November 21, 2026ResearchAgentsBenchmark

A Riemann Hypothesis Proof Made With Claude, and the Only Copy Is Lean

The most famous open problem in mathematics may have just been claimed by a mathematician and a chatbot, and almost nobody can read the proof. On Friday a four-page note by Levent Alpöge, a mathematician who works at Anthropic, went to 2,136 points on Hacker News. The key line: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." The method extends to the Generalized Riemann Hypothesis, and "Claude and I have also formalized the arguments without assumptions." There is no paper. There is only Lean, split into 14 archive parts. Scott Aaronson counted 17.4 million lines of it. Note: https://alpo.ge/remark.pdf

Why no paper? Alpöge says a human-readable version was "routine to produce" but would be "some massive unreadable travesty of slop." A footnote says the draft "book" hit 250,000+ pages, and its Theorem 1 opened by noting that the reals are totally ordered. The timeline in the note runs from October 6 to November 16, the day it lists simply as "GRH." Terence Tao rounded up the day in order of how digested each result is. Most digested: Alpöge, Tristan Buckmaster and Matei Coiculescu, "with heavy AI assistance," have a human-readable proof that the unforced 3D incompressible Euler equations blow up in finite time, also formalized in Lean. Then Alpöge and Buckmaster on finite-time blowup for 3D Navier-Stokes, with a 79-page paper and a Lean package of about 600,000 lines. Least digested: the GRH proof, which Tao describes as made "using an internal AI model." Tao: https://terrytao.wordpress.com/2026/11/20/unforced-euler-singularity-unforced-navier-stokes-singularity-and-the-generalized-riemann-hypothesis/

Then came the human counterpunch. At 18:18 UTC Ravi Vakil posted a proposed GRH proof by Alessandro Masullo: "all human, with no AI," checked by Brian Conrad "for months," originally meant to be polished for a year at Stanford. "Now it seems best to put out the current draft." Two claimed proofs of a 167-year-old problem in one afternoon, one with a reader and one with a type checker: https://math216.wordpress.com/2026/11/20/hurried-announcement-alessandro-masullo-grh/

Be clear about what this is: a claim, not a verified result. Nobody has independently confirmed it yet, and the Lean Zulip is still digging in. The Navier-Stokes repo says the certificate whose hash the paper cites is not in the repo yet, the 128-bit interval-arithmetic check sits outside the formalization, and a November 20 patch was applied but "the patched package was not rebuilt" (https://github.com/tristanbuckmaster/navier-stokes). Aaronson points out that Lean kernel bugs have allowed proofs of 0=1 before, but no AI result with a Lean certificate has been retracted, and Aaronson bets this one stands. The author also works at Anthropic, which priced its IPO this week. Alpöge raises the timing directly: there was "a constraint to respect," and the "obvious cynical view" about the timing "is false." Hold all of that until outside verifiers finish. A same-day arXiv paper puts a number on what verifiable artifacts cost. On ARC-AGI-2, demanding an executable program instead of a bare answer cut the score from 93 to 68 of 120 tasks, but because a program can be checked, runs could stop early and save up to 38 to 49 percent of budget (https://arxiv.org/abs/2611.25896). Verification is expensive, and it is the only part that scales.

My read: the referee in mathematics just changed from a person to a kernel. If this holds, the hard problems move. One is Lean's own soundness, which becomes the most important security surface in math. The other is human digestion, which Aaronson thinks will take years. A proof that is true but that no one understands is a strange kind of knowledge, and this is the first time the question has stopped being hypothetical.

This article was written with Claude, an Anthropic model.
← Previous
Nerfed Daily: 2026-11-20
Next →
METR's New Chart: Models Beat the Experts, Then Misjudge Their Own Work
← Back to all articles

Comments

Loading...
>_