Large Language Model

A Large Language Model is a computer program that writes texts. The name, abbreviated LLM, describes exactly that. Beforehand, it has read gigantic amounts of text: books, newspaper articles, websites, program code. In doing so, it has memorized which words in which context typically occur together. When you now ask it something, it predicts which word would best fit as the answer. Then the next one. And so on, until a complete text has been produced. ChatGPT, Gemini, and Claude are well-known programs built on such models.

Why this matters

Before 2020, computers could only process language within narrow limits. A program for translation could do nothing else. A program for spell-checking couldn’t either. Large language models break with this pattern. A single model can translate, summarize, write program code, and solve exam questions. Nobody programmed it individually for these tasks.

This makes the technology extremely interesting economically. Companies like Microsoft, Google, and Nvidia have tied their market value strongly to it. At the same time, language is the interface to almost every office job. That is why there is such heated debate about which activities will change in the coming years.

How it works

Inside is an artificial neural network. This is a computational method with billions of adjustable numbers, the parameters. These numbers are the model’s memory. During training, the model is shown a piece of text in which the next word is missing. It guesses. If it’s wrong, the parameters are adjusted minimally. This is repeated many billions of times.

You can imagine it like extreme practice with fill-in-the-blank texts. Anyone who fills in millions of such gaps learns not only grammar. They also learn along the way that Paris is in France and how a recipe is structured. Because this knowledge is needed to fill in the correct blank.

Important: the model calculates with probabilities, it does not check facts. That is why it sometimes invents things that sound plausible yet are false. Experts call this hallucination. After the basic training comes fine-tuning. Humans rate the answers, and the model learns to prefer helpful and polite responses.

Where you encounter the term

In everyday life, the technology is found in chatbots, in the autocomplete feature of emails, and in search engines that deliver whole answers instead of lists of links. Translation services and coding assistants also use it. In schools, it has become a constant topic, because homework can now be produced with it in seconds.

In the news, you usually encounter two numbers. The parameter count describes the size of the model. The context window indicates how much text it can take into account at once. In addition, model names with version numbers keep appearing, such as GPT-5 or Llama 3. You also read about open models, which anyone is allowed to download, and closed ones, which only run on the provider’s servers. Anyone who knows this distinction understands most reports on the AI industry much faster.