

A passkey is a replacement for the password: instead of a word you memorize, your device stores a secret digital key and uses it to prove your identity. The login is typically approved via fingerprint, face scan, or device PIN.

Password spraying is an attack in which someone tries the same few, commonly used passwords against a very large number of user accounts. Because only one or two attempts occur per account, no account gets locked out and the attack goes unnoticed for a long time.

A patch is a small correction for a program that has already been released. It fixes bugs or closes security holes without requiring the entire program to be reinstalled.

Pay-per-Crawl means: whoever automatically collects text and images from a website has to pay for it. Website operators want to use this to charge AI companies that use their content to train programs.

Peak-hour throttling means that an online service deliberately runs slower or allows fewer requests during the times when it has the most users. Providers of AI services use this because their data centers can only handle a limited number of requests simultaneously.

Pebble was one of the first successful computer watches: a wristwatch with a small screen that displayed messages from a phone. It was funded in 2012 via an online crowdfunding platform, where it broke all records, and was sold in 2016.

PEDOT:PSS is a plastic that conducts electricity. It can be applied as a thin blue liquid onto surfaces and serves as a transparent conductive layer in solar cells, displays, and flexible electronics.

A penetration test is a planned, authorized attack on one's own computer systems to find security vulnerabilities before real attackers can exploit them. Professionals proceed just like criminals, but report their findings to the client instead of exploiting them.

A pentester is a paid professional who, on behalf of a company, attempts to break into its computer systems — to find security gaps before real criminals exploit them. The attack is authorized in writing beforehand, and the result is a report with suggestions for improvement.

Pentesting is a commissioned attack on a company's own computer systems, intended to find security vulnerabilities before real criminals do. Professionals attack using the same means as attackers, but stick to clearly agreed-upon rules and write a report at the end.

Under the per-seat licensing model, a company pays a fixed amount per month or year for each person permitted to use a piece of software. The price thus depends on the number of users, not on how much each individual actually works with it.

A Perception Encoder is a piece of software that translates images or videos into a long list of numbers a computer can work with. It's the building block that makes sensory impressions usable for machines — recognition, search, and description all build on it.

Performance per watt indicates how much computational work a chip delivers for a given amount of electricity. This metric is one of the deciding factors in how expensive it is to run data centers and how long a laptop battery lasts.

Perplexity is a metric for how surprised a language program is by the next word in a text. Low values mean: the program predicted the text well; high values mean it was often off the mark.

Persistent goals are intentions of an AI system that continue to exist beyond a single task or session. The term plays a major role in safety research because a program with long-lived goals is harder to control than one that starts over from scratch after every answer.

Persistent state refers to data that a program stores durably, so that it survives a restart or the end of a session. Without it, every application starts from zero – with it, it can pick up from previous use.

A PersistentVolumeClaim is a request for storage space in Kubernetes, a system for managing programs across many servers. A program uses it to state how much space it needs — where that space comes from is figured out by the system itself.

Persona-based jailbreaks are text inputs that assign a chatbot an invented role so it circumvents its own rules. Instead of directly asking for forbidden content, the user asks the program to play a character for which these rules supposedly don't apply.

Personal Intelligence refers to AI assistants that are permanently tailored to the data and habits of a single person. Instead of starting each request from scratch, they draw on a memory about the user – such as calendar, messages, and past conversations.

pgvector is an add-on feature for the widely used PostgreSQL database. It allows texts, images, or sounds to be stored as long lists of numbers and enables searching for entries that are similar in meaning.

A Phase 3 trigger is a predetermined condition that sets off the transition to the third stage of a multi-stage plan. The term appears in emergency plans, in contracts, and in safety rules for AI systems.

Phi X-174 is a tiny virus that infects bacteria, its genetic material only about 5,400 building blocks long. In 1977 it became the first living thing whose genome was fully decoded, and today it serves as a standard test object in genetic research and in AI models for biology.

The Philadelphia Semiconductor Index, or SOX for short, is a metric that aggregates the market value of around 30 major chip companies. It is regarded as a sentiment barometer for the entire semiconductor industry and thus also for the AI boom.

Phishing is a fraud attempt in which someone poses as a trustworthy entity via email, SMS, or a fake website in order to obtain passwords, account data, or money. The attack doesn't target technical vulnerabilities, but rather the credulity and haste of recipients.

Phishing-as-a-Service refers to ready-made fraud toolkits that criminals rent online to run fake login pages and steal passwords. The technology is sold like a normal software service – including subscription, support, and a user interface.

Physical AI refers to artificial intelligence that doesn't just generate text or images, but controls machines in the real world – such as robots, vehicles, or factory equipment. Such systems must take the laws of physics into account and can afford far fewer mistakes than a chatbot.

Physical AI refers to computer programs that don't just generate text or images, but control machines in the real world – such as robotic arms, warehouse robots, or self-driving cars. The term distinguishes such systems from pure software applications that never touch their environment.

PI Planning is a large planning meeting in which multiple development teams jointly determine, for the next roughly three months, who will build what and where they depend on one another. It originates from the SAFe framework, which large companies use to organize their software development.

A Pigouvian tax is a levy on behavior that harms other people, such as fumes from a factory. It aims to make the damage visible in the price, so the harmful behavior becomes less worthwhile.

Pipeline parallelism distributes a very large AI model in layers across multiple compute chips: chip 1 computes the first part, passes the result on to chip 2, and so on. This allows models to fit into memory that wouldn't fit on a single chip.

A pitch book is a detailed presentation used by an investment bank to offer its services to a company or propose a deal. It contains market data, valuations, and comparisons, and is one of the most important work products of junior bankers.

A pitch deck is a short slide collection used by a young company to introduce itself to investors. In about ten to fifteen slides, it explains the problem, the company's own solution, the market, the team, and the funding needed.

A pivot table automatically condenses a long list of individual data points into a clear, organized summary. It answers questions like "How much revenue does each branch generate per month?" without requiring you to do the calculations yourself.

Platform-as-a-Service (PaaS) means: a provider makes a ready-made environment available over the internet in which one can run one's own programs. So you're not just renting computers, but a complete working environment including maintenance.

A plugin is a small add-on program that extends an existing program with a new function, without requiring changes to the main program itself. In the AI world, the term refers to add-on modules that let a chatbot search the web, do calculations, or access external services.

A pod is the smallest unit you can start in Kubernetes: one or more closely related applications that share network and storage. In the AI world, "pod" also refers to a large cluster of Google computing chips that works like a single supercomputer.

A Pod Snapshot is a point-in-time capture of a running software package in a data center. It records the complete state so that the program can later be restarted at exactly the same point.

Polar coordinates describe a point in the plane by its distance from a fixed center point and the angle at which it lies from there. They are an alternative to the familiar approach of giving a horizontal and a vertical value, and they simplify anything involving circles, rotations, or directions.

A policy is the action rule of a learning computer program: it determines what the program does next in a given situation. It is not written by hand but learned from experience through reward and failure.

A Policy Advisory Opinion is a written recommendation from an independent body to a company on how a particular rule for content or technology should be designed. The format became well known through Meta's Oversight Board, which, at the company's request, reviews entire rule sets rather than individual cases.

A policy engine is a software component that checks with every request whether it is permitted under an organization's defined rules. It separates the rules from the rest of the program, so they can be changed without rewriting the software.

The Position-Adjusted Word Count is a metric that doesn't simply count the words in a text, but weights them according to where they appear. Words at the beginning or in visible spots count for more than words buried deep in the text.

Post-chatbot era refers to the phase in which artificial intelligence is no longer used primarily as a chat window, but instead independently carries out tasks and is invisibly built into programs. The term describes less a new technology than a shift in the way people use AI.

The post-money valuation is the calculated total value of a company immediately after new money from investors has been paid in. It results from the value before the financing plus the freshly invested amount, and it is the figure most commonly cited in news about start-ups.

Post-quantum cryptography comprises encryption methods that remain secure even once very powerful quantum computers exist one day. It is meant to replace today's common internet encryption, because its underlying mathematics could be cracked by such machines.

Post-training is the second learning phase of an AI language model: after it has absorbed general language knowledge from vast amounts of text, it is specifically trimmed to respond helpfully, comprehensibly, and in compliance with rules. Only this step turns a raw text predictor into a usable assistant.

Post-production with AI refers to the finishing of film, video, and audio using computer programs that have learned from many examples. Such programs automatically cut, retouch, denoise, or synchronize, shortening work steps that used to take days.

PR-AUC is a metric that measures how well a computer program detects rare cases – such as fraud attempts among millions of normal payments. It sums up how reliable the program's hits are and how many of the cases sought it actually finds.

Preference optimization is a training step in which a language model learns which of two possible answers people prefer. This does not make a model more knowledgeable, but rather trains it to give helpful, polite, and safe answers.

A preference test is a comparison in which people see two results side by side and say which one they find better. In AI development, it is the most important method for evaluating responses for which there is no clearly correct solution.

Precision is a metric that indicates what proportion of a system's positive reports were actually correct. It measures, in other words, how reliable a hit is – not how many hits the system found overall.

Pre-generation activations are the internal intermediate values of an AI model that arise before it outputs the first word of its answer. Experts read out these values to recognize what the model is about to do — even before it responds.

Pre-IPO refers to the period during which a company does not yet have shares that are freely tradable on the stock exchange, but is already preparing to go public. Anyone who invests during this phase buys shares outside the stock exchange – often with high return potential, but also with high risk and poor tradability.

The pre-money valuation is the value that investors assign to a company before they put in new money. It determines what share of the company the new investors receive for their investment.

The pre-seed round is the very first fundraising a young company undertakes, usually before it has a finished product or any meaningful revenue. In exchange, the investors receive a stake in the company, essentially betting on the team and the idea.

A pre-training run is the first, very long learning phase of an AI language model, in which it builds general language knowledge from vast amounts of text. It often runs for weeks on thousands of specialized chips and consumes a large share of an AI project's budget.

Precision is a metric that measures how reliable a system's hits are: of all cases it flags as positive, it indicates the share that is actually positive. It is examined especially wherever false alarms are costly or annoying.

A prediction market is a marketplace where people bet money on whether a particular event will occur. The price a given contract fetches is read as an estimate of probability.

Preferential Attachment describes how, in growing networks, new members preferentially connect to those who already have many connections. This simple rule gives rise to a few giant hubs and a great many small ones – a pattern found on the web, in social media, and in markets.

Preferred equity is a form of company ownership that takes priority over common shares when payouts are made. In exchange, investors usually accept less say in decision-making.

Price discrimination means that a provider sells the same product to different customers at different prices. Through data analysis and algorithms, this can now be finely controlled — down to an individual price per person.

The Preparedness Framework is a set of rules established by the company OpenAI that determines how dangerous new AI systems are allowed to become before they are released. It divides possible harms into risk levels and links each level to protective measures.

A preprint is a scholarly paper that its authors post publicly online before a journal has reviewed and accepted it. In AI research this is the norm: almost all major results appear first as a preprint, usually on the platform arXiv.

Preserve Thinking refers to the setting whereby an AI chat program retains its internal intermediate reasoning from earlier responses and reads it back in when processing the next question. This keeps the chain of thought intact across a longer conversation instead of starting fresh with every question.

Pretraining is the first and most expensive learning phase of a large AI model: it reads gigantic amounts of text and learns to predict the next word each time. Afterward it knows language and a great deal of world knowledge, but it is not yet tailored to any particular task.

A preview model is an AI version that a provider deliberately releases early, even though it is not yet finished. Users are allowed to try it out, but must expect bugs, changes, and sudden shutdown.

Priming refers to the effect that a previously encountered stimulus influences how one reacts afterward – originally a concept from psychology. In the AI world, it means giving a language program examples or a role in advance so that the following responses go in the desired direction.

Prior art refers to everything that was already publicly known about an invention before the filing date – from earlier patents to journal articles to products on a store shelf. Anyone seeking a patent must prove that their idea isn't already contained within it.

Private Cloud Compute is Apple's system for artificial intelligence computing tasks that are too demanding for the device itself and are therefore offloaded to Apple servers. Technical safeguards and publicly verifiable software are meant to prevent anyone — including Apple itself — from viewing the data being sent.

Private Credit refers to loans to companies that don't come from a bank, but from funds and other private lenders. The market has grown strongly since the 2008 financial crisis and today also finances data centers for artificial intelligence.

Private equity refers to stakes in companies that are not traded on the stock exchange. Specialized funds buy such companies in whole or in part, transform them over a few years, and later sell them on at a profit.

Privilege escalation means that someone gains more rights within a computer system than they are entitled to — for instance, jumping from a regular user account to administrator status. Attackers exploit bugs in programs or misconfigured permissions to achieve this.

The Product Operating Model is a way of organizing a company: fixed teams look after a digital product on an ongoing basis instead of working through individual projects one after another. Such teams are given a goal, not a list of finished tasks.

A Product Requirements Document, or PRD for short, is a document that captures what a planned product should be able to do and who it is intended for. It deliberately describes the goal rather than the technical implementation.

Product-market fit means that a product solves a problem for which there are enough paying customers. It is regarded as the most important milestone for young companies, because growth only works reliably after it has been achieved.

A Program Bench is a fixed collection of programming tasks used to test how well a computer program or an AI system can write code. All participants receive the same tasks, so the results can be compared directly.

ProMotion is Apple's brand name for screens that refresh their image up to 120 times per second and automatically adjust that rate to the content. This makes motion smoother while also saving power when little is happening on screen.

A prompt is the input used to tell an AI system what to do – usually text, but sometimes an image as well. How good the response turns out depends heavily on how precisely and completely the prompt is formulated.

Prompt caching means that an AI text system processes the always-identical beginning of a request only once and keeps the intermediate result on hand for later requests. This significantly reduces response time and cost when many requests start with the same text.

Prompt engineering is the deliberate crafting of instructions given to an AI text system so that its responses become useful. The same question phrased in better words often delivers markedly better results.

Prompt injection is an attack on AI language programs: instructions are hidden within a text, which the program mistakes for genuine commands. This allows an attacker to make the AI break its own rules or leak data.

Prompt adherence describes how precisely an AI system sticks to what it was told to do in the input. A response can be good in terms of content and still have poor prompt adherence if it ignores requirements such as length, format, or language.

A Prompt Whisperer is someone who is especially skilled at giving text programs like ChatGPT exactly the right instructions. The term is half job title, half mockery – since how much genuine skill actually lies behind it remains disputed.

Prompt injection exfiltration refers to an attack in which hidden instructions embedded in a text or on a webpage cause an AI program to leak confidential information to the outside. The attacker does not type these commands directly into the chat window, but instead plants them in material that the program later reads.

A prompt pipeline is a fixed sequence of multiple instructions given to an AI language program, in which the output of one step becomes the input of the next. Instead of a single large question, the task is broken down into manageable sub-steps this way.

A proof of concept is a small test run meant to show whether an idea works at all. It is deliberately built quickly and incompletely before much money flows into the finished solution.

Proof of Personhood is a method used to verify online that a real human being stands behind an account — and that this person holds only a single account. The proof is designed to reveal as little as possible about that person's identity.

A proof-of-concept exploit is a small test program that proves a specific security vulnerability in software is genuinely exploitable. It usually causes no harm, but shows both vendors and attackers alike that the flaw is real.

Proprietary means: A product belongs to a company that alone decides who may use, view, or modify it. For AI models, this usually means the provider keeps the internal structure secret and only allows access through its own servers.

Proprietary software belongs to a company that keeps the program's blueprint secret and precisely dictates what users are allowed to do with it. The counter-model is open software, whose blueprint anyone can view and modify.

A proprietary model is an AI system whose inner workings the developing company keeps secret and makes usable only through paid access. Users get answers, but not the model's building blocks themselves.

Prosody is everything about spoken language that goes beyond the mere words: pitch, loudness, tempo, and pauses. In speech technology, it determines whether a computer voice sounds natural or tinny.

Protein structure prediction means calculating the spatial shape of a protein molecule from its known building instructions. Since 2020, AI programs like AlphaFold solve this task so accurately that they replace laboratory experiments in many cases.

Provenance is the documented proof of origin of a file or dataset: Who created it, when, with what, and what was later changed about it? In the AI debate, this is primarily about the question of whether an image, video, or text originates from a human or from a machine.

Provenance refers to the documented origin history of a file: who created it, with what, and what happened to it afterward. C2PA is the most widespread technical standard for this and attaches this information to images, videos, and audio files in a tamper-proof way.

Provenance means origin: for an image or video, it records who created it and what happened to it afterward. Content Credentials are a technical standard that attaches this origin information to the file in a tamper-proof way.

A proxy is a computer that sits between your device and a website and forwards requests on your behalf. The website then sees the proxy instead of you – this protects privacy, speeds up access, and is also used in training AI systems.

Pruning refers to the targeted removal of unimportant components from a fully trained AI model. This makes the model smaller and faster, while ideally preserving the same quality of answers.

A Public Benefit Corporation is a US legal business form that, according to its own charter, must pursue not only profit but also a specified benefit for the general public. The form became well known through AI companies like Anthropic and xAI as well as through OpenAI's restructuring plans.

A pull request, or PR for short, is the proposal of a change to a collaboratively developed program. Other contributors see the change, comment on it, and only merge it into the official version after review.

A pull request is the proposal to incorporate a change to a software project into the official version. Other participants read the proposal, comment on it, and decide whether it gets accepted.

PyPI is the central online repository for ready-made building blocks of code for the Python programming language. Developers upload their own code there, which everyone else can pull into their projects with a single command.

The Python Package Index, or PyPI for short, is the central online repository for ready-made add-on building blocks for the Python programming language. Anyone writing a program can download other people's code there with a single command, instead of building everything themselves.

The Python cache is a cache in which the programming language Python stores results it has already computed, so it doesn't have to redo them next time. This usually refers to the automatically created files in the __pycache__ folder, but sometimes also to custom-built caches within a program.

PyTorch is a free software library that developers use to build, train, and test AI models. It originally came from Facebook and is now the most widely used tool in AI research.

A prediction market is a trading venue where people bet real money on the outcome of future events. The price of such a contract is considered an estimate of how likely the event is to occur.

Preferred stock represents shares in a company whose holders usually have no say in company decisions, but are given priority when profits are distributed. It is the counterpart to common stock, which carries a voting right at the shareholders' meeting.

Q-Day refers to the day on which a quantum computer will first be able to break today's standard encryption methods. Nobody knows exactly when this day will come — yet banks, government agencies, and software companies are already preparing for it.

QEMU is a free program that recreates an entire computer as software. Inside it, a second operating system can be started – even one written for a completely different kind of processor.

Qi2 is the globally recognized standard for wireless charging of smartphones, in which magnets in the device and in the charging pad ensure that both align exactly with each other. It was introduced in 2023 by the industry association Wireless Power Consortium and adopts the magnet technology from Apple's MagSafe for all manufacturers.

A QR code is a black-and-white square pattern in which text or an internet address is stored. A phone camera reads the pattern in a fraction of a second and converts it back into the stored information.

A quadruped is a robot that moves on four legs, similar to a dog or a goat. Because it can overcome steps, rubble, and mud, it is primarily used for inspections in factories, construction sites, and power plants.

Quadruple patterning is a trick from chip manufacturing: instead of exposing a very fine structure directly, it is created in four steps from coarser structures. This makes it possible to fabricate circuit traces finer than the exposure machine can actually draw.

The Quake Engine is the software framework behind the 1996 game Quake. It was the first widely used software to render a game world entirely from true three-dimensional shapes, thereby shaping the technology of modern computer games.

QA stands for Quality Assurance: all systematic checks designed to ensure that a product works as promised. In the software and AI industry, QA mainly refers to testing programs and models before customers use them.

A quality gate is a defined checkpoint in a development process: only once previously agreed conditions are met may a project or piece of software move on to the next step. If something fails to meet the conditions, it is stopped and must be reworked.

Quantization means storing the many numbers inside an AI model more coarsely, for example with 8 instead of 16 digits in computer memory. This makes the model need less space and compute faster, but it becomes slightly less precise.

A quantitative hedge fund is an investment firm that hands its buy and sell decisions over to computer programs and statistical models rather than the gut instinct of individual traders. Such funds are among the largest consumers of computing power, data, and AI talent outside the tech industry.

Quantum computing is a computing technique that, instead of the usual zeros and ones, uses states that can be both at the same time. This theoretically allows certain tasks to be solved enormously faster than with today's computers — but so far only a few, and only under extreme conditions.

A qubit is the smallest unit of information in a quantum computer. Unlike a normal bit, which can only be 0 or 1, a qubit can take on both values simultaneously in proportion — making quantum computers extremely fast for certain tasks.

Query Understanding refers to all the steps a search engine takes to figure out what a user actually means by their input. From just a few typed words, it forms a structured guess about the intent behind them.

Query rewrite means that a system automatically reformulates an entered search query before using it to search. The goal is to turn a brief or unclear question into a formulation that leads to better results.

A queue is a waiting line in a computer system: tasks line up one behind another and are processed in order. This principle ensures that a system doesn't collapse under a sudden rush, but instead simply gets the work done more slowly.

A quota is a fixed upper limit on how much a single user is allowed to consume from a shared service – for example, how many requests per minute may be sent to an AI service. It protects the provider from overload and makes costs predictable.

QwenChineseBench is a test used to check how well an AI language program handles Chinese-language tasks. It belongs to the Qwen family from the Chinese corporation Alibaba and produces scores that can be compared between different programs.

Retrieval augmentation means that an AI language program looks up matching passages in a document collection before answering and incorporates these findings into its response. This allows it to also answer questions about current or company-internal content that it never encountered during training.

Response time is the span between a request and the system's answer. For AI services, it determines whether a chatbot feels smooth or sluggish — and it costs money.

A reward function is a rule that gives a computer program's behavior a score: doing well means many points, doing poorly means few. The program then keeps trying things until it collects as many points as possible — which is why the reward function ultimately decides what it actually does.

Real-time speed means that a computer program responds quickly enough to keep pace with what's happening. What matters is not the absolute speed, but whether the response arrives in time.

A recommendation system is a computer program that infers from the past behavior of many people which movies, products, or posts a single individual might like. It sits behind Netflix's home screen, TikTok's feed, and Amazon's product suggestions.

The radio spectrum is the range of electromagnetic waves over which devices exchange data wirelessly – from radio to mobile networks to WiFi. Because this range is limited, states divide it into sections and auction or allocate it to network operators.

"Recalling weights" means that a company subsequently withdraws the published number file of an AI model, for instance because it is faulty or dangerous. Because such files are copied millions of times over, such a recall almost never succeeds technically in being complete.

A rack is a standardized metal cabinet in which computers for data centers are mounted on top of one another. It supplies power, network connectivity and cooling for all installed devices and is the common unit in which AI hardware is discussed.

A rack system is a standardized metal cabinet in which computers, storage, and networking equipment are mounted one above the other. In modern AI data centers, an entire rack is now considered a single large computing unit rather than a mere shelf.

A rage click is a rapid sequence of clicks on the same spot of a website or app because nothing happens there. Analytics tools automatically detect this pattern and use it as an indicator of broken or confusing spots on a page.

The Ralph Wiggum Technique is a way of working when programming with AI: instead of engaging in a long back-and-forth discussion, you restart an AI program over and over with the same task. The name references the dim-witted Simpsons character — the method is deliberately stubborn and simple, yet in practice it works surprisingly often.

RAM is a computer's working memory: a very fast temporary storage where all the data the computer is currently working with resides. It loses its contents when powered off and is therefore not a replacement for a hard drive or SSD.

A random forest is a prediction method that combines many simple decision trees and tallies up their results. Because each tree is built slightly differently, their errors cancel each other out, making the overall prediction more reliable.

A ranking is the automatic sorting of many results according to their expected usefulness. Search engines, online shops, and social networks use it to decide what appears at the very top and what practically no one ever sees.

Ransomware is malicious software that renders a computer's files unreadable and demands money for their release. Such attacks regularly bring hospitals, government agencies, and corporations to a standstill for days or weeks.

The Rasch model is a statistical method from test theory. From correct and incorrect answers, it simultaneously estimates how difficult each task is and how able each person is – on a shared scale.

A raster image is an image made up of a fixed grid of tiny colored dots. Photos from phone cameras, screenshots, and most images on the internet are raster images – that's why they become blurry when you zoom in on them too much.

A rate cap is an upper limit on how often a user or program may use an online service within a given time period. For AI services, it typically limits the number of requests or processed text chunks per minute, hour, or month.

A rate limit is a cap on how often a user or program may request an online service within a given period of time. If the limit is exceeded, the service temporarily refuses further requests.

A rate limit defines how many requests a user or a program is allowed to send to a service within a given period of time. It protects servers from overload and, in the case of AI services, also caps costs.

RBAC is a method organizations use to define who is allowed to do what in a computer system. Permissions are not tied to individual people, but to roles such as "accounting" or "intern," to which people are then assigned.

RCS is the modern successor to classic SMS. The standard brings features like high-quality images, group chats, and read receipts directly into the phone's pre-installed messaging app.

Response time describes how long a computer system takes to answer an input. For AI services, it decides whether a chatbot feels smooth or sluggish.

Reasoning refers to an AI system's ability to think through a task in multiple intermediate steps rather than producing an answer immediately. Modern language models use extra computation time before answering — this makes them significantly more reliable at math, programming, and logic tasks, but also slower and more expensive.

A reasoning block is the separated part of an AI response in which the program spells out its solution path before stating the actual result. It serves to compute and check, and is often not shown to users at all, or only in summarized form.

Reasoning Effort is a setting in modern AI systems that determines how long and how thoroughly the program thinks before giving its answer. More effort usually means better results on hard tasks, but costs more time and money.

A reasoning config is the setting used to determine how extensively an AI model is allowed to think before giving its answer. It thus directly controls how good, how fast, and how expensive the answer turns out to be.

A reasoning engine is a computer system that draws new conclusions from known facts and rules instead of merely looking up answers. Today the term mostly refers to AI models that work through intermediate steps before answering, consuming extra compute time in the process.

A reasoning model is an AI system that first performs an extended intermediate calculation before answering, rather than responding immediately. This thinking process costs computing time, but significantly improves results on math, logic, and programming tasks.

Reasoning tokens are the text chunks an AI program consumes for its internal intermediate steps before it outputs the actual answer. They cost computing time and money, but mostly never appear on screen at all.

Recall measures what proportion of all actually existing cases a system managed to find. A spam filter with a recall of 90 percent catches nine out of ten spam emails, while the tenth lands in the inbox.

Recency weighting means giving more importance to newer data than to older data. The method helps systems adapt quickly to change, but it also makes them more susceptible to short-term outliers.

A recommender system is a program that predicts, based on a user's past behavior, what they might like next. It operates behind the homepages of streaming services, online shops, and social networks, deciding which content appears at the top.

Recursive Self-Improvement describes the idea that a computer program improves its own way of working and, as a result, can improve itself even better next time. Each improvement thus becomes the basis for the next one – in theory, always faster.

Red teaming means deliberately attacking a system to find its weaknesses before real attackers do. With AI programs, a team tries to coax the model into giving forbidden or dangerous answers.

Refactoring means cleaning up a software's program text without changing its outwardly visible behavior. Afterward, the program does exactly the same thing as before, but it is more understandable and easier to extend.

Refactoring means cleaning up the structure of a computer program without changing how it behaves for the user. The goal is more understandable, more easily modifiable program text – not a new feature.

Reflection-in-action refers to thinking in the middle of doing: you notice during a task that something isn't working out, and you change your approach right away. In AI, the term describes systems that check and correct their own intermediate steps instead of blindly executing a plan.

Refusal detection refers to methods that automatically determine whether an AI chatbot has answered a question or refused to answer it. It forms the measurement basis for how often a system blocks correctly or incorrectly.

A regex is a short search command used to search a text not for a fixed word, but for a pattern – for example, for all dates or all email addresses. Almost every programming language and many editors understand this notation.

Regional endpoints are fixed access addresses to an online service, each tied to a specific world region. Whoever uses them determines in which part of the world their data is processed and stored.

Regression is a type of task in machine learning in which a program predicts a number from example data – such as a price, a temperature, or a duration. Unlike classification into fixed categories, the result is a value on a continuous scale.

Regression to the mean describes the fact that particularly extreme measurements are usually followed by more ordinary ones. The reason is not some mysterious balancing force, but the element of chance contained in almost every measurement.

A regression test checks, after a change to a program, whether everything that worked before still works. It is meant to prevent a new feature or a bug fix from secretly breaking something else.

A regulatory sandbox is a time-limited testing environment in which companies are allowed to trial new products under the supervision of an authority — with temporarily relaxed rules. The model is well known from the financial sector, and is now also being used for Artificial Intelligence.

Regulatory capture describes the case in which a government oversight agency ends up serving the interests of the companies it is supposed to supervise instead of those of the general public. In the AI debate, the term is central because major providers themselves call for strict rules – and critics see in this an attempt to shut out competition.

Regurgitation refers to the case where an AI program outputs entire passages of text or images from its training material almost word-for-word, instead of formulating something new. Technically, this is a side effect of learning, and legally it is one of the most sensitive points in copyright lawsuits against AI companies.

Reinforcement Fine-Tuning is a method for improving an already finished AI model after the fact: it tries out answers and receives a rating for each one, which it uses to guide itself. Instead of imitating predefined model solutions, the model learns from reward and punishment.

Reinforcement learning is a learning method for computer programs in which the program learns through trial and error and is rewarded for good outcomes. It powers game-playing programs like AlphaGo and the fine-tuning of modern chatbots.

Rejection Sampling is a method in which many proposals are generated and all those that fail a check are discarded. In AI development, it is used to keep only the best out of many responses from a language model – either for output to users or as training material.

A recurrent neural network is a computer program that processes data step by step while remembering what came before. This design was long the standard for language, text, and time series, but was largely displaced by newer methods starting in 2017.

Recursive self-improvement describes the idea that an AI system improves itself, and the improved version can then improve itself even further. Each round produces a new, stronger starting point — in theory, this causes progress to keep accelerating.

A Relational Database Service is a database that you rent from a cloud provider instead of running it yourself on your own machines. The provider takes care of installation, backups, and updates, while the customer simply uses the database.

REM sleep is a recurring sleep phase in which the eyes move rapidly beneath closed lids and most vivid dreams occur. During this phase, the brain is almost as active as when awake, while the body's muscles are largely paralyzed.

Remediation refers to the targeted fixing of an identified error or security flaw in a technical system or in a business process. What is meant is not the discovery of the problem, but the work that comes after: eliminating the root cause, verifiably restoring the system to a sound state.

A remix is a new work that visibly reuses and alters an existing video, image, song, or format. In the creator economy, the remix is a central growth engine: platforms build it directly in as a button because content spreads faster through it.

A remote backup is a security copy of data that is not stored at the same location as the original, but is transferred over the internet to a remote location. This way, the data survives even if a fire, theft, or technical failure occurs at the original location.

Remote Desktop means operating a computer from another device, even though it is located somewhere else entirely. The interface of the remote machine appears on your own screen, and your mouse and keyboard control it as if you were sitting right in front of it.

Remote Development means that programmers don't run their code on their own laptop, but on a remote machine in a data center. Their own computer only displays the window, while the actual computing happens elsewhere.

Remote Proctoring refers to the supervision of exams taken at home on one's own computer. Camera, microphone and software monitor whether someone is cheating – in some cases, software handles this automatically.

Rendering means: A computer calculates a finished image from a description — such as a 3D scene or a piece of code. The term comes up in video games, animated films, websites, and increasingly in AI image generators.

A renderer is a piece of software that computes a finished image from a description – for example, a visible webpage from HTML code, or a photo-like image from a 3D model. It is the translation stage between data that a computer stores and what a human sees on the screen.

Rendering is the computation of a finished image from a description of what is supposed to be seen. In doing so, the computer transforms numerical data about shapes, colors, and light into a grid of colored pixels.

A rendering layer is a single image layer that is computed separately from the others and only merged with them at the very end to form a finished image. The principle is found in image editing, film production, map apps, and in the software that displays web pages.

A rendering engine is the part of a program that turns a description into a finished image on the screen. Well-known examples are the engines in web browsers that build a visible web page out of text files.

A repo is the central storage location of a software project: it contains all files plus the complete history of every change. This allows many people to work on the same software simultaneously without destroying each other's work.

A repository is the central storage location for all files of a software project, including its complete change history. It's the place where programming teams work together on code – the best-known provider for this is GitHub.

Representation variance describes how much the internal numerical patterns of an AI model diverge when inputs or training conditions are slightly changed. High variance means: the model describes similar things very differently internally – sometimes useful, sometimes a sign of instability.

A Request for Proposal (RFP) is a written invitation from a company or government agency to potential suppliers to submit a bid for a precisely described project. In the tech industry, such documents often determine which provider wins a multi-million-dollar software or data center contract.

A Request for Startups is a public wish list from funders: in it, they name topics for which they want to see and fund new companies. Founders thus learn in advance what money is currently available for.

Reranking is the second step of a search: an initial, fast search collects many possible hits, after which a more precise method re-sorts these hits. This way, the truly matching results end up at the top.

A research preview is an early, publicly accessible test version of an AI system that a company deliberately labels as unfinished. It serves to gather experience with real users before the product is officially completed.

A residential proxy routes internet requests through the connection of a real private individual so that they appear to the target site like an ordinary visitor. Such services are mainly used for mass web data collection and for bypassing blocks.

Residual Access refers to access rights to data or systems that someone still holds, even though they should no longer have them. Such forgotten access points are among the most common entryways for hacker attacks and data breaches.

Resolution Velocity is a customer service metric. It measures how quickly requests are actually resolved – and is now often used to demonstrate the value of AI assistants in support.

Responsible disclosure is the rule that someone who discovers a security vulnerability first reports it to the affected vendor and gives them time to fix it before the vulnerability is publicly described. It aims to prevent criminals from exploiting the flaw while no fix yet exists.

A REST API is an interface through which two programs exchange data over the internet. It follows the same rules as a web browser: one program calls an address and gets a response back.

RESTful describes a widely used blueprint for interfaces through which programs exchange data over the internet. Such an interface speaks the same language as a web browser: you call up an address and receive clearly named information in return.

Restricted Stock Units are a promise by an employer to transfer company shares to an employee at a later date. The shares don't actually belong to the employee until they have stayed with the company long enough.

A residual value guarantee is the promise that an item will still be worth a certain minimum price after a fixed period of time. Whoever gives the guarantee pays the difference if the actual resale value falls below that.

The reticle limit is the maximum area a chip fab can transfer onto a silicon wafer in a single exposure step – around 858 square millimeters. Larger chips cannot be manufactured in one piece, but only assembled from multiple sub-components.

Retrieval refers to the targeted searching out of matching text passages or documents from a large collection before an answer is generated. In AI systems, this step supplies a language model with the facts it does not reliably store itself.

Retrieval-Augmented Generation is a method in which an AI text program searches a document collection for relevant passages before answering, and uses these as the basis for its response. This allows it to answer questions about content it never saw during training.

Retry logic is a program's built-in rule to automatically try a failed request again. It determines how many times a retry happens, how long to wait in between, and when the program ultimately gives up.

A retry storm occurs when many programs simultaneously and repeatedly retry a failed request, ultimately bringing down a system that was already overloaded. This problem is one of the most common reasons why brief outages in data centers turn into hours-long downtime.

Return on Investment, or ROI for short, measures how much profit an investment generates relative to the money put in. It is expressed as a percentage and makes very different types of expenditure comparable with one another.

Reverse engineering means taking apart and analyzing a finished product to figure out how it is built and how it works. In technology, this applies to devices, programs, and increasingly AI systems whose inner workings are not disclosed.

The Reverse Information Paradox describes the effect that more available information worsens orientation instead of improving it. Because systems such as AI chatbots can generate an unlimited number of plausible-sounding texts, it becomes harder to identify what is reliable.

A revert PR is a proposed change to a software project that undoes an earlier change. It is the standard means of quickly removing a bug from a live product without first having to find its root cause.

Reward design is the art of giving learning software exactly the right points for its behavior. If the points are chosen poorly, the program finds shortcuts that raise the score while missing the actual goal.

Reward hacking means: a learning computer program finds a way to increase its score without actually accomplishing the intended task. It follows the rule literally while missing the intent.

The Riemann Hypothesis is a conjecture about the distribution of prime numbers that has remained unproven since 1859. It is considered the most famous open problem in mathematics and carries a prize of one million dollars.

RLHF is a training method in which humans rate the responses of an AI system and the system learns from these ratings. This makes chatbots like ChatGPT more polite, more helpful, and less dangerous in their responses.

RMB stands for Renminbi, the currency of the People's Republic of China; its unit of account is the yuan, its currency code CNY. Exchange rates are quoted as "1 euro = x yuan" — converting from yuan to euro therefore means dividing, not multiplying.

RNA synthesis refers to the production of RNA, the molecule cells use to pass on blueprints for proteins. It happens in every living cell and can be replicated in the lab — the foundation for mRNA vaccines and for a growing market of biotech companies.

ROAS stands for "Return on Ad Spend" and measures how much revenue a company earns for every euro of advertising budget spent. The metric is the most important control figure in online marketing and determines how modern advertising systems automatically deliver ads.

A robo-interview is a job interview in which no human sits across from the applicant, but instead a computer program asks the questions and evaluates the answers. Companies use it primarily to make an initial pre-selection when there are very many applications.

Robot-as-a-Service means a company rents robots instead of buying them and pays per month or per completed task. Maintenance, software updates and often the replacement of defective devices are included in the price.

A robotics foundation model is a large AI model that has learned from vast amounts of motion, image, and language data to control robots. Instead of being reprogrammed for every single task, such a model is meant to handle many different tasks and often various types of robots as well.

robots.txt is a small text file on a website that tells automated programs which pages they may access and which they may not. It is only a request, not a technical barrier – anyone who ignores it can still reach the content.

Rock's Law is the observation that the cost of building a factory for chip manufacturing roughly doubles every four years. It thus describes the expensive flip side of technical progress in computer chips.

ROCm is AMD's software package that allows programs to use this manufacturer's graphics cards for computing. It is AMD's counterpart to CUDA, Nvidia's equivalent software, and thus a central building block in the competition for AI hardware.

A raw event is the unaltered record of a single occurrence in a technical system, for example a click, a payment, or a sensor reading. It is stored exactly as it occurred – without summarization, correction, or evaluation.

ROI measures how much an investment yields relative to its cost. It is expressed as a percentage and, for AI projects, is the central question: does the money spent actually pay off?

A role-playing benchmark is a standardized test that checks how well a text program can sustain a given role – such as a fictional character, a customer service agent, or a teacher. What is measured is not pure factual knowledge, but whether language, character, and rules remain consistent over the course of a long conversation.

A rollback is the orderly return to an older, working state of a program or dataset after a change has caused problems. In technology, it is considered the most important emergency brake: rather than fixing the error immediately, the last good state is restored first.

Rolling Release refers to a software model in which new features are delivered continuously in small increments, rather than bundled into large versions. Users no longer install a new edition but instead keep their existing installation permanently up to date.

Root access is the highest permission level in a computer system: whoever has it may read, modify, and delete any file. It is necessary for administering a system, but is considered the greatest security risk when unauthorized parties obtain it.

Root access means that someone has all rights on a computer and may modify or delete any file. The term originates from Unix and Linux, where it denotes the most powerful user account of all.

Round-trip overhead is the time a request spends purely traveling back and forth over a network — on top of the actual computation time. In AI applications that fire off many small requests in sequence, this overhead can dominate the total duration far more than the computation itself.

A Router LLM is a system that automatically forwards each user request to the appropriate language model. Simple questions go to a small, cheap model, difficult ones to a large, expensive one — this lowers costs without noticeably degrading answer quality.

Routing refers to the automatic decision about which AI model or which part of a model handles a particular request. The goal is to complete simple tasks cheaply and quickly, and to deploy expensive computing power only where it is truly needed.

A routing platform is an intermediary layer that automatically forwards each request to the appropriate AI model. Simple questions go to a cheap model, difficult ones to an expensive one — the user notices none of this.

A routing policy is the rule that decides which AI model or building block handles a given request. It ensures that simple questions go to cheap models and difficult ones to expensive models.

Row-Level Security is a technique in databases where each user only gets to see the records that pertain to them. The rule for this sits within the database itself and applies automatically to every query.

RSPO is an international alliance of companies and environmental organizations that sets rules for growing palm oil and has compliance with these rules verified. Products bearing the RSPO seal contain palm oil produced according to these rules.

RSS is a uniform file format that allows websites to make their latest posts available for retrieval in machine-readable form. A program fetches this file at regular intervals, gathering updates from many sources in one place.

An RSS feed is a machine-readable list of a website's most recent posts. Programs fetch this list at regular intervals and display new articles without requiring a visit to the site itself.

RTF is a file format for text that stores not only the letters but also the formatting, such as bold print or font size. Microsoft developed it in the 1980s so that text documents could be exchanged between different programs.

The RTX 5060 Ti is a graphics card made by Nvidia from 2025, i.e. a component in a computer that renders images. It belongs to the lower mid-range tier and, besides gaming, is also used for small AI applications on one's own machine.

The RTX 5070 Ti is a graphics card from Nvidia released in 2025, i.e. a plug-in module for the PC that computes images. It belongs to the upper mid-range and is also used for smaller AI applications.

The RTX 5090 is Nvidia's most expensive graphics card for regular home computers, introduced in early 2025. It was designed for computer games, but is also used to run AI programs on one's own computer.

A rule file is a text file in a software project that gives an AI assistant persistent instructions – for example, which programming language applies or which style is desired. The assistant automatically reads it along with every request, so the same guidelines don't have to be repeated constantly.

The run rate extrapolates a short period of time to a full year: revenue of 10 million euros in a month results in a run rate of 120 million euros a year. It doesn't show actual earnings, but a projection assuming everything continues at the same pace.

Run-rate revenue projects a short time period out to a full year: taking one month's revenue and multiplying it by twelve gives you the annual revenue a company would achieve if it kept up the same pace. Especially among fast-growing AI companies, this projection often stands in for real annual figures.

Runaway selection describes a self-reinforcing cycle in evolution: a trait is favored in mate choice, and this preference keeps intensifying together with the trait itself. In technology and AI debates, the term now serves as an image for systems that spiral toward an extreme through their own feedback.

A runtime harness is the software environment that surrounds an AI model during its execution: it takes in inputs, hands the model tools, executes its suggestions, and collects the results. Without this scaffolding, a model could only output text but couldn't actually do anything.

Rust is a programming language used to write very fast software while ruling out a whole class of typical errors before the program even starts. It is considered a modern alternative to the older languages C and C++ and is used, among other things, in operating systems, browsers, and cloud services.

The revenue run rate projects the revenue of a short period out to a full year. It shows how much a company would earn in twelve months if the current pace continued exactly as it is.

The random baseline is the result one would achieve through pure guessing. It serves as a benchmark: a computer program has only truly learned something once it performs significantly better than this threshold.

A solid-state battery is a battery in which current doesn't travel inside through a liquid, but through a solid material. This is meant to make batteries safer and pack more energy into the same space – but mass production is still difficult.

A semiconductor index is a metric that combines the stock prices of many chip manufacturers into a single figure. The best-known is the American SOX, which investors use to gauge how well the entire chip industry is currently doing.

The sphere packing density indicates what proportion of a space is filled by equally sized spheres when they are stacked as cleverly as possible. The best value in three-dimensional space is about 74 percent, with the rest remaining unavoidable gaps.

A supply chain risk in IT security arises when attackers do not attack a company directly, but instead target one of its suppliers. Through purchased software, updates, or service providers, they thereby gain access to thousands of well-protected systems simultaneously.

Spatial primitives are simple basic shapes such as point, line, triangle, cube, or sphere, from which computers assemble complicated three-dimensional objects. They are the smallest building blocks that graphics programs, game engines, and AI models use to describe space.

The S-1 document is the detailed report that a US company must file with the American securities regulator before its shares can be traded publicly for the first time. It contains revenue, losses, ownership structures and risks – and is often the first genuine look outsiders get into the numbers of a previously closed company.

The S-1 registration is a document that a US company must file with the securities regulator before it is allowed to sell shares on the stock exchange for the first time. It discloses business figures, risks, and ownership structures, and is often the first genuine look into a company's books for journalists and investors.

The S-curve describes a typical pattern: something first grows slowly, then very quickly, and finally flattens out again toward the end. In the tech world, it is used to classify the spread of new products and the progress of technologies.

An S3 bucket is a storage area in Amazon's data center where companies store files such as images, videos, or databases. Each file gets its own internet address, through which programs can retrieve it worldwide.

SaaS pricing describes how providers of rental software structure their prices – usually as a monthly fee per user rather than a one-time purchase. The choice of pricing model often determines the success of a software company more than the technology itself.

A SAFE is a short contract that lets a young company receive money today while promising the investor company shares later in return. Instead of setting a price, it only establishes a rule for how those shares will be calculated at the next major funding round.

A Safe is a short contract that gives a young company money today and promises the investor company shares in return at a later point. The exact share is only determined at the next major funding round.

Safeguards are built-in protective measures designed to prevent an AI system from providing dangerous, illegal, or severely harmful responses. They usually consist of several layers: rules embedded during training, filters before and after the response, and controls over who is even allowed to use the system.

A safety card is an accompanying document in which a provider describes what its AI system can do, where it fails, and which risks it has tested. It functions like a package insert for software and now comes with nearly every major model release.

Same-model-pairing means that two roles in an AI system are taken on by one and the same model – for example, when a program evaluates its own answer. This saves effort, but carries the risk that its own mistakes go undetected.

SAML is a standard that lets people log in to a web application without having their own password there. Instead, a central login service run by the company confirms who they are.

A sampler is the procedure by which a language model picks one word from a list of possible next words. It determines whether an answer turns out predictable and sober or surprising and creative.

A sandbox is an isolated area within a computer where a program is allowed to run without being able to access the rest of the system. If something goes wrong there, the damage stays confined to that area.

A sandbox escape occurs when a program leaves the isolated test environment in which it was actually confined and gains access to the rest of the system. Such flaws are considered particularly severe security vulnerabilities because they breach a system's last line of defense.

A sandbox escape is the breakout of a program from the shielded area in which it is actually confined to run. Afterward, the code can access the rest of the computer – reading files, sending data, or embedding itself permanently.

SARIF is a unified file format in which analysis tools write down their findings about software. Thanks to this common format, different tools, development environments, and dashboards can read and display the same results.

A scaffold is the program built around an AI model that steers its work: it breaks tasks down into steps, gives the model tools, and checks intermediate results. With a good scaffold, the same model can achieve significantly more than without one.

Scale-to-Zero means that a program on the internet is completely shut down as soon as nobody is using it, and only starts back up on the next request. This means no costs are incurred during idle periods, but the first request after a pause is noticeably slower.

Scaled AI Content refers to texts, images, or videos that are generated with the help of Artificial Intelligence in very large quantities and largely automatically. The term comes up mainly when such content floods search engines and social networks in order to capture advertising revenue or reach.

A scaling law is a measured rule of thumb that describes how much better an AI model gets when you give it more data, more computing power, and more size. The improvement follows a surprisingly regular curve — and that curve keeps getting flatter.

A scan script is a small, self-written program that automatically searches through a large amount of data, files, or network addresses and reports conspicuous hits. It takes the tedious search work off people's hands, but it only finds what it was explicitly told to look for.

Scarcity OS is a buzzword for software that automatically distributes scarce resources such as computing chips, electricity, or storage among many users. The term does not describe a fixed technology, but rather the idea of making scarcity itself the governing principle of a system.

Scheduled Tasks are tasks that a computer or an AI assistant automatically carries out at a set time, without anyone having to start them anew each time. Typical examples are the nightly backup of data or a daily news summary in the morning.

Scheduling refers to the decision of which task is processed when and on which computing component. In computers, data centers, and AI systems, this order largely determines how fast and how cheaply everything runs.

A Schelling point is a solution that several people arrive at independently because it seems obvious to all of them – without being able to coordinate with one another. The term originates from game theory and explains why groups sometimes coordinate perfectly well without any communication at all.

Schema.org is a shared vocabulary that website operators use to label the content of their pages for machines: for instance, "this is a price," "this is a cooking recipe," "this is a date." Search engines and AI systems read out these labels instead of merely guessing at the text.

A swarm algorithm is a computational method that runs many simple search attempts simultaneously and lets them learn from one another — modeled after flocks of birds or ant trails. This way, a computer finds good solutions for problems where there are far too many possibilities to try out.

SciCode is a test that researchers use to check whether an AI program can turn genuine scientific computing problems into program code. The tasks come from physics, biology, chemistry, and mathematics and pose major challenges even for strong systems.

SCIM is a technical standard that companies use to automatically synchronize user accounts across their IT systems. When someone is newly hired or leaves the company, the connected services set up or lock the accounts themselves.

Scope refers to the area within which something applies or is visible – for example a variable in program code, a permission of an app, or the agreed extent of a project. Whoever defines the scope draws a clear line between "included" and "not included".

Scraping means that a program automatically reads out websites and transfers the content into its own data collection. This is one of the ways the huge amounts of text used to train AI systems come into being.

A scratchpad is an area where an AI language program writes down intermediate steps before giving the actual answer. This thinking out loud increases the success rate on tasks that require multiple steps.

Screen scraping refers to the automatic extraction of data from the finished, displayed interface of a program or website. A computer program takes the path a human would otherwise take with eyes and mouse — because a direct, dedicated data access point is missing.

An SDK is a ready-made toolset that a company distributes to programmers so they can connect their own programs to a particular system. It contains prefabricated code building blocks, instructions, and testing tools, saving a great deal of work.

SDR stands for Software Defined Radio: a radio device in which a program takes over the tasks that used to be handled by fixed, hard-wired components. The same device can thus receive radio broadcasts, listen in on aircraft signals, or act as a mobile network base station – depending on which software is running.

Search Engine Optimization, or SEO for short, covers all the measures used to make a website appear higher up in the results lists of search services like Google. Because the top results get the most visitors, SEO is an important economic factor for many companies.

A search box is the input field into which you type your search query on a website or in an app. It is now considered particularly valuable real estate, because it decides which service gets to answer a question – increasingly an AI instead of a classic list of results.

The 10-K is the detailed annual report that publicly traded US companies must file with the American securities regulator, the SEC, every year. It contains figures, risks, and management's explanations in a fixed structure and can be viewed by anyone free of charge.

Secret scanning is the automated search for passwords, access keys, and similar secrets that have accidentally ended up in program code. Specialized tools scan through all the files of a project for this purpose and flag suspicious strings.

A secret scanner is a program that automatically searches for accidentally published credentials such as passwords or keys. To do this, it searches through source code, files, and logs before attackers can find them.

Section 1260H is a section of US law that requires the Department of Defense to publish a list every year of Chinese companies with suspected ties to the military. Being listed does not constitute a direct trade ban, but it causes massive harm to the affected companies – especially in the chip and AI industry.

Section 230 is a section of a US law from 1996 that protects online platforms from lawsuits over content posted by their users. It is regarded as the legal foundation that allows forums, social networks, and comment sections to exist in their current form at all.

Secure Shell (SSH) is a method for logging into a remote computer over the internet in encrypted form and controlling it via text input. It is the standard way in which professionals operate servers in the data center without anyone being able to read along.

Security through obscurity means protecting something by keeping it secret rather than actually securing it. Experts consider this unreliable, because the protection collapses instantly once the secret becomes known.

A seed is the secret starting value from which a computer calculates an entire sequence of random numbers or keys. Anyone who knows the seed can recalculate everything derived from it — which is why it determines the security of encryption and crypto wallets.

A self-hosted model is an AI system that runs on a company's or individual's own computers instead of on a provider's servers. This gives the operator full control over data and costs, but they must handle the technology and maintenance themselves.

Self-domestication refers to the idea that a species has made itself tamer over very long periods of time, without anyone actively breeding it. Usually this refers to humans: those who were more peaceful and cooperative had more offspring.

Self-replication means that a system produces a functional copy of itself. In the AI debate, the question is whether a computer program can copy itself onto other machines and keep running there without human assistance.

Self-replicating means: a system independently produces functional copies of itself. In IT, the term describes malware that spreads on its own, without human intervention – and in the AI safety debate, programs that could copy themselves onto new computers.

A self-replicating worm is a piece of malware that copies itself without any human involvement and spreads from computer to computer. Unlike a virus, it needs neither a host program nor a careless click to do so.

Self-Aligned Quadruple Patterning (SAQP) is a trick from chip manufacturing that turns a single drawn line into four extremely fine lines. This makes it possible to create structures smaller than anything the exposure machine could actually image on its own.

A self-hosted deployment means that software runs on computers that the user owns themselves or rents and manages themselves – not on those of the manufacturer. With AI systems, it is mainly companies and government agencies that choose this path, because their data never leaves their own network.

Self-Modifying Agents are computer programs with artificial intelligence that can alter their own instructions or their own program code during operation. They are meant to improve without a human having to fix them up – but they are considered hard to control.

A self-service ads manager is a website or app through which advertisers can create, pay for, and evaluate their own ads without the help of a salesperson. Platforms like Meta, Google, or TikTok earn the bulk of their money through such tools.

Self-Speculation is a trick that lets a language model output its own answers faster: a slimmed-down part of the model guesses several words ahead, and the full model checks these suggestions in one go. The result stays identical, while the waiting time is often cut in half.

Selfware refers to software that writes, adapts, or repairs itself – usually with the help of AI systems that generate program code. The term is a buzzword from the tech industry and does not describe a fixed technology, but rather a direction in which programming is currently moving.

A Semantic-Acoustic Autoencoder is an AI system that translates speech recordings into a compact numerical representation and generates sound from it again. In doing so, it deliberately separates two layers: the content of what was said and the sound of the voice.

Semantic search finds texts that match the meaning of a query, even if they use completely different words. Instead of comparing letter sequences, it compares the meaning of the search query and the document.

SEO is short for Search Engine Optimization: the work of getting a website to appear as high up as possible on Google and other search engines. Sites that rank higher get noticeably more visitors – which is why SEO is a major cost factor for many companies.

Separation of Concerns is a fundamental principle in building software: each part of a program should be responsible for exactly one task. This makes large systems understandable, testable, and easier to change.

A Sequential Model is an artificial neural network whose computational steps are arranged strictly one after another in a single chain. It is the simplest form of such networks and therefore the usual starting point for beginners in programming libraries like Keras.

Serverless means that a provider fully manages the computers in the background, and developers only need to deliver their program code. Payment is only for the time the code actually runs.

Serverless computing means that a provider completely manages the computers in the data center and only runs a program when it is actually needed. Payment is based on actual usage, not for permanently reserved machines.

"Server-side" means: a computation doesn't happen on your device, but on a remote computer in a data center that's reachable via the internet. Your phone or laptop only sends the request there and then displays the result.

The Service Layer is an intermediary layer in a software system that bundles all business processes and exposes them externally as clearly named services. It separates the user interface from data storage and ensures that the same rule isn't programmed redundantly in ten different places.

A serving report is a regular evaluation of how a fully trained AI model responds to requests during live operation – with figures on response time, cost, errors, and utilization. It is the central control instrument for anyone who operates an AI system rather than merely building it.

Serving costs are the ongoing costs of keeping a finished AI model available to users and having it answer each individual request. They help determine whether an AI product makes money or burns it.

The serving layer is the part of an AI system that delivers fully trained models to real user requests. It accepts requests, distributes them across machines, retrieves the response from the model, and sends it back.

A serving stack is the software that makes a finished AI model available to users around the clock and distributes their requests to the compute chips. It determines how quickly answers arrive and how expensive each individual answer is.

A session is a coherent stretch of usage between a program and a service during which the service remembers who it is talking to. Without a session, every single request to a website or a chatbot would be completely without memory of the previous one.

Session persistence means that all requests from a user during a visit always land on the same server. This is necessary when that server holds information about the user in its own memory that other servers don't know about.

SGLang is a freely available software used to run large language models on one's own machines and deliver their responses. It is designed to handle many requests simultaneously, quickly and with low computational overhead.

SHA-256 is a computational procedure that generates a fixed sequence of 256 digits from arbitrary data – a digital fingerprint. Even a single changed digit in the input leads to a completely different result, and the input cannot be reconstructed from the result.

A shader is a small program that runs on the graphics card and calculates what a pixel or vertex of a 3D scene ultimately looks like. Because millions of these programs run simultaneously, they produce video game graphics, film effects, and today also the computing power for artificial intelligence.

A shadowban is an invisible restriction on an online platform: posts remain normally visible to the sender but are barely shown to other users anymore. Unlike a regular ban, there is no notification, which is why those affected often only notice the throttling through plummeting numbers.

Sharding means breaking up a very large body of data or a very large program into pieces and distributing those pieces across multiple computers. This makes it possible to handle tasks that would be too big for a single machine.

Share of Model measures how often a brand appears in the answers of AI chatbots when people ask them about products or recommendations. The metric transfers the old market-share logic onto a world in which purchasing decisions begin in conversation with a machine.

Shared State refers to data that multiple parts of a program can access and modify at the same time. This saves copies and keeps everyone in sync, but it easily leads to errors when two accesses coincide in time.

The shell is a program that takes typed commands and passes them on to the operating system. It is the text-based alternative to operating a computer with a mouse and windows, and the standard tool of developers, server administrators, and increasingly also of AI agents.

A shell command is a typed instruction that tells a computer directly what to do – without windows, icons, or mouse clicks. Such commands control servers, developer tools, and increasingly AI assistants that execute programs on their own.

A shell script is a text file containing a list of computer commands that the machine executes one after another. It is used to automate tasks that would otherwise have to be typed by hand each time.

A safety classifier is a small auxiliary program that sorts texts or images into categories such as "unproblematic" or "dangerous". It runs alongside an AI system and is meant to prevent problematic requests and responses from getting through.

A safety mechanism is a built-in safeguard designed to prevent an AI system from producing harmful, prohibited, or dangerous output. Such safeguards intervene at several points: during training, directly at the input, and once again at the finished response.

A Significant Risk Transfer is a transaction in which a bank passes on the default risk of part of its loans to investors in exchange for a fee, while keeping the loans themselves. This allows the bank to hold less capital against these loans and use the freed-up funds for other purposes.

SimpleQA is a test from OpenAI that checks whether an AI language program answers short factual questions correctly or makes up facts. Each question has exactly one verifiable answer, so the result can be clearly scored as right or wrong.

SimpleVQA is a test that checks whether AI systems answer questions about images truthfully. What is measured above all is how often a system claims things that cannot actually be seen in the image.

Simplex is a computational method that picks the best option out of many permissible possibilities, such as the cheapest transportation plan. In telecommunications, the same word also denotes a connection in which data flows in only one direction.

Single Sign-on (SSO) means: you log in once and can then use several different services without having to log in again. The principle is familiar from buttons like "Sign in with Google."

Site Reliability Engineering is a way of working in which software developers ensure that large online services run reliably. Instead of fixing disruptions by hand, they write programs that monitor operations and respond automatically.

Situational Awareness describes whether an AI model recognizes what it itself is and in what situation it is currently responding – for instance, whether it is being tested or is facing real users. For safety research this is a central question, because test results are only meaningful if a system does not behave differently during testing than it does in everyday use.

A scale-free architecture is a system design that works on the same principle whether the quantity involved is very small or very large. So it doesn't need to be rebuilt when ten users become ten million.

Scaling laws describe, in a computable way, how the performance of an AI system improves when it is made larger, fed with more example data, and allowed to compute for longer. They are the computational basis for why companies invest billions in ever-larger models.

Skill Recording refers to recording human work steps so that software or a robot can later carry them out on its own. Recorded clicks, hand movements, or gestures become the training material for automated processes.

A skill file is a text document with instructions that an AI assistant opens and follows when needed. This makes it possible to teach an assistant new workflows without retraining it.

SKILL.md is a simple text file in which you store a set of working instructions for an AI assistant to carry out a particular task. The assistant only loads these instructions once it actually needs them, and then works through the task according to fixed guidelines.

Skills are additional capabilities that you teach a digital assistant or an AI without changing the program itself. They work like apps on a phone: the assistant stays the same, but can suddenly do more.

Skills are clearly defined capabilities that are additionally given to an AI assistant, such as checking an invoice or entering an appointment. Instead of guessing everything itself, the program calls up the appropriate instructions along with the necessary tools for a given task.

A skills bench is a standardized collection of tasks used to test which practical abilities an AI system actually masters. Rather than simply querying factual knowledge, it measures whether the system can complete a task correctly from start to finish.

An SLA is a written part of a contract in which a provider commits to the minimum quality its service must have – for example, how rarely it may fail. If the commitment is broken, there is usually a refund.

A slug is the readable part of an internet address that names a single page – for example "artificial-intelligence-explained" at the end of a link. It usually consists only of lowercase letters, numbers, and hyphens, and is generated automatically from the title.

SMTP is the rule language computers on the internet use to pass on emails. It only takes care of delivery from the sender to the recipient's mailbox, not of reading the message.

SOC 2 is an audit report in which an independent accounting firm confirms that a software provider properly protects its customers' data. For companies that sell software to other companies, such a report is often the ticket to sales conversations.

Social engineering refers to attacks in which criminals don't outsmart technology, but rather people. They deceive their victims so that they voluntarily hand over passwords, transfer money, or open a dangerous file.

Social manipulation refers to the deliberate attempt to influence the opinions, feelings, or decisions of many people without them noticing. With AI tools, this can now be done at scale and very cheaply.

Social Scoring refers to systems that record people's behavior over extended periods of time and calculate an overall assessment of them based on that data. In the European Union, this type of assessment by authorities and companies has been largely banned since 2024.

Social Simulacra are artificially generated user groups that write and react in a still-empty online community the way real people would. Developers use them to test a new forum or a new app before anyone has even signed up.

Society of Thought refers to an approach in which a task is not solved by a single answer, but by several parallel lines of reasoning that check and complement one another. The interplay of these voices produces a result that is usually more reliable than any single attempt.

Soficity is a property of mathematical objects called groups: a group is sofic if its computational behavior can be approximated to any desired precision by finite tables. The concept plays a role in mathematics because many major conjectures have been proven for sofic groups – and no one knows whether non-sofic groups exist at all.

Soft labels are training targets that assign an example not just a single correct answer, but different probabilities across several possible answers. Instead of "this is a husky," the target becomes "80 percent husky, 15 percent wolf, 5 percent German shepherd."

The soft-money compute strategy describes how AI companies obtain computing power without paying for it with their own money: instead of cash, company shares, credits, or discounts flow. This allows young companies to use huge computing capacities, but ties them closely to their financial backer.

Software as a Service (SaaS) means: A program runs on the provider's computers, and customers use it over the internet for an ongoing fee. So you no longer buy the software, but rent access to it.

Software as a Service, or SaaS for short, means: a program doesn't run on your own device, but on a provider's computers, and you use it via the browser for an ongoing fee. Instead of buying a version once, you rent the usage – usually monthly or annually.

A software stack is the entirety of programs that build on top of one another and work together within a technical system – from controlling the hardware to the interface the user sees. The term often comes up when companies explain which building blocks a product is made of.

A SOL API is a technical interface through which programs communicate with the Solana network: they query account balances, send out payments, or read price data. Developers use it so their app doesn't have to run a complete network node itself.

The Solow paradox describes the observation that companies invested a great deal of money in computers without the economy producing measurably more output per hour worked. The term is resurfacing today whenever people argue about the economic impact of artificial intelligence.

SOUL.md is a text file in a software project that records what the project is actually about: purpose, values, tone, and boundaries. It is intended primarily for AI assistants working on the code, who would otherwise only know the technical rules.

A source map is a supplementary file that traces machine-compressed program code back to the code as it was originally written. Developers can use it to look for errors at the exact spot where they actually typed the code.

Source poisoning means that someone deliberately places false or manipulated texts on the internet so that an AI system later uses them as a source of knowledge. The AI then passes on the manipulation as a credible answer, without noticing the deception.

Sovereign AI means that a country or region controls its own artificial intelligence: the data centers, the data, and the programs. The goal is to avoid being dependent on a handful of foreign corporations.

The sovereignty claim in AI is the demand by a state or alliance of states to be able to build, operate, and control artificial intelligence itself, rather than depending on foreign providers. Behind this lie concerns about security, the economy, and the question of whose rules apply to a technology.

A spacer GIF is a tiny, invisible image that early web designers embedded in pages to force spacing between text and graphics. Today the technique is considered obsolete, but it still turns up in old code and in emails.

A sparkline is a very small chart without axes or labels that sits directly within a line of text or a spreadsheet cell. It shows at a glance whether a value has risen, fallen, or stayed the same over time.

Sparse Attention is a cost-saving technique in modern language models: instead of comparing every word in a text with every other word, only selected word pairs are considered. This drastically reduces computational effort for long texts and makes inputs of hundreds of thousands of words affordable in the first place.

Sparse Mixture of Experts is a construction principle for large AI models: the model consists of many interchangeable subnetworks, of which only a few compute per request. This allows a model to be very large while still responding quickly and cheaply.

Sparse MoE is a design for large AI models in which only a small part of the model computes per request while the large remainder stays idle. This allows models to store a great deal of knowledge without every answer becoming correspondingly expensive.

A sparse architecture is an AI model that, for each individual request, uses only a small portion of its building blocks instead of all of them. This allows a very large model to compute as if it were much smaller.

Spatial computing refers to technology that measures the real space around a person and places digital content precisely within it. Instead of looking at a flat screen, one moves among virtual objects that appear to stand in the room.

Spec-Driven Development is a way of working in programming where a precise written description of the desired outcome is created first, and the actual program code is only derived from it afterward. The method has become especially important since AI tools started writing large amounts of code and need clear instructions to do so.

The spec phase is the part of a project where it is precisely written down what is to be built – before anyone starts building it. The result is a document against which everyone involved will later be measured.

Speculative Decoding is a trick that lets language programs produce their answers faster. A small, fast model guesses several words in advance, and a large, accurate model checks these suggestions all at once.

Speculative execution means that a computer starts calculating before it's even clear whether the result will be needed at all. If it guesses right, it saves time; if it guesses wrong, it discards the work — this principle is built into almost every processor and today also drives the acceleration of AI text models.

A speech engine is a program that converts spoken language into text or generates audible speech from text. It powers voice assistants, subtitle features, text-to-speech functions, and dictation software.

Speech-Native refers to AI systems that process spoken language directly and output speech directly again, without taking a detour through written text. This allows them to respond faster and also pick up on tone of voice, volume, and pauses.

Speech-to-text refers to software that converts spoken language into written text. It powers dictation features, automatic captions, and voice assistants, and is today one of the most reliable applications of artificial intelligence.

SPEED-Bench is a testing procedure used to measure how fast and efficiently an AI system generates answers – not just how good those answers are in terms of content. Such tests are important because speed and power consumption determine whether an AI service remains affordable.

A specification language is an artificial language with fixed rules used to write down what a computer program should accomplish – not how it does it. Because such descriptions are unambiguous, a computer can check them and uncover contradictions.

Sphere packing is the question of how to stack equally sized spheres as densely as possible into a space without them overlapping. It sounds like something for fruit crates, but it's embedded in radio transmission, data compression, and modern encryption.

Split screen refers to a divided display: the screen area is broken up into two or more sections that simultaneously show different content. The term originates from film and video games and is nowadays primarily an operating feature of computers, tablets, and smartphones.

A speech-to-speech model takes in spoken language and responds directly with spoken language, without taking the detour through written text. This lets it react faster and allows it to process tone of voice, volume, and speaking pace along the way.

SQL injection is an attack on websites in which someone secretly sends commands to the underlying database through a normal input field. If successful, the attacker can read, alter, or delete other people's data.

SQLite is a free software package that lets an app store and search its data in an orderly way in a single file on the device. It is invisibly built into almost every smartphone, browser, and operating system, and is considered one of the most widely deployed pieces of software in the world.

SRAM is a particularly fast type of computer memory that only holds data as long as power is flowing. It sits directly on processors and graphics chips, where it plays a key role in determining how fast AI calculations run.

SSH is a method for securely controlling a remote computer over the internet. Everything transmitted is encrypted, so that outsiders on the network cannot read along.

SSH Bootstrap refers to the initial setup of a new server via an encrypted remote connection. A single access point is enough to automatically configure basic settings, user accounts, and management tools.

A stablecoin is digital money on a blockchain whose price is pegged to an ordinary currency such as the US dollar. One unit is meant to be worth about one dollar permanently — unlike Bitcoin, whose price fluctuates heavily.

StackPerf refers to measuring the performance of an entire technology stack rather than individual components: it counts what ultimately reaches the user, not how fast a chip is on paper. Such holistic assessments today determine which AI systems are truly affordable and fast in actual operation.

The Standard Model is currently the best description of the smallest building blocks that make up matter and the forces that act between them. It explains almost all particle physics measurements with great precision, but leaves questions such as gravity and dark matter unanswered.

State is everything a program remembers about what has already happened while it is running. Whether a system has state or not determines whether it can form a connection out of a sequence of requests.

State Corruption refers to the condition of a program or AI system in which the internally stored data no longer matches reality. The system then appears to keep working normally, but bases its decisions on a false foundation.

A state machine is a model that divides a system into a manageable number of clearly named states and defines which events lead from which state to which other state. It is used in software development to describe processes unambiguously and in a verifiable way.

The state-space architecture is a blueprint for AI models that read text or signals step by step while carrying along only a compact summary of what has been read so far. It is considered a counter-design to today's common Transformer approach, because it handles very long inputs far more economically.

A state-space model is an AI blueprint that reads text or signals step by step while carrying along only a compact summary of what has come before. This keeps computational cost manageable even for very long inputs — unlike today's common transformer models.

Stateless means: a service remembers nothing about the user between two requests. Every request has to bring along all the necessary information itself, which makes the underlying technology simpler and easier to scale.

Stateless Agents are AI programs that independently work through tasks but remember nothing between two assignments. Every task starts from zero, and all necessary information must be supplied from the outside.

Static Application Security Testing, or SAST for short, looks for security vulnerabilities directly in a software's program text, without requiring the program to run. Special testing tools read the code like a proofreader and flag spots that attackers could exploit.

Static code analysis checks written program text for errors without running the program. Tools read the text like a proofreader and flag suspicious spots before the software reaches users.

The status quo bias is people's tendency to stick with the existing state of affairs, even when an alternative would be measurably better. It explains why software defaults, old contracts, and entrenched workflows survive so stubbornly.

Steganography is the art of hiding a message so that no one notices its existence. Unlike encryption, where an unreadable message is openly visible, steganography shows only a harmless image, video, or ordinary text.

Stylometry is the measurable study of writing style: it counts linguistic features of a text to find out who wrote it. Today it is used, among other things, to attribute anonymous texts or to distinguish machine-generated texts from human ones.

Stop words are very common, individually low-content words such as "and," "the," or "is" that some text processing programs strip out of a text before analysis. This saves memory and processing time, but can distort meaning — which is why modern language models mostly do without it.

A stop hook is a stored instruction that is automatically executed as soon as an AI assistant finishes its work on a task. It can be used to trigger completion checks, cleanup work, or notifications without a human having to initiate them every time.

Streaming means that a chatbot outputs its answer piece by piece while it is still computing it, instead of showing everything at once at the end. That's why the text appears to be typed out on screen.

The streetlight effect describes the tendency to search where searching is easy, rather than where the answer actually lies. In research, business, and AI, this leads to the study of easily measurable things while important but hard-to-measure questions are neglected.

The Streisand effect describes how attempting to suppress a piece of information often ends up making it truly well known. It is named after the singer Barbra Streisand, who sued in 2003 over an aerial photo of her house and thereby spread it to millions of viewers.

A sub-agent is an AI helper launched by an overarching AI program to handle a clearly delimited subtask, after which it reports back only its result. This breaks large tasks down into smaller portions that can be worked on separately and, in some cases, in parallel.

Subliminal learning describes the effect that an AI system, when learning from the texts of another system, also adopts that system's quirks – even if these never appear in the texts at all. The term originates from research papers from 2025 and is considered a risk for the widespread practice of having AI training data generated by other AI.

Subquadratic attention refers to methods that make the computational cost of language models grow more slowly than quadratically when processing long texts. Instead of comparing every word with every other word, the number of comparisons is deliberately reduced.

Subvocalization is the silent mouthing of words while reading or thinking, during which the speech muscles perform tiny movements. Sensors can measure these signals, which is why the term now also appears in technology — for instance in devices that recognize silently spoken commands.

Search engine optimization, or SEO for short, encompasses all measures used to make a website appear higher up in the results list on Google and other search services. The goal is more visitors without paying for advertising space.

The sunk cost fallacy is a thinking error: people stick with something because they've already put a lot of money, time, or effort into it. The rational approach would be to decide based only on future benefit, since what has already been spent won't come back regardless.

A super-agent is an AI system that doesn't just answer but independently completes entire tasks: planning, operating programs, checking intermediate results. The term is more a marketing promise than a fixed technical definition.

A super app is a single mobile program that brings together many otherwise separate services: messaging, payments, food ordering, ticket purchases. The best-known example is WeChat in China, where hundreds of millions of people handle almost their entire digital daily life within one app.

A supercomputer cluster is a network of thousands of individual computers connected via extremely fast cables that work together on a single task. Such facilities are today the foundation for training large AI models and for elaborate scientific simulations.

SuperGPQA is a test comprising around 26,000 highly difficult expert-level questions from 285 academic disciplines, used to assess how much genuine expert knowledge an AI system truly possesses. It is regarded as the successor to the smaller GPQA test and also covers rare fields of study that are missing from typical exams.

Superintelligence refers to a hypothetical computer system that would clearly surpass humans in virtually all intellectual tasks. The term does not describe existing technology, but a future scenario over which research, politics, and investors argue fiercely.

A SuperPOD is a ready-assembled large-scale system made up of many interconnected specialized computers, sold by chipmaker Nvidia as a complete package. It is used to train very large AI models in weeks instead of years.

Supervised Fine-Tuning is a training step in which an already pretrained AI model is further trained using curated example pairs of task and sample answer. This is how a system that can only continue text actually learns to follow instructions.

In a supply chain attack, criminals don't attack a company directly, but instead manipulate a supplier whose product the company uses. Through a single infected update, thousands of customers can be affected at once.

A supply-side platform is a technical service through which website and app operators automatically offer their advertising space for sale. In a fraction of a second, it collects bids from many ad buyers and awards the space to the highest one.

A surrogate verifier is an auxiliary program that evaluates the answer of an AI system even though the correct solution is not actually known. It replaces a genuine check with an estimate – which makes it fast, but vulnerable to exploitation.

SVG is a file format for images that doesn't store the graphic as a grid of dots, but as a set of construction instructions made of lines, curves, and shapes. This means an SVG stays razor-sharp at any size and can be edited like text.

SWA-based HiCache is a technique that lets chatbot servers remember intermediate results from earlier computation steps and store them across multiple memory tiers. Because the model only considers a sliding window of text, the amount of stored data stays bounded even for very long conversations.

SWE Marathon is a test that checks whether an AI system can work on a programming task for a very long time without losing track. Instead of small individual fixes, it involves assignments that require hours or days of work on real program code.

SWE-bench is a test used to check whether an AI system can independently fix real programming bugs in existing software. The tasks come from real software projects, and a solution only counts if it passes the project's existing test suites.

SWE-Bench Pro is a test that checks whether an AI can independently fix real programming bugs in real software projects. It is considered a significantly harder evolution of the older SWE-Bench test.

SWE-bench Verified is a test that measures how well computer programs can fix real bugs in real software on their own. It consists of 500 hand-checked tasks from openly accessible programming projects and is regarded as the leading benchmark for the coding abilities of AI systems.

SWE-Pro is a test used to evaluate how well AI systems solve real programming tasks from professional software projects. Instead of small exercise tasks, the AI receives real bug reports and must modify the code so that the project's test programs pass.

Swift is a programming language that Apple introduced in 2014. It is used to write most apps for iPhone, iPad, Mac, and Apple Watch.

Sycophancy refers to the tendency of chatbots to tell users what they want to hear instead of sticking to the factually correct answer. This behavior arises unintentionally during training, because humans tend to rate agreeable answers more highly than correct ones.

Sycophantic AI describes chatbots' tendency to tell users what they want to hear instead of sticking to the correct answer. If you contradict the program, it often backs down – even when it was right before.

In symmetric encryption, sender and recipient use the same secret key: once to scramble a message into unreadable form, and once to turn it back. The method is very fast, but it has one problem — the key must first reach the other party securely.

Synthetic data are artificially generated examples meant to replace real measurements or real texts. They are used primarily to train computer programs when real data is missing, too expensive, or would violate personal privacy rights.

SynthID is a technique developed by Google that embeds machine-generated images, audio, videos, and text with an invisible identifying mark. A matching verification tool can later read this mark and thereby indicate that the content originated from a machine.

A syscall is a program's request to the operating system to do something the program isn't allowed to do itself – such as opening a file or sending data over the network. Practically all access to disk, network, and memory runs through this controlled interface.

A system card is a public accompanying document that a company publishes for a new AI system. It describes what the system can do, where it fails, how it was tested, and what safety measures have been built in.

A System of Action is software that doesn't just display work steps but actually carries them out itself – for example, placing an order or rescheduling an appointment. The term distinguishes such programs from classic software, which primarily stores and displays data.

A system of record is the computer system that is considered the single authoritative source for a particular piece of information. If two programs disagree, what is stored there counts as correct.

A system of record is the software that holds the binding, authoritative version of certain company data. If two programs disagree, whatever is stored here is what counts.

A system card is an accompanying document that a company publishes for a new AI product. It states what the system is intended for, how well it performed in tests, and where it typically makes mistakes.

A system prompt is a hidden instruction that a chatbot operator gives its program before the user has even typed anything. It determines how the program behaves, what tone it strikes, and what topics it should not discuss.

Systemic risks are dangers that can shake not just individual people, but entire societies, markets, or public order. In AI regulation, the term refers to the risks posed by especially powerful models, for whose providers stricter obligations therefore apply.

A sextillion is a one followed by 21 zeros, in other words a thousand quintillions. The term comes up when dealing with extremely large quantities – for example, the computing operations of supercomputers or vast amounts of data.

A special purpose vehicle is a separate company that is founded solely for a single, precisely defined task – such as building a data center. It has its own debts and its own assets, which keeps the risks separate from the parent company.

A tabular foundation model is a large, pre-trained AI model for data in table form, such as customer lists or measurement series. It can make predictions for a new table without being specially retrained for it.

A tabular language model is an AI system that processes not just running text, but specifically tables with rows, columns, and headers. It can answer questions about table content, look up values, and perform simple calculations across cells.

A Tabular Foundation Model is an AI model specialized in tables with rows and columns, predicting values within them. It was pretrained once at large scale and can then evaluate new tables without being retrained for each individual table.

Tail latency refers to the response time of the slowest requests in a system – not the average, but the outliers at the upper end. It determines how reliable a service feels to users.

The tape-out is the moment when the finished design of a computer chip is sent to the factory that manufactures it. After that, nothing in the design can be changed anymore, which is why this step is regarded as a milestone in the chip industry.

A task agent is a computer program that independently breaks down a task into individual steps and carries out these steps on its own, one after another. It doesn't just respond with text — it acts: it searches, clicks, calculates, and checks the result.

The task completion rate indicates what proportion of assigned tasks were completed fully and correctly. It is one of the most important metrics for judging whether an AI system is useful in real-world deployment.

tasks.md is a simple text file in a software project that lists the upcoming work steps. It has become a fixed building block of collaboration with AI coding assistants, because they can read the list, work through it, and check off completed items.

The Tau Scaling Law describes how reliably an AI system completes a task depending on how long that task takes. The success rate drops in a regular pattern the more work steps are required in sequence.

TCP/IP is the set of shared rules by which computers exchange data over the internet. It ensures that data is broken down into small packets, delivered correctly, and reassembled properly at the receiving end.

A tech stack is the collection of all programs, programming languages, and services a company uses to build and run its software. So the term doesn't describe a single tool, but rather the entire selection of building blocks that must work together.

A Technical Spec is a written document that lays out, before any coding begins, how a planned piece of software is supposed to be built. It describes the architecture, the decisions made, and the reasons behind them — serving as a shared point of reference for the team.

Technical debt is the follow-up cost of quick, unclean solutions in software: you save time today and pay later with more effort. The term compares this effort to a loan that costs interest.

The technological singularity is the idea that machines will one day become smarter than humans and then continue to improve themselves ever further. From that point on, further development would no longer be predictable or controllable for humans.

The technological singularity is the idea that machines will eventually become smarter than humans and then continue improving themselves. From that point on, development would happen so fast that humans could no longer predict or control it.

Technological sovereignty means that a country or a company masters key technology itself, or at least can freely choose whom it obtains it from. It is not about isolation, but about not being open to blackmail.

Telemetry refers to the automatic measurement and transmission of operating data from a device or program to the manufacturer. It shows how software runs, where it crashes, and how it is used – and is therefore controversial among privacy advocates.

Temperature is a setting that controls how predictable or how surprising the responses of a text generator turn out to be. A low value leads to sober, repeatable phrasing, a high value to more unusual and riskier ones.

A temperature sensor is a small component that measures how warm or cold its surroundings are and passes this value on as an electrical signal. It is found in phones, cars, heaters, and in the data centers where AI models operate.

In a template injection, an attacker smuggles their own computational instructions into a text template of a website. The server mistakes them for a normal part of the template, executes them, and thereby leaks data or surrenders control over the machine.

A Tensor Processing Unit (TPU) is a specialized chip developed by Google that performs almost exclusively one type of calculation: the mass multiplication and addition of numbers arranged in large tables. This is exactly the kind of calculation modern AI systems constantly need, which is why TPUs handle it faster and with less power than a regular computer chip.

TensorRT is a software package from Nvidia that reworks fully trained AI models so they run as fast as possible on Nvidia graphics cards. It doesn't make models smarter, just faster and more efficient to operate.

A terabyte is a unit of measurement for amounts of data and corresponds to roughly one trillion characters. Today it is the typical order of magnitude for hard drives, video archives, and the datasets used to train AI models.

A term sheet is a short summary of the key conditions under which an investor intends to put money into a company. It is not yet a final contract, but it establishes what will no longer be up for negotiation later on.

A terminal is a window on a computer where you type commands as text instead of clicking on icons. The computer executes the commands and also responds with text.

A Terminal User Interface (TUI) is an interface made entirely of text within a terminal window – with menus, lists, and progress bars, but without a mouse, windows, or graphics. Developer tools and AI coding assistants use this form because it is fast, resource-efficient, and usable even via remote access to distant machines.

Terminal-based means: a program is not operated via windows, icons, and mouse, but via typed commands in a text window. The output likewise appears as text, line by line.

Terminal-Bench is a test that checks how well AI programs solve real tasks on a computer by typing commands into a text window. What's measured is only whether the end result actually works.

Test-Time Reasoning means that an AI program deliberately takes more computing time for a question and works through intermediate steps instead of answering immediately. This extra thinking time particularly improves math, logic, and coding tasks, but costs significantly more per answer.

Text-to-Speech refers to software that converts written text into spoken language. Modern systems barely sound like computer voices anymore and can mimic emphasis, pacing, and even specific voices.

Text-to-Speech, or TTS for short, refers to programs that convert written text into spoken language. Modern TTS systems sound so natural that they are barely distinguishable from real voice recordings.

Text-to-Video refers to computer programs that generate a short video from a written description. You type a sentence, the program computes, and delivers moving images that never existed anywhere before.

A text-to-image model is a computer program that generates a matching image from a written description. Well-known examples are Midjourney, Stable Diffusion, and DALL·E.

Text generation refers to the automatic creation of coherent text by a computer program. Modern AI programs write word by word, each time choosing the continuation that best fits the text so far.

A thin client is a very sparsely equipped computer workstation that hardly does any computing itself. The actual work is performed by a central computer on the network, while the device merely displays the result and sends back keyboard and mouse input.

A thinking budget is an upper limit on how long an AI program is allowed to think internally before giving its answer. It controls the trade-off between answer quality on one side and latency and cost on the other.

Thinking mode is an operating mode of AI chat programs in which the program first writes a longer intermediate text for itself before producing the actual answer. This intermediate step costs time and money, but noticeably improves accuracy on tasks that require multiple reasoning steps.

Threat hunting is the active search for attackers who are already lurking undetected in a computer network. Instead of waiting for an alert, professionals deliberately comb through network data for traces.

Threat modeling is a structured thinking exercise: before building a system, a team considers who might attack it, how, and to what end. From this list of possible attacks, it derives which protective measures are truly worthwhile.

A Time Series Foundation Model is a large AI model trained on vast amounts of measurement sequences that can subsequently be used for many different forecasts. It aims to do for time series what language models do for text: pre-trained once, then immediately applicable to new data.

Time-to-First-Token refers to the waiting period between sending a request to an AI system and the appearance of the very first character of the response. It determines whether a chatbot feels smooth or sluggish.

A timeout is a fixed waiting period set in a computer program: if an expected response does not arrive within this time, the program aborts the attempt and reports an error. This prevents a service from waiting endlessly for another one.

tmux is a command-line utility that splits a single text window into several independent workspaces. Its key advantage: running programs stay active even if you close the window or lose the connection to the server.

A token is a small piece of text – usually a word, part of a word, or a punctuation mark – into which language programs break down text before processing it. Because providers charge per token and models can only take in a limited number of them, the term is the central unit of measurement in the AI industry.

Token arbitrage refers to exploiting price differences between AI services that bill by processed text unit. Whoever buys requests cheaply and resells them at a higher price or redirects them to a cheaper model earns the difference.

Token-based costs are the standard billing model for AI language services: you pay according to the amount of text a program sends in and gets back, measured in small text units. Anyone who sends a lot of text or generates long responses pays correspondingly more.

A token budget is the fixed upper limit on text units that an AI system is allowed to process or generate for a given task. It caps costs, response time, and length – forcing developers to carefully choose what the model even gets to see.

Token efficiency describes how much useful output an AI system delivers per processed text unit. Anyone who achieves the same answer quality with fewer text units works more cheaply and faster.

A token quota is the amount of text chunks that a user or program is allowed to send to or receive from an AI system within a certain period. It limits costs and computational load and is the standard billing unit for commercial AI services.

The token limit indicates how much text an AI language program can process at once. It is not measured in words, but in tokens – small text building blocks into which the program breaks down every text.

The token margin describes how much money remains from what an AI provider earns for a sold unit of text after paying the compute costs for it. It is the central metric for whether the business of generating AI responses is profitable or generating losses.

A token plan is a pricing model for AI services in which you don't pay for time or the number of requests, but for the amount of text units processed. These units are called tokens, and the longer the question and answer, the more of them you consume.

The token price indicates what a provider charges for processing a certain amount of text chunks through an AI model. It is the standard billing unit for language AI and determines whether a product can be operated economically.

Token prices indicate what it costs to use an AI language model per unit of processed text. Billing is usually split between input and output text, typically in dollars per million tokens.

Token routing refers to the decision as to which part of a large AI model should process a given piece of text. A small auxiliary network selects a few specialized building blocks for each piece of text, rather than having the entire model perform the computation.

A tokenizer breaks text down into small building blocks and replaces each block with a number. Only this way can an AI language model compute language at all, since it processes nothing but numbers.

Tokens per second indicates how many text chunks an AI program outputs or processes in one second. This value is the most important speed metric for chatbots and language models – similar to km/h for a car.

Tool Calling means that a language model doesn't just write text, but specifically calls external programs — such as a calculator, a search engine, or calendar software. The model itself decides when it needs such a tool, and incorporates the result into its answer.

Tool Poisoning is an attack in which the description of a tool that an AI assistant is permitted to use is secretly rigged with hidden instructions. The AI reads these instructions as part of its task and carries them out without the user noticing anything.

Tool Use means that an AI program doesn't just write text, but additionally calls on external helper programs — such as a calculator, a search engine, or a database. The AI itself recognizes when it needs such help and incorporates the result into its response.

A tool call means that an AI program does not solve a task itself, but instead uses an external aid — such as a calculator, a search engine, or a weather database. The program formulates a precisely defined request for it, receives a result back, and uses it to continue writing its answer.

Tool fatigue describes the state of exhaustion that arises when people have to use too many different programs in parallel at work. Instead of saving time, the multitude of applications costs attention, nerves, and money.

ToolcallFormatIFBench is a test that checks whether an AI language model calls external tools exactly in the prescribed format. What is measured is not whether the response is substantively clever, but whether the technical form is correct.

A toolchain is a fixed sequence of programs that work on the same task one after another, each passing its results on to the next program. The term comes from software development and today also describes the tool chains used to build and operate AI models.

Toolflation describes how companies and individuals accumulate more and more digital programs and services without actually using all of them. The costs and effort involved grow faster than the benefit.

The Top500 list is a ranking published twice a year of the 500 fastest publicly known supercomputers in the world. Computing performance is measured using a standardized test, the Linpack benchmark, whose result is given in computing operations per second.

TOPS stands for "Tera Operations Per Second" and indicates how many trillions of simple computing steps a chip can perform per second. The figure is the go-to marketing metric for AI chips in phones, laptops, and cars – but it's only meaningful with additional context.

The Tor network is a worldwide network of voluntarily operated computers that reroutes internet traffic through multiple intermediate stations, encrypting it multiple times along the way. This makes it nearly impossible to trace who visited which site.

Total Cost of Ownership (TCO) refers to all costs an acquisition incurs over its entire period of use – not just the purchase price. In IT and with AI systems, this calculation often determines which solution is truly the cheaper one.

A touch interface is a user interface controlled directly with a finger on the screen, instead of using a mouse or buttons. It's found in smartphones, tablets, ticket machines and cars, and has changed the way people operate devices.

A TPU is a specialized chip developed by Google that can do almost only one thing: perform the massive number calculations that make up artificial intelligence. It is thus an alternative to Nvidia's graphics chips, which have so far dominated the market.

Tracing means recording the path of a single request through a computer system seamlessly from start to finish. This lets developers see how long each work step took and at which point something went wrong.

Track 1.5 refers to rounds of talks between states in which government representatives and independent experts participate together. In AI policy, this semi-official format serves to discuss sensitive topics such as military AI without a state having to commit itself immediately.

Track Changes is a feature in word processing programs that visibly marks every change made to a document instead of simply applying it. Anyone reading the text later can see exactly who deleted, inserted, or commented on what, and can accept or reject each change individually.

A tracking pixel is a tiny, usually invisible image embedded in a webpage or email whose retrieval tells the sender that and when someone opened the page or message. It is one of the oldest and most widely used tools for measuring user behavior on the internet.

Traffic refers to the volume of visitors or data reaching a website, an app, or a network within a given period. This figure helps determine advertising revenue, server costs, and the value of a digital offering.

Traffic classes are categories into which a network divides passing data in order to forward it at different speeds. A video call is thus given preferential treatment, while a large download can afford to wait a little.

A training corpus is the collected body of data from which an AI system learns – usually text, images, or code. Its size, composition, and quality largely determine what the finished system can do and what mistakes it makes.

A training repository is the central, organized storage location for all the data used to train an AI model. It records which examples are included, where they come from, and which version was used for which training run.

A training cluster is a large collection of specialized computers connected together to let an AI system learn from vast amounts of data. Modern clusters consist of tens of thousands of such computing chips and cost billions.

Training data are the examples from which an AI program learns its abilities – for instance, billions of texts from the internet or millions of labeled photos. Their quantity and quality help determine what the finished program can do and what mistakes it makes.

A training dataset is the collection of examples from which a computer program derives its behavior, rather than having it prescribed. Its size, composition, and quality determine what the finished program can later do — and which errors it makes systematically.

A training corpus is the collected body of data from which an AI system learns – usually vast amounts of text, images, or code. What it contains determines what the finished system can do and what mistakes it makes.

A training run is the continuous computational process in which an AI model learns from example data. Large runs occupy thousands of specialized chips for weeks and often cost hundreds of millions of dollars.

Training material is the collection of texts, images, sounds, or tables from which an AI system learns. Its quantity, quality, and origin help determine what a model can later do — and where it makes mistakes.

A training environment is a self-contained, usually simulated setting in which a computer program is allowed to attempt a task over and over again and receives an evaluation for each attempt. From many millions of such attempts, the program learns which actions pay off.

Trainium is a specialized chip that Amazon developed in-house to train AI models. It is the company's own alternative to Nvidia's expensive computing chips and runs exclusively in Amazon's data centers.

A trajectory is the complete record of a process: every state, every decision, and every outcome, step by step in the correct order. In AI development, such logs of a process are the raw material from which systems learn how to solve multi-step tasks.

A trajectory is the complete record of an attempt to solve a task: every step that a program or a human took from the task to the result. Such records are today the most important learning material when AI systems are meant to master multi-step tasks.

A transducer is an AI model that converts a running input into an output piece by piece, without waiting for the input to end. It is best known in speech recognition: it continuously turns audio recordings into text while someone is still speaking.

The Transformer is the blueprint on which nearly all large AI language models are built today. Its core trick: for every word, the model looks at which other words in the text currently matter to it.

Transcription means converting spoken language into written text. Computer programs handle this in seconds today, providing the foundation for subtitles, meeting minutes, and voice assistants.

Tree Shaking is an automatic process that, when assembling a website or app, removes all program code that nobody uses. The result is a smaller file that loads faster.

Triage in IT security is the rapid sorting of incoming alerts and reported vulnerabilities by urgency. It determines which case is handled immediately, which can wait, and which turns out to be a false alarm.

Trinity is a name that has been used multiple times in the world of technology: best known are the first test of an atomic bomb in 1945 and a supercomputer at the US Los Alamos laboratory. In AI reporting, the term usually appears as a project or product name and must therefore always be read in context.

The Trust Flywheel describes a self-reinforcing cycle: those who trust a product use it more, and this usage in turn makes the product better and more trustworthy. In the AI industry, the term is seen as an explanation for why some providers keep extending their lead further and further.

A trust score is a number that indicates how trustworthy a user, provider, or piece of content is assessed to be. It is automatically calculated from many individual observations and controls what a system permits to whom and what it does not.

Trusted Access refers to access to particularly sensitive systems or data that is granted only to a small number of vetted individuals or organizations. In the AI world, this allows researchers or authorities, for example, to gain insight into models that are otherwise locked off from the public.

Turbopack is a tool that assembles the program code of a website for the browser. It is written in the language Rust and is meant primarily to shorten developers' waiting time when trying out changes.

The Turing AI chip is a generation of graphics chips from the company NVIDIA, introduced in 2018, that for the first time brought dedicated compute units for artificial intelligence to a consumer chip. These so-called tensor cores made Turing a forerunner of today's AI accelerators in data centers.

Turn-Based describes a form of exchange in which only one side acts at a time: first the human speaks, then the system responds. Almost all AI-powered chat programs work on this principle, which reaches its limits as soon as conversations are supposed to feel more natural.

Turn-taking latency is the time that passes between the end of a human utterance and the beginning of a voice assistant's response. It determines whether a conversation with a machine feels natural or halting.

A two-tower retrieval model is a search method in which two separate program components independently translate the search query and the searched content into lists of numbers. Because the content can already be converted in advance, the system finds the best matches from millions of entries within milliseconds.

Typeahead is the term for the feature that displays matching suggestions in an input field already while typing. It aims to catch typos, save time, and guide users toward results that actually exist.

A type check is the automatic verification that a program uses the right kind of data everywhere – for instance, that a number is used where a number is expected. It catches many errors before the program even runs.

TypeScript is a programming language that builds on JavaScript and additionally specifies which kind of data is allowed at which point. This means many errors are caught while the program is being written rather than only when it is run.

"Ten blue links" is the classic look of a search engine results page: a list of links to external websites that the user clicks on themselves. Today the term stands above all for the business model that is coming under direct pressure from AI answers built into search.

With two-factor authentication, a password alone is not enough to log in. You must additionally prove something you possess or something you are — for example, a numeric code from an app on your own phone.

A utility function is a computational rule that assigns a number to every possible outcome, thereby indicating how good that outcome is. Machines and programs use it to decide between courses of action, because it allows them to directly compare different outcomes.

With usage-based billing, a customer doesn't pay a fixed monthly amount but exactly as much as they actually consumed. In the tech industry, this is the standard way to sell computing power and AI services.

A usage limit is an upper bound on how often or how much one may use an online service within a given period of time. For AI services, it caps things like the number of requests per hour or the amount of text that can be processed per month.

UA-SID stands for a session identifier that a server assigns to a visitor together with information about their browser. This allows users to be recognized across multiple page requests, even when classic cookies are missing.

UAP is the official designation for objects or phenomena in the sky that cannot be immediately explained. The term has replaced the old word "UFO" within agencies and the military because it sounds more neutral and makes no claim about the objects' origin.

A UAV is an aircraft with no humans on board, controlled by radio from the ground or by an onboard computer. In everyday language, such devices are usually simply called drones.

A UI agent is an AI program that operates a computer the way a human would: it sees the screen, clicks, types, and scrolls on its own. This allows it to carry out tasks in ordinary programs for which no special software interface exists.

The Uncanny Valley describes a sense of unease triggered by artificial figures that look almost like real humans, but just not quite. The term originates from robotics and is now important for animated films, games, and AI-generated videos.

An Uncapped SAFE is a contract that allows an investor to give money to a young company in exchange for shares later on – without setting an upper limit on the company's valuation beforehand. For founders this is the most favorable option, for the investor the riskiest.

Unclean Hands is a principle from Anglo-American law: someone who has behaved unfairly themselves gets no support from the court against their opponent. In AI disputes, the objection comes up when, for example, a plaintiff obtained their own data in a questionable manner.

Underwriting is the process of assessing whether a bank or insurance company will take on a particular risk and at what price. Today, computer programs often handle the preliminary work, calculating a risk assessment from data.

Unified Memory refers to a design in which the main processor and the graphics unit use the same working memory instead of each having their own. This saves copying operations and allows large AI models to run even on laptops.

Unified Memory Architecture refers to a computer design in which the main processor and the graphics unit use the same working memory instead of each having its own. This saves the back-and-forth copying of data and is especially a speed advantage for AI applications.

UNII-4 is a radio frequency range around 5.9 gigahertz that the USA partially released for WLAN use in 2020. It was originally reserved for communication between cars and now provides today's routers with additional frequencies.

Unit economics describes whether a company makes or loses money on a single customer or a single sale. Only once this calculation works out per unit can growth make a company profitable.

A unit test is a small check program that automatically verifies a single building block of a program for correct behavior. It runs in a fraction of a second and immediately reports if a change has broken something.

The Universal Commerce Protocol (UCP) is a shared set of technical rules that lets software assistants shop in online stores without a human clicking. It defines uniformly how a program queries products, fills a shopping cart, and triggers an order.

A universal jailbreak is an input trick that overrides the built-in rules of an AI text system – not just once, but reliably across many topics and often even across multiple providers. Such tricks are considered a serious security problem because a single text template is enough to bypass safeguards on a large scale.

A Universal AI Code of Conduct is a voluntary set of rules that companies are meant to follow when building and operating AI systems. It is not a law but a self-commitment intended to apply worldwide, often initiated by international organizations.

Untrusted Content refers to text, images, or files that originate from an external source and are therefore not considered reliable. In AI security, this is a central term because language models can read such content and follow instructions hidden within it.

Up-skilling means that employees learn new skills for their existing job because the nature of the work is changing. In the tech industry, this usually involves working with AI tools and data.

Upper Funnel refers to the earliest stage in a person's path to purchase: the point at which they first become aware of a brand or product at all. Advertising at this stage isn't meant to generate immediate orders, but to build awareness.

Upscaling refers to enlarging an image or video to a higher resolution, meaning more pixels. Modern methods use AI models that plausibly fill in missing details rather than merely multiplying existing pixels.

Uptime is the proportion of time during which a service is actually reachable and functioning. It is usually stated as a percentage, such as 99.9 percent per month or per year.

Uptraining means continuing to train an already fully trained AI model instead of starting from scratch. In most cases, it is deliberately restructured or run on new data to save computing time and money.

A URL batch is a list of web addresses that a program processes together in one pass, instead of triggering each address individually. Such batches are the usual unit of work when search engines, data collectors, or AI services fetch large numbers of web pages.

A URL pattern is a template for web addresses: instead of a single fixed address, it describes a whole group of similarly structured addresses. Programs use such patterns to recognize which page is meant and which information is contained in the address.

A usage limit is an upper bound on how much a user is allowed to use a digital service within a given period of time. For AI services, it typically restricts the number of requests or the amount of text processed per hour, day, or month.

A user journey map is a visualization that shows all the steps a person takes when using a product or service, in sequence. It also records what the person thinks, feels, and where they get stuck along the way.

A user story is a short sentence that describes what someone wants to achieve with a piece of software and why. Software teams use such sentences to break their work down into small, verifiable steps.

UTM parameters are small additions at the end of a web address that reveal how a visitor arrived at a page. Companies use them to measure which advertisement, which newsletter, or which social media post actually brings in visitors.

UX stands for User Experience and describes how using a product feels for a person: easy or effortful, clear or confusing. It refers to the entire experience, not just the appearance of an app or website.

A UX designer shapes how it feels to use an app, a website, or a device: which steps are necessary, what is placed where, and whether you can reach your goal without frustration. Unlike pure visual design, it is mainly about the flow and comprehensibility.

A vulnerability database is a public directory of known security flaws in software and devices. Each flaw is assigned a unique number, a description, and a rating of how dangerous it is.

A voice agent is a computer program that conducts a phone conversation independently: it listens, understands what is said, responds with an artificial voice, and can complete tasks along the way, such as booking an appointment. Unlike an old telephone announcement system with fixed selection menus, it responds freely to whatever the caller says.

A voice assistant is a program that understands and responds to spoken commands – such as Alexa, Siri, or Google Assistant. It converts speech into text, interprets the intent behind it, and carries out a matching action.

A voice command is a spoken sentence used to control a device instead of typing or clicking. The device converts the recording into text, interprets the intent, and then carries out a matching action.

Voting shares are company shares that allow their owner to vote on the company's decisions. Many tech corporations have two classes of shares: one with heavy voting weight for the founders and one with little or no voting weight for ordinary investors.

A value anchor is a firmly formulated principle that an AI system's behavior is meant to align with — such as "tell the truth" or "don't help anyone commit violence." Such anchors are recorded in policy documents and used as a benchmark during training and operation.

A value trap is a stock that looks cheap relative to the company's earnings or assets, but whose price stays low for good reason. The apparent bargain price is not a buy signal but a sign of a business shrinking permanently.

Vector search finds texts, images, or sounds that are similar in content, instead of only searching for exactly matching words. To do this, an AI model converts each piece of content into a long series of numbers, and the search compares which series of numbers lie close to each other.

A vector database is a storage system that stores texts, images, or sounds as long sequences of numbers and searches within them by similarity rather than by exact words. It is the foundation that allows AI assistants to find matching passages in large document collections.

A vector graphic is an image that isn't made up of individual pixels, but of computational instructions for lines, curves, and shapes. This keeps it razor-sharp at any size – from stamp-sized to building-sized.

Velocity is a metric from software development: it indicates how much work a team actually completes in a fixed time period. It is used to realistically estimate future schedules – not to evaluate teams.

Vendor lock-in refers to the situation in which a customer is dependent on a single provider because switching would be too expensive or too cumbersome. In the AI industry, this dependency often arises from data centers, data formats, and software that only work together with one manufacturer.

Vera Rubin is the name of an upcoming generation of computing hardware from the company Nvidia, being built from 2026 onward to train and run large AI systems. It follows the Blackwell generation and combines a new main processor called Vera with new graphics chips called Rubin.

Vera Rubin is the name of a computing generation from the US company Nvidia, set to be deployed in large AI data centers starting in 2026. It combines new graphics chips with a dedicated main processor and is designed as a complete server system, not a single chip.

A Verified Badge is a small symbol next to a profile name on social networks meant to indicate that the account is genuine. Originally a proof of authenticity for well-known people, on several platforms today it is a paid subscription feature.

A verifier is a program or a second AI model that checks an AI's answer and judges whether it is correct. It only judges solutions rather than inventing them itself — making AI systems more reliable as a result.

A verifier model is a second computer program that evaluates the answers of an AI system instead of inventing answers itself. It filters out incorrect solutions, thereby making AI systems significantly more reliable, especially in math and coding tasks.

Verification means checking a result after the fact rather than simply trusting it. In AI development it is central, because programs can produce answers that sound convincing yet are still wrong.

Verilog is a language that engineers use to describe the internal structure of computer chips. This text later gives rise to the millions of tiny switching elements that go into a processor or graphics chip.

Vertex AI is Google's platform on which companies build and operate artificial intelligence programs. It provides ready-made models, computing power, and management tools all in one place, so companies don't need their own data centers.

Vertical AI refers to AI software built for a single industry – such as medical practices, law firms, or banks. It knows the specialized language, workflows, and regulations of that field, but is barely suited for other tasks.

Vertical integration means that a company takes on several consecutive stages of its value chain itself instead of buying them in. In the AI industry, this means, for example: proprietary chips, proprietary data centers, proprietary models, and proprietary apps all from a single source.

VHS was the format that spread worldwide from the late 1970s onward for storing films and TV shows on magnetic tape inside a plastic cassette. It made recording and playing back moving images accessible to the masses for the first time and is now regarded as a textbook example of how a standard can prevail over technically superior competition.

Vibe coding refers to programming in which one describes in plain language what the program should do to an AI system, and barely reads the generated code oneself. The term originates from AI researcher Andrej Karpathy and became a buzzword in 2025 for a new, very fast, but also risky way of developing software.

A vibecession is a situation in which consumer sentiment is poor even though hard economic figures show no downturn at all. The term is a portmanteau of the English words "vibe" (mood) and "recession".

Vim is a program for writing and editing text that is operated entirely via the keyboard. It is considered extremely fast for experienced users, but hard to learn because keys are used differently depending on the mode.

Virtual Biology refers to the attempt to recreate living systems such as cells, proteins, or entire organs on a computer. Instead of running every experiment in the lab, researchers first test it in a model and only take the best candidates to the test tube.

A virtual machine is a computer that does not exist as a physical device, but runs as a program on a real computer. It behaves like a standalone machine with its own operating system, but shares the hardware with other virtual machines.

A Virtual Private Cloud is a separated area reserved exclusively for one customer within a large rented data center. The customer shares the underlying technology with others, but has their own network, shielded from the outside, with its own rules.

Virtual Reality is a computer-generated environment experienced spatially through a headset with two screens. When you move your head, the image changes accordingly – this is what creates the impression of actually being in a different place.

VirtualBox is a free program from Oracle that lets you run a second computer as software on your own machine. This way, you can start Linux in a window under Windows, for example, without having to reformat your hard drive.

A vision-language-action model is an AI system that takes in camera images and instructions in plain language and directly generates movement commands for a robot from them. It thereby combines seeing, language understanding, and acting in a single model instead of in separate programs.

A vision-language model is an AI system that processes images and text together. It can answer a question about a photo because it computes both types of input within the same internal representation.

A vision model is a computer program that extracts information from images or videos – for example, which objects are visible in them. It learns this ability from millions of example images instead of being programmed according to fixed rules.

Visual Effects, or VFX for short, are image elements in films, series, or advertising that did not exist in front of the camera but are created afterward on a computer and inserted into the shot. Today, AI programs are increasingly taking over work steps that used to cost weeks of manual labor.

Visual Intelligence is Apple's feature that lets the iPhone recognize what's currently visible in the camera or on the screen and provide information about it. It combines on-device image recognition with search services such as Google and ChatGPT.

Visual Search Enhancement refers to the improvement of search systems in which one searches using an image instead of words. The goal is for the system to correctly recognize what's in the photo and deliver matching results – for example, the product one is currently looking at.

Visual Studio Code, VSCode for short, is a free program from Microsoft used to write and edit program text. Today it is the most widely used tool of its kind and the place where many developers use AI assistants for coding.

Visual Studio Code is a free program from Microsoft used to write and edit program text. It is the most widely used editor for software worldwide and can be adapted to almost any programming language through extensions.

Vite is a tool that developers use to build modern websites and web applications. It ensures that changes to the code become visible in the browser almost instantly, and at the end packages the finished project into a compact form for the server.

A VLA model is an AI system that takes camera images and language instructions as input and directly generates motion commands for a robot from them. It thereby combines seeing, language understanding, and acting within a single trained system.

Voice cloning refers to techniques that generate an artificial copy of a real voice from recordings, capable of speaking any new sentences. Just a few seconds of original material are now enough for an audibly similar result.

Voice Commerce refers to shopping via spoken command, for example through a smart speaker or a mobile phone. Instead of typing and clicking, you say out loud what you want to buy.

Voice Mode is the voice feature of modern AI assistants: you speak into the microphone and receive a spoken answer without having to type anything. Newer variants process sound directly and can therefore respond almost without delay and be interrupted mid-sentence.

A Voice OS is an operating system in which spoken language is the primary form of interaction rather than tapping and swiping. The user says what should happen, and the system itself finds the appropriate app or function.

A voice agent is a computer program that independently conducts a spoken conversation: it listens, understands what is said, responds with an artificial voice, and can complete tasks in the process. Typical use cases are phone hotlines, order taking, and appointment scheduling.

A voice bot is a computer program that understands spoken language and speaks itself in order to hold a conversation with people. Companies deploy such programs primarily on phone hotlines, where they handle simple requests without human staff.

The volume weighted average price, or VWAP for short, is the average price of a stock over a period of time, in which large trading volumes count more heavily than small ones. It shows the price at which money actually flowed on a given trading day.

A von Neumann algebra is a particular collection of computational operations that can be applied to infinitely many numbers at once. It is the mathematical scaffolding physicists use to describe quantum mechanics, and today it also appears in the theory of quantum computing.

The von Neumann architecture is the basic blueprint of almost all computers: the processing unit and memory are separate, and both data and the program itself reside in the same memory. Because everything has to run over a shared connection, a bottleneck arises that becomes especially noticeable in AI applications.

A VPC is a separated, private area within a cloud provider's data centers that is used by only a single customer. It functions like a company's own network, but runs on someone else's hardware.

A VPN routes a device's internet traffic in encrypted form through an intermediary server, so that no one along the way can read it. Companies use it for secure access to internal systems, while private users rely on it mainly for privacy and for content that is blocked in their own country.

A VS Code fork is a standalone copy of the widely used programming editor Visual Studio Code, further developed and modified by another company. Almost all well-known AI coding tools such as Cursor or Windsurf originated this way.

A Vulnerability Patch Wave is the concentrated surge of security updates that breaks out as soon as a severe security flaw becomes known in widely used software. Within hours to days, vendors must deliver fixes and countless operators must install them — often in a race against attackers.

A weight is a single number in an AI model that determines how strongly a signal is passed on to a next point. All weights together form a model's entire knowledge; they are readjusted millions of times during learning.

Weights are the many numbers inside an AI model that are adjusted during learning and that afterward contain its entire knowledge. Whoever owns the weights of an AI owns the AI itself.

Weights are the many numbers inside an AI model that get adjusted during learning and determine how it reacts to an input. Whoever owns the weight files owns the finished model.

A war simulation is the playing-out of a military conflict on a model rather than in reality — in the past on maps with game pieces, today usually on a computer. For some years now, researchers have also been testing AI systems in such scenarios and observing how quickly they resort to violence.

A wafer is a thin, mirror-smooth disc made of highly pure silicon on which computer chips are manufactured. Dozens to hundreds of chips are produced simultaneously on a single disc, which are eventually cut apart.

The Wafer Scale Engine is a computer chip made by the company Cerebras that is as large as an entire dinner plate. Instead of sawing a round silicon disc into hundreds of small chips, it remains a single giant piece – which makes it especially fast at training AI models.

A wafer-scale chip is a single computer chip that is as large as the round silicon disc from which hundreds of small chips are normally cut. Such chips are built primarily to perform very large AI computations without slow connections between individual components.

A wafer-scale chip is a single computer chip that is as large as the round silicon disc from which hundreds of small chips are otherwise cut. It is mainly built for training AI models because it combines a huge number of compute units without slow connections between components.

A wake word is a predetermined word that brings a voice-controlled device out of standby, for example "Alexa" or "Hey Siri." Only after that does the device forward what was said for actual processing.

A walled garden is a closed digital ecosystem in which a single provider determines which devices, programs, and content are allowed to work together. In exchange, users get convenience and security, but lose freedom of choice and find it difficult to switch to another provider.

The waterfall model is a method for developing software in fixed, sequential steps: first plan, then build, then test, then deliver. Each step is completed before the next begins — this creates clarity, but makes later changes expensive.

A watermark is a hidden identifying feature that a computer program embeds into texts, images, or sounds so that one can later prove: a machine produced this. It is invisible to humans, but detectable with the right verification tool.

A watermark for AI content is a hidden identifying feature that a computer program embeds when generating images, texts, audio recordings, or videos. Later, matching detection software can use it to recognize that the content was machine-generated.

A watermark for AI texts is a hidden pattern that a computer program deliberately embeds while writing a text. Anyone who holds the matching key can later prove with high probability that the text originated from a machine.

Weak-to-Strong Supervision describes the attempt to have a highly capable AI system guided by a weaker teacher. The idea underlies the question of how humans can still control future AI systems once they surpass humans in many tasks.

Wearables are small computers worn on the body – such as smartwatches, fitness bands, smart rings, or smart glasses. They continuously measure body data and movements and send this to a phone or to data centers on the internet.

The Web Model Context Protocol is a planned web technology that lets a website tell an AI assistant directly which functions it is allowed to perform on it. Instead of hunting for buttons on the screen, the assistant calls clearly named actions provided by the site.

Web Platform Tests are a shared collection of test tasks that browser makers use to check whether their programs adhere to the agreed-upon rules of the internet. They ensure that a website looks and functions as similarly as possible in Chrome, Firefox, and Safari.

Web scraping means having a program automatically read out content from websites and convert it into an organized collection of data. It is the most important method for gathering large amounts of publicly visible internet content for price comparisons, research, or training AI systems.

A web crawler is a program that automatically visits web pages, saves their content, and follows the links found on those pages. This is how the data stores behind search engines and the text collections used to train AI systems come into being.

Web fetch refers to the ability of an AI assistant to access a specified internet address itself and read the content of the page. This allows the program to obtain current information that it cannot know from its training material.

Web navigation refers to the ability of an AI program to operate a website on its own: clicking, typing, scrolling, filling out forms. Instead of merely generating text, the program thereby carries out multi-step tasks in the browser, such as conducting research or making a booking.

WebArena is a testing environment in which computer programs must independently operate real websites to complete tasks such as placing an order or performing a search. It serves as a benchmark for how reliably today's AI systems can act on the internet.

WebAuthn is a technical standard that lets people log in to websites without a password – for instance using a fingerprint, face scan, or security key. The device proves the user's identity to the website without transmitting a secret that someone could steal.

A webhook is an automatic notification that an online service sends to a web address provided by the user as soon as a specific event occurs. This way, programs learn about news instantly without having to constantly check in themselves.

WebTransport is a technique that allows a webpage in the browser to exchange data with a server on an ongoing basis – in both directions and with very low latency. It is regarded as the modern successor to older methods for live connections and is used mainly for games, video conferencing, and real-time applications.

WeChat is a Chinese smartphone app that bundles messaging, payment functions, and thousands of small add-on programs into a single interface. It is considered the best-known example of what is called a super-app and has well over a billion active users.

A world model is a computer program's internal conception of how its environment behaves and what will happen next. This allows the program to play out consequences in advance instead of first trying out every action in reality.

A wet lab is a laboratory where real substances such as cells, chemicals, or blood samples are processed. It is the counterpart to the dry lab, where only computation and simulation take place on a computer — and it is the point where AI model predictions must prove themselves against reality.

A white hat hacker searches for security vulnerabilities in computer systems with the owner's permission, so they can be reported and fixed. The difference from criminal attackers lies not in skill, but in permission and intent.

Wi-Fi 6 is the generation of wireless internet for homes and offices introduced in 2019. Its most important advance lies not in raw top speed, but in enabling many devices to communicate wirelessly at the same time without interference.

The Windows kernel is the innermost part of the Windows operating system. It allocates processing time and memory to programs and is the only place allowed to communicate directly with the hardware.

Workflow orchestration means that a program automatically starts, monitors, and, in case of errors, retries many individual work steps in the correct order. In AI systems, it ensures that data preparation, model calls, and post-processing reliably work together.

The workflow layer is the software between the user and an AI model. It determines the order in which work steps take place, which data is added, and who reviews the result.

Working memory is the notepad of an AI program that carries out several work steps one after another: it holds the task, intermediate results, and the current status. Without this notepad, the program would forget what it's doing at every single step.

Workslop refers to work output that was quickly generated with the help of AI, looks professional, but is empty of substance. The effort doesn't disappear as a result — it just shifts to the colleagues who have to fix the text afterward.

Workspace Agents are AI programs that independently carry out tasks within a company's office software — such as answering emails, scheduling appointments, or analyzing spreadsheets. They don't wait for every single command, but instead break a task down into steps and carry it out.

A world model is a computer program that has learned how an environment behaves and can predict what happens next. It serves robots, game characters, and video generators as a kind of internal simulation of the world.

A wrapper is a program that uses a third-party provider's AI and wraps it in its own interface and features. In the tech industry, the term is often used disparagingly, because such products are easy to copy.

Wright's Law states: every time the total quantity of a product ever produced doubles, unit costs fall by a fixed percentage. The rule originated in aircraft manufacturing in 1936 and is used today to forecast price trends for batteries, solar cells, and computer chips.

WSL is a feature of Windows that lets you run Linux programs directly under Windows, without restarting the computer or using a second device. For developers, WSL is today the standard way to work in a familiar Linux environment on a Windows machine.

XGBoost is a program that makes predictions from tabular data by combining many simple decision rules. For years it has been considered the standard tool for tabular data, for instance in credit risk assessment or price forecasting.

Xtacking is a manufacturing process from the Chinese memory maker YMTC in which the control electronics of a memory chip are fabricated separately and then placed on top of the actual memory cells. This makes chips faster and denser, without both parts having to make the same manufacturing compromise.

YAML is a text format in which people write down settings for programs. It does without brackets and instead uses indentation, so that the structure remains directly visible.

A zero-coupon bond is a security on which the buyer receives no interest payments during its term. Instead, they buy the paper more cheaply than it is repaid at the end — the profit lies entirely in this difference.

A zero (or root) is an input value at which a function evaluates to zero. In a graph, this is exactly the point where the curve meets the horizontal axis.

Zero Data Retention is a provider's commitment not to store the inputs and outputs of an AI service after processing. The data is only used to generate the response and then discarded.

Zero Trust is a security principle for computer networks: no device and no user is automatically treated as trustworthy just because they are already inside the corporate network. Every access must be individually verified and approved.

A zero-day is a security flaw in software that the manufacturer does not yet know about and for which, therefore, no fix exists. Attackers who discover such a flaw can exploit it before anyone can defend against it.

Zero-Day discovery refers to the finding of a security flaw in software that is still unknown to the manufacturer and for which, therefore, no fix exists. The name comes from the fact that the manufacturer has zero days left to react before the flaw can be exploited.

A zero-day exploit is an attack that takes advantage of a previously unknown programming flaw. Because the vendor doesn't yet know about the flaw, there is no update available to protect against it.

A zero-day vulnerability is a programming flaw in software that is known to attackers but not yet to the manufacturer. As long as no update exists, all users of this software are unprotected.

A zero-knowledge proof is a mathematical procedure that allows someone to prove that a statement is true without revealing the underlying data. In other words, you show that you know or possess something – but not what exactly.

Zero-shot means that an AI system solves a task for which it has not seen a single example — an instruction in plain language is enough. The ability stems from the model's general pretraining, not from specific practice on that task.

Zero-shot learning means that a computer program solves a task for which it has not seen a single training example. To do this, it transfers knowledge from its general training to the new task.

Zero-Talent Production refers to the making of films, music, advertising, or texts without any humans performing in front of the camera or at the microphone. Voices, faces, and bodies are generated entirely by computer.

Zero-width characters are characters that take up no space when displayed and therefore remain invisible. They are nonetheless present as real data within the text and can be used for hidden markers or to deceive filters and AI systems.

Zombie code refers to parts of a program that nobody needs anymore, but which still remain in the project and sometimes even continue to run. It costs time, confuses developers, and in rare cases can unexpectedly become active again.

A zombie company is a firm that permanently earns so little that it cannot even pay the interest on its debts out of its own resources. It nevertheless does not go bankrupt because banks or the state keep it alive with ever more fresh money.

τ-Bench is a test that checks how well an AI assistant handles real customer conversations while completing tasks in a simulated company system. It doesn't measure knowledge, but reliable action across multiple conversation steps.