Wink Pings

Has the Riemann Hypothesis Been Proven? An Anthropic Researcher Says Claude Found a Proof

Anthropic researcher Levent Alpöge published a note claiming that he and Claude discovered a proof of the Riemann Hypothesis, along with a Lean proof package. There has been no independent review yet; the key question is whether it can pass machine checking.

On November 20, Levent Alpöge, a number theorist and current member of Anthropic’s technical team, published a short note on his personal website (alpo.ge/remark.pdf), stating that he and Claude had discovered a proof of the Riemann Hypothesis. Unlike a verbal hint, the first page’s footnote of this note directly provides split archive Lean proof code: lean.tar.gz.part01 through part14, for outside download and verification.

The Riemann Hypothesis concerns the distribution of prime numbers and is one of the central unsolved problems in mathematics. Alpöge is no outsider. He studied under Manjul Bhargava, was formerly a junior fellow at the Harvard Society of Fellows, and has a track record in arithmetic statistics and number theory. In August this year, Anthropic publicly described Claude’s progress on a lower bound for the proportion of zeros of the Riemann zeta function, moving the known proportion from 41.6% to 67.2%; in July and August, he himself also released a counterexample to the Jacobian conjecture, a new construction of Hadamard matrices, and work related to the Hopf problem. From these advances to directly claiming a proof of the Riemann Hypothesis is a huge leap.

Some online commenters relayed details of the proof process:

60 subagents, 31 million tokens, 650 failures, and finally the proof ran through in Lean.

These numbers are more noteworthy than the statement itself. Lean is a formal verification tool in which proof steps are written as machine-checkable code. Being able to pass completely in Lean means that every step of reasoning can withstand mechanical checking, and readers do not need to fill in gaps by intuition. One commenter said that it does not matter if humans cannot understand a 250,000-page proof, as long as machines can verify it; in the future, review of major results may go through the compiler before going through people. The premise for this judgment to hold is that the proof code itself contains no hidden assumptions or holes.

At present, this set of Lean files has not undergone independent peer review. The mathematics community and the AI community are discussing verifiability. Some people in the comments hope it is not a false alarm, while others admit that ordinary people have no way to confirm something at this level. Whether the formal code can pass completely in Lean and whether there are unnoticed holes will always be more critical than the claim. Only after external researchers unpack these 14 split files will the conclusion truly become clear.

发布时间: 2026-11-21 12:35