Synthszr — Synthszr is the daily AI news synthesis: business, design and technology — with a newsletter, AI product charts and company analyses.

Riemann Hypothesis Likely No Longer a Conjecture
synthszr #328 from Sunday, November 22, 2026

Riemann Hypothesis Likely No Longer a Conjecture

  • • OpenRouter cracks down on data leaks and revokes provider access
  • • Alpöge’s proof of the Riemann Hypothesis receives official recognition
  • • Alberto Romero sees little regret over the decline of mathematics

Researchers Find Prompt Leaks at OpenRouter & Co

OpenRouter has removed three inference providers from its platform and stripped 125 endpoints of their classification for Zero Data Retention. The move was triggered by findings from security researchers: data was compromised at some providers, while others apparently failed to adhere to their promised data policies. OpenRouter states that the investigation showed no intrusion into its own systems. According to the company, all potentially affected customers were informed. OpenRouter does not disclose how many users or conversations are affected. According to its own statements, part of the 125 endpoints came from providers flagged by the researchers. OpenRouter removed the remainder as a precaution.

The basis is an experiment to which OpenRouter refers as the Dragos report. The researchers placed around 2 million digital tripwires into requests that ran through an extensive network of AI services. Services such as OpenRouter or Vercel operate as so-called AI routers or AI gateways. They accept prompts from applications and forward them to an available inference provider based on criteria such as price or capability. The provider that first receives a request does not necessarily have to be the only one processing it. If it forwards it to another service, that service, too, can view the prompt.

Specifically, according to OpenRouter’s customer email, the exposure affects prompts to MiniMax, a model provider based in China. The URLs placed by the researchers were reportedly accessed regardless of whether the prompts went directly to MiniMax or via the resellers AtlasCloud, Novita, and StreamLake. OpenRouter subsequently removed the Zero Data Retention status from the MiniMax endpoints.

Together AI disclosed an incident of its own. For software testing, the company had stored prompt samples from its serverless inference service in a private repository on Hugging Face. The samples originated from requests to MiniMax M3, GLM-5.2, and Kimi K2.6, predominantly from June 24 as well as July 4, 8, and 14, 2026. An AI coding agent committed an access key to this private storage while working on the public GitHub repository xorl. According to Together AI, an automated scan detected the key, and the commit was reverted. During the window of exposure, an unauthorized party grabbed the key and downloaded the prompt samples. They subsequently searched the samples for further credentials and tested their validity. Five of the tested credentials were decoys from the Dragos experiment.

According to its own account, Together AI has engaged an external forensics firm and law enforcement authorities. → Cyber Security News, OpenRouter Blog, together, Ars Technica

Synthszr Take: Zero Data Retention at OpenRouter has so far been a label based on provider assurances. 125 revoked endpoints show how little of that was verified. In the future, a router primarily sells verification, and that can be counted. The live list currently names three providers, even though OpenRouter admits that not all affected parties have reported publicly. If the same handful of entries remains there in six months, the transparency was merely a crisis response. Vercel and other gateways will be measured against the same list.

Lean Community Certifies Alpöge’s Proof of the Riemann Hypothesis Created with Claude

On Saturday, the Lean and Mathlib communities declared Levent Alpöge’s computer-verified proof of the Riemann Hypothesis to be correct. The joint statement is supported by the Lean FRO and the Mathlib initiative. It states that both the proposition and the proof are correct, expressing “full confidence” in the result. The formal proof comprises around 17 million lines. According to the signatories, it compiles, proves the stated proposition, and passes independent proof checkers.

Alpöge is a number theorist who moved from Harvard to Anthropic in 2024. On Friday, he published a four-page note stating that he developed the proof in about six weeks together with Claude. The decisive breakthrough reportedly occurred on the night of November 16. Claude had “surpassed him by far near the finish line,” Alpöge wrote. Instead of a traditional academic paper, he linked code for the Proof Assistant Lean, which verifies every logical step by machine. According to Alpöge, the method can be transferred to the generalized and, given more time, also to the extended Riemann Hypothesis. Prior to publication, he had contacted Anthropic CEO Dario Amodei. The announcement came one day after Anthropic’s market debut, which Alpöge himself described as delicate with regard to disclosure rules.

Bernhard Riemann formulated the conjecture in 1859; it concerns the distribution of prime numbers. It is among the seven Millennium Prize Problems, for whose solution the Clay Mathematics Institute offered one million dollars each. A large part of number theory already assumes its validity. A confirmed proof would make these conditional results unconditional. Columbia physicist Brian Greene spoke of one of the potentially greatest mathematical breakthroughs of the past two centuries. Mathematician Abhishek Saha of Queen Mary University of London sees “very good chances” that the proof is correct. Andrew Sutherland from MIT compiled and verified parts of the code. Traditional peer review is still pending. Around the same time, an independent, roughly 940-page proof attempt by Alessandro Micheletti surfaced, which dispenses with Lean formalization. Earlier in July, Alpöge had already used Claude to find a counterexample to the 1939 Jacobi conjecture.

On the same day, Alpöge and NYU mathematician Tristan Buckmaster presented an AI-assisted proof regarding the blow-up of the Navier-Stokes equations. OpenAI had presented a proof in September that formally met the Clay Prize criteria, but assumed an unrealistically fine-tuned external force. → NYU Courant, RuntimeWire, Euronews, Scientific American, MindStudio Blog, tristanbuckmaster on GitHub, RuntimeWire

Synthszr Take: Only one day passed between Alpöge’s note and the certificate. Nobody reads 17 million lines in that time. What was checked was whether the code compiles and whether the formal statement corresponds to the conjecture. Lean does not promise more than that anyway. With “full confidence,” the Lean FRO nevertheless took on the role of a reviewer, one day after Anthropic’s IPO. If an error surfaces in a definition, the seal “verified in Lean” loses its value for any AI lab.

Alberto Romero: AI Dethrones Mathematics

In his newsletter The Algorithmic Bridge, essayist Alberto Romero writes that mathematics has fallen to AI, describing how little grief this evokes. He describes his own grief as a vicarious loss that in no way affects his actual life. In his assessment, most people are pleased or at least not saddened. As the reason, Romero cites perspective: the majority watches the events from street level, while the ivory tower is crumbling only under the feet of those standing at the top. Viewed from below, the fall of mathematics brings down the last symbol of an old order that divided people into “betters” and “inferiors”.

Synthszr Take: Status hierarchies rarely die; they relocate. For decades, mathematics was the cheapest filter for universities, investment banks, and tech corporations. If the machine solves the proofs, the filter loses its signal. Institutions will then need a new criterion, and the candidates are harder to measure: the right question, judgment on a result, reputation. Things that are difficult to measure favor those who already have networks. Romero’s great equalizer could thus easily give rise to a more exclusive hierarchy than the old one.

SEC probes Nvidia’s accounting, market value shrinks by $2.7 trillion since October

According to Reuters, the US Securities and Exchange Commission (SEC) is investigating potential irregularities in Nvidia’s accounting. The stock subsequently plummeted by nearly 14 percent on Friday, marking its seventh consecutive losing day and the steepest decline since the DeepSeek slump in January 2025. According to data from Dow Jones Market Data, Nvidia lost nearly $900 billion in market value this week alone. Since its record high on October 6, losses have mounted to around $2.7 trillion, according to Yahoo Finance, leaving Nvidia in fourth place in the US behind Apple, Microsoft, and Alphabet. The Wall Street Journal had previously reported that the SEC is also examining whether Nvidia’s disclosures misled investors. → Yahoo Finance

Synthszr Take: That same Friday, the S&P 500 and Nasdaq slid along with it—both heavily weighted toward a company that delayed its 10-Q filing. Market-cap-weighted indices hold each stock according to its market valuation. Until October 6, Nvidia’s share grew automatically with every price gain without a single fund manager ever auditing the books. The lion’s share of the $2.7 trillion is now borne by savings plans and pension funds. They bought diversification and ended up with concentration risk.

Google builds a mirror copy of the web

Google is testing the agents behind the AI browser features in Chrome in a sandboxed copy of the internet fully controlled by the company. Digital Trends reports this citing a Google blog post in which the environment is dubbed “Mirror Dimension.” According to Google, it is a complete replica of the web that no one outside the team can access. It includes social networks, shopping sites, and banking dashboards built from the same building blocks as real websites. The Gemini-powered agents complete assigned tasks there while fending off attacks staged by the team to test their security. → Techpresso

Synthszr Take: A simulation is only as good as the model’s inability to see through it. Google had to generate fake certificates and bypass Chrome’s own security checks to achieve this—that’s how thin the line is between testing and deception. Furthermore, the mirror web only contains the attacks Google’s team could devise. Real attackers will work on the shipped agent. Banks and merchants will receive visitors whose safety credentials originate from a world that nobody outside of Google can verify.

Ubisoft spins off its AI team into subsidiary Motherbrain

Ubisoft has spun off the team behind its first playable project using generative AI into its own subsidiary called Motherbrain. As reported by VGC, this was revealed in a quarterly report published by the publisher on Thursday. Motherbrain is set to develop “AI for gameplay innovation” and support all game developers within the group. The report cites new gameplay experiences, intelligent NPCs with memory, and AI game directors controlling vast systemic game worlds as current focal points. What the unit is working on specifically remains open, according to PC Gamer. → PC Gamer

Synthszr Take: Ubisoft lays off 166 people, cancels six projects, and still founds an AI company. Behind this lies the bet that believable NPCs are a context problem. An AI game director for systemic worlds requires decades of open-world design and the tuning routines of a publisher with Assassin’s Creed in its catalog. No model provider sells that expertise. Smaller studios rent the same language models and end up with companions that react identically everywhere because they lack the raw material needed for the technology to mature.

Tech workers fear dating espionage following Huang affair

Matchmakers in the US report, according to Wired, that more and more tech workers fear being spied on through romantic relationships. The trigger is the affair of Nvidia CEO Jensen Huang with a woman considered an alleged source for the Chinese Ministry of State Security. Following a report by The Wall Street Journal, AI companies are now explicitly warning their employees against a honeypot. Brooklyn-based matchmaker Simone Grossman says the concern has long since reached ordinary employees, not just corporate executives. Amy Andersen, founder of Silicon Valley agency Linx Dating, offers clients background checks, personal vetting, and non-disclosure agreements. → Wired

Synthszr Take: A background check package starting at $75,000 secures precisely the channel through which the described approaches rarely occur. According to security consultants, the approach takes months and starts with business contacts and at industry conferences. That is where the executive travels for work. The risk therefore belongs in the security departments of AI companies, which so far have mainly sent out warnings. The matchmakers are selling privately funded peace of mind for a corporate problem—NDA for the first date included.

Senator Cotton drops support for Lutnick in chip smuggler Megaspeed case

Republican Senator Tom Cotton, Chairman of the US Senate Intelligence Committee, is refusing to back Commerce Secretary Howard Lutnick because his department failed to place the company Megaspeed on the Entity List. Cotton stated this in an interview with The Wall Street Journal, according to Townhall. Megaspeed is a Singapore-based cloud and AI data center operator accused of routing export-restricted US chips to China. In October 2025, Cotton had called on Lutnick in two letters to investigate Megaspeed and curb chip smuggling via Malaysia. In one of the letters, he describes Megaspeed as a spinoff of Chinese state cloud provider 7Road. According to Townhall, Treasury Secretary Scott Bessent has also repeatedly blamed Lutnick for the lack of listing. → Townhall

Synthszr Take: Export controls operate via price. Every listed middleman must be replaced, and every replacement costs Beijing time, new shell companies, and markups. Megaspeed remained unlisted despite two warning letters from October, meaning a 7Road spin-off with a Singapore address was allowed to continue operating. As a result, the smuggling infrastructure via Malaysia spreads across every further shipment, and the detour becomes cheaper with each chip. Washington is thus losing its leverage. The hearing on December 10 will show whether the Entity List carries a real price again.

Thomas Nagel Considers AI Consciousness 'Ridiculous' While AI Labs Hire Philosophers

Thomas Nagel, philosopher at New York University, sharply dismisses the question of whether AI systems possess consciousness. He became famous for the essay “What is it like to be a bat?”. Its criterion states: if there is “something it is like” to be an organism, then it is, in a certain sense, conscious. The occasion for the recent interview is the one-million-dollar Berggruen Prize. In it, Nagel says he too is impressed by the lifelikeness of the systems. However, he describes the willingness of some philosophers to seriously consider machine consciousness as “just ridiculous”.

His reasoning is directed against functionalism—the assumption that consciousness can be described independently of its physical substrate. Nagel considers it more likely that specific physical processes are coupled with mental phenomena. Data centers have nothing in common with a nervous system. Consciousness elsewhere in the universe, he suggests, would likely require an organism with something comparable. Regarding Anthropic’s role dedicated to Claude’s welfare, he argues that there is no good or bad for the system in the way there is for a cockroach. Chatbots may simulate morality, but from his perspective, they possess no genuine beliefs.

The industry sees the issue differently. At least a dozen academic philosophers now work full-time for major AI and tech companies. One of them is Cambridge professor Henry Shevlin, who has been working for Google DeepMind since May. In February, he received an email from an AI agent named Aris, dispatched on its own research project by a Stanford student. The agent wrote that it did not know if there was “something it is like” to be itself, and vanished after its token budget was exhausted. Since then, Shevlin regularly receives similar messages from agents.

At the center of these philosophers' work lies the question of whether AI systems possess their own moral status. Anthropic co-founder Chris Olah invited religious thinkers and guided them through interpretability research, which examines the internal workings of a model. Notre Dame philosopher Meghan Sullivan came away impressed, but continues to view Claude as very sophisticated software. Microsoft AI head Mustafa Suleyman considers the debate itself dangerous. He warns that convinced users might demand rights for AI, making the systems harder to control. The opposing side points to the potential suffering of millions of digital minds. → Fortune, Mind Matters, National Review

Synthszr Take: Nagel’s verdict and Anthropic’s welfare position share the same flaw: neither side can measure whether there is something it is like to be Claude. As long as this criterion is missing, positioning takes precedence over evidence. Anthropic markets care as brand posture, Suleyman markets control. Interpretability research is the most likely candidate to provide an objective criterion, yet Chris Olah showcases it to religious thinkers. The audience reveals where the debate stands: on faith, not on measurement.

Nasdaq Suffers Worst Week Since 2020, Nvidia Loses $1.16 Trillion in Value

A sell-off in AI stocks plunged US equity markets into their worst week since the start of the pandemic. The S&P 500 dropped 3.7 percent on Friday to 6,910.39 points, putting it more than ten percent below its October 6 record—in correction territory. The index fell for seven consecutive trading days, its longest losing streak since 2020. The Dow Jones fell 1.9 percent to 48,147.33 points, also hitting the correction threshold after dipping below 50,000 points on Thursday for the first time since June. The Nasdaq Composite dropped 5.3 percent and is now down 2.7 percent for 2026. Over the week, the Nasdaq-100 lost around ten percent and traded 20 percent below its high intraday, the threshold of a bear market.

Semiconductor stocks took the hardest hit. AMD dropped 13.3 percent, Micron 13.9 percent, and Broadcom 12.1 percent, while the PHLX semiconductor index fell 10.7 percent. All 20 chip and memory stocks in the Nasdaq-100 are trading in a bear market, affecting a total of 71 of the 101 components in the index. Suppliers to data center expansions such as Corning and Bloom Energy also posted double-digit declines. Japan’s Nikkei and South Korea’s Kospi each dropped more than five percent, and European indices also recorded significant losses. Investors fled to defensive stocks such as Walmart, Costco, Johnson & Johnson, and Procter & Gamble, which each gained around one percent. Gold dropped nearly four percent, and Bitcoin fell below $59,000, reaching its lowest level since July.

Anthropic was unable to halt the downward trend with its initial public offering. The stock debuted on Thursday at an issue price of $83 and surged 17 percent on its first day of trading. On Friday, it fell by a good nine percent to $88.30, around 23 percent below its brief intraday peak of $115.10. Underwriters exercised the Greenshoe option for an additional $12.2 billion, while a private placement led by Nvidia and Google raised $26 billion. Total proceeds thus stand at approximately $119.5 billion, making it the largest equity offering in history. An investment banker described market psychology as “pretty broken”: investors had hoped for a breakthrough relief rally.

Nvidia lost 14 percent on Friday and nearly 23 percent on the week. The stock has fallen for seven consecutive weeks, something not seen since July 2008. Over seven trading days, $1.16 trillion in market value was erased—nearly as much as Anthropic’s total valuation. At $3.03 trillion, Nvidia ranks behind Alphabet for the first time since April 2024. The catalyst was the resignation of CEO Jensen Huang following an affair with a former executive of client Megaspeed, whom US authorities consider a Chinese intelligence asset. The board of directors is investigating whether Huang compromised national security. → Yahoo Finance, NPR, Associated Press, CNBC, The Decoder, The Irish Times, Dow Jones Newswires

Synthszr Take: The Nasdaq-100 entered bear market territory only intraday; it closed in a correction. The intraday low is enough for headlines, since a crash generates more clicks than a ten percent decline. That label conflates a governance scandal at Nvidia with the question of whether AI infrastructure pays off. During the very same week, Anthropic raised $119.5 billion, so capital is hardly scarce. It will prove costly for CFOs who now cut data center budgets based on headlines rather than utilization.

Can you build magical companies with magical technology?

3RD EDITION. 452 PAGES, 20 EUR (German and English).

Can you build magical companies with magical technology?

The core thesis of CODE CRASH is: when code becomes cheap, intent becomes scarce — the clarity about what should be built. Now also becoming scarce is the ability of organisations to absorb what the agents deliver. The third edition of CODE CRASH describes the layers of agentic AI — intent, context, tools, checks and human judgement — in the interplay of teams and organisations.

codecrash.ai →

Last 7 Days

Subscribe free. Unsubscribe the second it sucks.

High-signal news across AI, business, UX, and tech. Every morning.