Subscribe
AI & Math
4 min read

Two Proofs of the Riemann Hypothesis, One Day: Anthropic's Claude and a 940-Page Human Manuscript

Within hours on November 20, math's most famous open problem got two proposed solutions, one machine-formalized with Claude and one written entirely by hand at Stanford. Neither is verified yet, and the community is now judging both.

Mid-century Olivetti-style poster, teal chalkboard with zeta marks beside bold text reading TWO PROOFS ONE DAY
Mid-century Olivetti-style poster, teal chalkboard with zeta marks beside bold text reading TWO PROOFS ONE DAY

On November 20, Levent Alpöge, a number theorist on Anthropic's technical staff, posted a four-page note to his personal website with a sentence that has since ricocheted through every mathematics chatroom on the internet: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." Hours later, Ravi Vakil, a Stanford professor and president of the American Mathematical Society, posted a rival manuscript to his course blog with an equally loaded qualifier: "It is all human, with no AI."

One of the most famous open problems in mathematics, posed by Bernhard Riemann in 1859, therefore received two proposed solutions within the same news cycle. Neither has been verified. Both will be judged, and the episode is already a working referendum on what counts as mathematical evidence in 2026.

Two artifacts, two regimes of trust

The Anthropic side ships as machine-checkable text. Alpöge's note is accompanied by a formalization in Lean, the open-source proof verification language, which the MIT professor Scott Aaronson counted at 17.4 million lines in a GitHub archive. There is no journal submission and no conventional paper. The claim also lands one day after Anthropic's record stock market debut, a sequencing Alpöge himself flagged: "The precise timing is horrible," he wrote in the essay, which per TechCrunch describes messaging Anthropic CEO Dario Amodei on Sunday night to ask how to announce the result. He has since said he consulted Amodei before releasing it and acknowledged the timing could be read cynically, according to AFP.

Early expert reaction is leaning toward belief. Aaronson, writing on his blog, noted that no AI result carrying a Lean certificate has had to be retracted so far, while several without certificates have been, and said his bet is that the proof will stand. Number theorists have begun unpacking the argument in a Lean Zulip channel; one, Chris Birbeck of the University of East Anglia, described the core step as an automorphy-lifting argument drawing on a stack of heavy machinery from the field, a comment TechCrunch reported drew about twenty astonished reactions from readers of the channel.

Formalized machine proof versus a handwritten manuscript, Olivetti-style illustration

The Masullo side inverts the arrangement: a human-readable proof with no formal certificate. Vakil's post links a manuscript titled "The Music of the Primes," roughly 940 pages per Quanta, in which Alessandro Masullo, a Stanford-connected postdoc, proposes a proof of the Generalized Riemann Hypothesis, the stronger statement covering Dirichlet L-functions. Vakil wrote that Brian Conrad had been examining it closely for months and had given extensive feedback that was still being incorporated. The original plan was to bring Masullo to Stanford for a year to polish the writing, then announce after wider review. The Claude claim changed the calculus. "Now it seems best to put out the current draft," Vakil wrote.

Masullo, who per TechCrunch describes the work as twelve years in the making, added an introduction aimed at colleagues rattled by the AI results: "If any message at all this work is meant to send to the mathematician who feels discouraged by the pace and tone of the current climate, it is this: we are still here. We are very much alive. We are not going anywhere."

What happens next is different for each proof

The Anthropic result can in principle be machine-checked quickly. DataCamp's analysis of the release argued that whether the Lean files build could be settled within days, and whether the formalized statement actually matches the hypothesis within weeks, while noting nobody outside Anthropic had yet publicly reported a successful build. Human digestion is the slower part: mathematicians, Aaronson wrote, "will be satisfied when, and only when, they're able to explain and distill and reconstruct the ideas for themselves," and that will probably take years.

The Masullo draft needs old-fashioned refereeing. Vakil explicitly asked the community to check the manuscript, formalize it, and fill any small gaps that surface. The TLDR roundup of the release noted the same: the draft went out early specifically to invite prompt community checking.

Two proposed proofs awaiting judgment, Olivetti-style illustration

Some context on how fast the ground has moved. In August, Anthropic said an unreleased Claude model had improved a long-standing bound on zeros of the zeta function, from 41.6 to 67.25 percent of nontrivial zeros shown to lie on the critical line, and the company's own post said it did not expect those techniques to prove the hypothesis itself. In early November, Anthropic researchers settled a nearly 20-year-old conjecture on quantum speedup, and Alpöge, with NYU's Tristan Buckmaster, released a related Navier-Stokes result that had helped kick off the frontier labs' math race this summer. A full Riemann proof, if it survives, is a different tier of achievement, and the existence of a competing human manuscript means credit for it may not settle on either camp cleanly.

The Clay Mathematics Institute and journals now face a queue: two proposed proofs of the same problem, one written by a model and checked by a machine, one written by hand over twelve years, arriving hours apart on the same Friday.

Sources

AI & MathAnthropic