AI Trends

Levent Alpöge Says He and Claude Found a Proof of the Riemann Hypothesis; Lean Code Has Yet to Be Independently Verified

Unite.AIUpdated Nov 21 00:35

AI summary

Anthropic mathematician Levent Alpöge says he and Claude found a proof of the Riemann Hypothesis and a method for solving the Generalized Riemann Hypothesis, and released the Lean code in 14 compressed archives. The argument has no conventional paper; the generated version reportedly exceeds 250,000 pages, and the announcement cites no independent review of the complete proof. Whether it holds up depends on scrutiny by the mathematical community and successful verification of the released Lean code.

Why read this

AI score 74

The report distinguishes among three layers of evidence—formalized code, personal claims, and independent review—making it easier to understand the current credibility limits of this mathematical claim.

Topics

Tags

Body

Levent Alpöge said on November 20, 2026, that he believes he and Claude discovered a proof of the Riemann hypothesis and a method resolving the Generalized Riemann Hypothesis. The Anthropic mathematician released a personal statement linking to Lean code, but identified no independent validation of the completed argument.

Alpöge is also a junior fellow at Harvard’s Society of Fellows. The Riemann hypothesis dates to 1859 and concerns the zeros of the Riemann zeta function, which describe the distribution of prime numbers. It places the zeros that determine the primes on a particular vertical line; many mathematical results assume it to obtain a form of randomness in the primes.

Alpöge says he expects, with more time, to address the Extended Riemann Hypothesis as well. That is an expectation rather than another completed result.

Alpöge says the arguments were formalized without assumptions and released as code for Lean, a formal proof system, in 14 compressed archive parts. He chose not to add a conventional write-up, describing the argument as incredibly long and the prospect of a massive generated document as unreadable; a footnote puts the generated version at more than 250,000 pages.

The release therefore pairs a formalization presented as complete with limited human-readable explanation. Alpöge says he is presenting the argument before mastering a large body of surrounding mathematics. His next steps are to understand and explain it, pursue simplifications, speak about it and write it properly. His cautions include possibly strange or suboptimal choices and timeline descriptions paraphrased from Claude that may need editing.

Alpöge describes a division of labor in which he guided the conceptual move from a toy setting to its suitable analogue, while Claude supplied expertise he lacked, particularly on automorphic forms for higher-rank groups. The argument builds on Stepanov’s function-field method, which uses a divisor to obtain multiplicity; reproducing that effect over the integers was a central obstacle. Fourier-analytic ideas led the pair to repeated theta lifts in an automorphic framework, with Claude proposing a chain of length X for an underlying sum of length X.

The rough timeline in the statement, supplied by Claude after a search, lists a Stepanov determinant on October 6, 2026, iterated theta lifts and related computations on November 15, 2026, and the generalized result on November 16, 2026. Alpöge cautions that some descriptions paraphrase Claude and may need revision. The dates therefore document his retrospective account of the research, not independent verification of the argument.

Anthropic reported on August 10, 2026 that an unreleased research version of Claude increased a longstanding lower bound for the proportion of zeta zeros satisfying the Riemann hypothesis from 41.6% to 67.2%, while explicitly stating that Claude had not proved the hypothesis and that the techniques were not expected to yield a proof. The earlier work was a partial bound; the new statement presents a complete proof through a different approach.

That earlier effort used two Claude Code sessions and 31 million output tokens. Claude initially generated 650 ideas that did not work, then coordinated about 60 subagents in a second attempt. Alpöge and Ralph Furman examined the result; Brian Conrey and Dan Goldston reviewed the paper; and the Lean formalization passed the comparator validation tool.

The new statement names no equivalent review of the full proof it describes. Independent mathematical scrutiny and successful checking of the released Lean code are therefore the decisive tests before Alpöge’s announcement can be treated as a solution to the Riemann hypothesis.

Source

Unite.AI