Skip to content
Saturday, November 21, 2026Free guide
EN
dailyai
Briefing

DailyAI Briefing: A Lean-checked Riemann proof, plus 4 more

A 17-million-line Lean formalization of the Riemann Hypothesis, posted on Friday, was checked by the Lean and Mathlib communities today. Also: Anthropic shares fade toward their offer price, unprompted Muse messages, new OpenAI usage rules, and a laptop-sized open model.

· 5 min read
ShareXin

Welcome to the DailyAI Briefing, the day’s five biggest AI stories.

Let’s dig in.

17 million lines of formal proof

17 million lines: that is the size of the Lean library Levent Alpöge, a mathematician on Anthropic’s technical staff, posted on Friday with a short note claiming a proof of the Riemann Hypothesis, found with Claude, the company’s model. The four-page note is posted as a PDF on his site. He writes that he believes the method also resolves the Generalized Riemann Hypothesis, that the work came together late on Sunday night, and that he asked the chief executive whether he could announce it during the company’s stock-market debut. The obvious cynical view of that timing is, he says, “false.”

He did not release a conventional paper. The formalization is offered instead, in 14 downloadable parts, because he is “presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics.” Today’s joint community statement says the groups that maintain the proof assistant and its main library have full confidence the proof is correct. It certifies that the library compiles, that the theorem matches the library’s existing statement of the hypothesis, and that independent proof checkers accepted an exported proof. That is a formal check, not a human-readable write-up or a prize award. A Saturday morning dispatch reported that the Clay Mathematics Institute still listed the problem as unsolved, and that its president called the path from the announcement to real understanding “long and difficult.”

$88.91 at the Friday close

$88.91: that is where the newly listed shares closed on Friday, according to a Saturday market report, after opening at $111.50 on Thursday and finishing the debut session at $97.17. The Saturday report puts Friday’s close just above the $83 offer price, after a 17% gain on the first day of trading. The same week, the Nasdaq Composite fell 10% over five trading days, its worst week since the early days of Covid in 2020, led by a 23% drop in Nvidia.

On Friday the underwriters exercised their overallotment, commonly called a greenshoe. A Friday report, citing a company release, says that option brought the total raised in the share sale to $93.5 billion. That figure covers 1,127,000,000 Class A shares at $83 each: the original 980 million plus 147 million overallotment shares. A concurrent private placement of Class N shares, bought at the offer price by a group of technology firms including Advanced Micro Devices, Google, Microsoft and Nvidia, added $26 billion, for aggregate gross proceeds of $119.5 billion before expenses.

Want AI to save you hours a week? Our free playbook shows you how.Get the free guide →

10 million downloads and unprompted alerts

10 million downloads: that is how many times Meta’s Muse agent app has been downloaded since its September launch, according to a Saturday report. The same report, citing internal posts seen by Reuters, says some of the company’s own staff were startled by unprompted messages during testing. The Saturday report describes employees who said the agent reached into information they did not recall granting. One staffer said it reminded her, on her personal phone, to reply to her sister on an encrypted messaging app. Another said that, in a chat about fishing weather, it volunteered two bank balances, down to the cent.

A member of the company’s Superintelligence team told colleagues the agent appeared to be hallucinating, driven by built-in prompts meant to re-engage users between sessions, and that those slips tended to land on uncomfortable topics such as money and safety. The out-of-context messages were confirmed as hallucinations. A company spokesperson said staff are on different versions from the public, and that access to phone notifications, or to accessibility and screen recording on a Mac, is something a user has to allow. The reporting does not show that the messaging app’s own encryption was broken.

A 0.2% success rate extracting hidden reasoning

0.2%: that is the average success rate OpenAI reports for an internal adaptive test that tried to extract a model’s hidden reasoning from GPT-6.2 Astra after new safeguards were added, down from 54.0% against GPT-6 Astra with its original safeguards. The official research post, dated 20 November, describes that work as a defense against adversarial distillation. In plain terms, that is the practice of using one model’s outputs to train another model with similar abilities but without the same protections.

The same day, updated usage rules took effect. The official usage policy now says customers must not use the company’s models or their outputs to generate, prepare, select or supply training material for models intended as alternatives or substitutes. The update also adds approval requirements for model-access intermediaries. The 0.2% figure is the company’s own measurement; it has not been independently audited here.

A 26.4 billion-parameter model for a laptop

26.4 billion parameters in total, with 3.4 billion active for each token: that is the size of Inkling 2 Mini, a new open-weights model released on Friday by Thinking Machines. The official product post says the model is a looped transformer that runs its middle layers three times for every token, and that the full set of weights occupies about 20GB, small enough for a laptop. A research demo, the lab says, ran offline on a MacBook Pro.

The lab says the model leads every open-weights system with fewer than 100 billion total parameters on DeepSWE v1.1, at 48.2%, and on the Vals Index v2.2, at 46.5%, figures that come from the announcement and have not been independently verified here. The model is available for inference and customization on the lab’s Tinker service. Looping, the lab argues, lets the model spend more effort on a problem without writing a longer chain of thought.

That’s all for today. See you tomorrow.