Saturday, Nov 21 | --:--
Back to home

Riemann Proof Claim: More Lean Checks Pass, Including One That Avoids Mathlib's Zeta

Updated Nov. 21: MIT's Andrew Sutherland reports that every module of Levent Alpöge's claimed Lean proof of the Riemann Hypothesis has now been replayed through the Lean kernel, that a second kernel accepts every compiled module, and that an independent statement-matching check confirms all five headline theorems. Combined with a separate formalization of Robin's and Lagarias's criteria, the proof also yields two divisor-sum inequalities that do not mention the zeta function. The lake comparator run is still in progress, and there is no readable proof or peer review. Earlier: Alpöge, an Anthropic mathematician, announced the Claude-assisted result on November 20.

5 min read San Francisco View as Markdown
A handmade editorial collage on torn ivory paper in cobalt and vermilion: a very large slab of interlocking cobalt, slate-grey and cream cut-paper tiles sits on torn paper, joined along one single jagged vertical seam a little left of centre, with only tiny vermilion scraps at the far corners; an official Anthropic logo is glued on a torn-paper clipping whose top edge is tucked under the slab's lower edge, with a soft shadow from the slab falling on the clipping, the full logo visible.

Updated November 21, 2026 (evening): More of the mechanical checks on Levent Alpöge's claimed proof have now finished, and one of them tests the result in a way that does not depend on how Lean's library defines the zeta function. The new section directly below covers it; the original report follows unchanged.

November 21: the checks widen, the reading has not started#

The verification is moving faster than the mathematics. MIT's Andrew Sutherland updated his verification page at 12:55 UTC on November 21. It now reports that every one of the 175,187 compiled modules has been replayed through the Lean kernel with no rejections, and that lean4lean, another kernel implementation, accepts every compiled module; five needed a raised step limit to finish. A version of con-leche with a --compare mode, which Sutherland describes as an independent implementation of Comparator's statement matching, confirms all five headline theorems against the published challenge file, which imports only Mathlib. The lake comparator run itself and a Lean-kernel replay of the exported proof were still in progress; Sutherland says the comparator's statement comparison and axiom check had already passed in one run.

The new check is aimed at the obvious objection to any Lean proof of a famous statement: that the formal statement might not mean what it seems to. Sutherland combined Alpöge's proof with Robin1984, a separate formalization of Robin's and Lagarias's criteria by another author. Those classical results turn the Riemann Hypothesis into inequalities about the sum of a number's divisors. Importing both gives two theorems "with no hypotheses and no mention of zeta": Robin's inequality for every n above 5,040, and Lagarias's inequality for every positive n. In runs logged after his previous update, both compile using only Lean's three standard axioms, and con-leche and nanoda accept the combined export. A comparator run on that export was still going. Sutherland notes the statements still rely on Mathlib's definitions of the divisor sum, Euler's constant, harmonic numbers, exp and log, and on the Robin1984 proof, which he put through the same checks.

None of this is a reading of the proof. Terence Tao said on November 20, the BBC reported, that "the digestion process will take some time, though in this case at least we have a human author to communicate with." TechCrunch reported the same day that Stanford's Ravi Vakil posted a separate, human-written draft proof of the Generalized Riemann Hypothesis by postdoc Alessandro Masullo, which has not been formalized or checked. Anthropic did something smaller in September, when Claude agents formalized an existing proof of Fermat's Last Theorem in Lean; this claim is different in kind because no accepted human proof exists to follow.

Our read: the divisor-sum version is the most useful new evidence, because it makes the target statement checkable by anyone who knows Robin's criterion without trusting one library definition. It still says nothing about whether the argument can be understood or published.

Limits of the November 21 update#

  • We have not run any checks; all results are from Sutherland's page as of 12:55 UTC on November 21, and the comparator runs were unfinished.
  • Sutherland repeats that the build ran on cloud machines without a sandbox and that the exported proof came from that build.
  • Tao's remark is as quoted by the BBC; we did not locate his original post. Masullo's draft is described from TechCrunch's report and has not been checked.
  • There is still no human-readable proof from Alpöge, and Anthropic had not commented in the reports we read.

The biggest claim an AI-assisted result could make has arrived with the mechanical checks first and the mathematical ones still to come. Levent Alpöge, a mathematician at Anthropic, said on November 20 that he believes he has, "working with Claude, discovered a proof of the Riemann Hypothesis." He has published it as a Lean formalization rather than as a paper. Within hours, MIT mathematician Andrew Sutherland reported that the Lean code compiles and that independently written proof checkers accept it. What does not exist yet is an argument that number theorists can read.

What Alpöge published#

Alpöge's statement is a short document on his website, and it links the Lean source in 14 parts. He writes that the method extends to the Generalized Riemann Hypothesis and that he expects, with more time, to cover the Extended Riemann Hypothesis too. He credits the model directly: "Claude far outpaced me near the finish line."

He chose not to publish a long written version. "It being obviously inappropriate to attempt to post some massive unreadable travesty of 'slop', I decided to only present Lean," he writes, and he says he intends to write the argument up properly.

He also addresses the timing. The result came to him "late Sunday night," he writes; he messaged Anthropic chief executive Dario Amodei, and "there was a constraint to respect." The Verge reads this as holding the announcement until after Anthropic's initial public offering on Thursday, November 19. "There is an obvious cynical view, to do with the timing, but it is false," Alpöge writes.

Anthropic did not respond to requests for comment from The Verge or Nature. The Verge notes that when Anthropic described working on the problem with Claude in August, it said: "We don't expect that the techniques Claude used will lead to proving the Riemann hypothesis."

What has been checked#

Sutherland published a verification page, last updated at 22:05 UTC on November 20, that sets out what his runs, done with Claude Code, have and have not established. All 175,187 modules compile without errors under Lean v4.33.1 and Mathlib. The five headline theorems depend only on Lean's three standard axioms. Two independently written kernels, nanoda and con-leche, accept the full exported proof. The theorem being proved uses Mathlib's own definition of RiemannHypothesis, not one written for this submission. On that page, "GRH" means the statement for Dirichlet L-functions only. Two further checks, the lake comparator run and a Lean-kernel replay of the export, were still running.

Sutherland lists his own caveats. The build ran on cloud machines without a sandbox. And a passing check shows only that the formal statements follow from the axioms according to the kernels used, relying on Mathlib's definitions being what they are meant to be.

"This is indeed an historic moment," Sutherland told Scientific American after compiling the result. "There is much work to be done to understand the mathematical details of the proof."

What has not been checked#

No mathematician has yet worked through a human-readable proof, because none has been published. Andrew Granville of the University of Montreal told Nature that "what he released is garbled" and that Alpöge should have prepared a clearer write-up before announcing. Fields Medallist James Maynard told Nature: "If this fully checks out, it is clearly a historic milestone in mathematics."

The Clay Mathematics Institute, which offers $1 million for the problem, requires a proposed solution to be published and to withstand at least two years of scrutiny, gaining general acceptance, before it can be considered for the prize, The Verge reports.

Lean acceptance has been questioned before. After OpenAI's Navier–Stokes claim, a preprint argued that the Lean check did not match OpenAI's written proof. Our read: that objection is narrower here, because the target statement is Mathlib's existing definition rather than a new translation, but the mathematical community's acceptance of the result is still the decisive check, and it will take far longer than a kernel run.

Limits#

  • We have not run any of the checks ourselves; the verification status is Sutherland's, as of 22:05 UTC on November 20, and some runs were unfinished then.
  • There is no peer-reviewed or human-readable proof, and the Clay Mathematics Institute has not accepted the result.
  • Alpöge is an Anthropic employee, and Anthropic had not commented in the reports we read.
  • How much of the work Claude did, and which Claude model was used, rests on Alpöge's account; he does not name the model.

Sources#

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading