A mathematician at Anthropic says he and the company's AI, Claude, have found a proof of the Riemann hypothesis, and the Lean community has certified the formal proof as correct — a machine-checked result that still awaits a paper mathematicians can read.
Levent Alpoge, who works at the US company, announced the claim on his website on November 20 local time. He published the proof as a formal proof written in Lean, a proof assistant that checks each step of an argument by computer. On November 21, the community of Lean's developers and users issued a joint statement certifying the proof as correct.
The Riemann hypothesis was proposed by the German mathematician Bernhard Riemann in 1859. It holds that the points where the Riemann zeta function equals zero — its nontrivial zeros — all lie on a single straight line. That would tie the apparently irregular sequence of prime numbers to a deep underlying order.
It is regarded as the most important unsolved problem in number theory. The Clay Mathematics Institute has put a $1 million prize on it as one of its Millennium Prize Problems. Of those problems, only the Poincaré conjecture has been solved.
What Alpoge says Claude did
In his statement, Alpoge said the method also settles the generalized Riemann hypothesis, a far broader version of the problem. Of the work with Claude, he wrote: "Near the goal, Claude far surpassed me."
The statement includes a timeline of the work that Alpoge says Claude compiled. According to it, key ideas for the proof came one after another from October 6, and the proof of the generalized Riemann hypothesis was reached on November 16.
No paper in a form mathematicians can read and check has been published. Alpoge said the proof as it stands is extremely long. A detailed document can be generated from the formal proof, but a trial run produced more than 250,000 pages, which he described as unreadable. He has released only the Lean code and plans to work through the content and write it up as a formal paper.
Independent checks
Lean lets a computer mechanically check whether each step of a proof is logically valid. According to Scientific American, outside researchers including the MIT mathematician Andrew Sutherland compiled the published code themselves and confirmed it ran without errors. Sutherland said there was almost no chance of anything suspicious.
The communities behind Lean and Mathlib, Lean's mathematics library, published a joint statement on November 21 certifying Alpoge's proof as correct. Lean FRO, the organization that develops Lean, is among those backing it.
According to the statement, the proof runs to 17 million lines of Lean. Several dozen people reproduced the compilation. Two independent tools confirmed that the theorem proved matches the description of the Riemann hypothesis recorded in Mathlib. Multiple independent checkers also inspected the proof.
The checks applied four criteria for evaluating a Lean proof that Kevin Buzzard, a mathematician at Imperial College London, set out in a blog post on November 12, before the announcement, and the proof met all of them. The statement adds that the work of mathematicians understanding and explaining the proof's new ideas lies ahead.
Earlier AI results fell short of the full conjecture
Anthropic said in August that an unreleased Claude had produced a related result. That work raised the lower bound on the share of zeros lying on the line from 41.6% to 67.2%, and did not prove the Riemann hypothesis itself. In October, OpenAI published a batch of mathematics results from its in-house AI. They included a proof of the quasi-Riemann hypothesis, a weakened form of the problem. The new claim differs in asserting a proof of the Riemann hypothesis itself.
Scientific American noted that the announcement came the day after Anthropic's stock listing. Alpoge acknowledged in his statement that the timing could invite a cynical reading, and said that reading is wrong.
