Hours After Alpöge’s AI-Assisted Riemann Hypothesis Proof, AMS President Ravi Vakil Rushes Out A Rival “All Human” Proof By Alessandro Masullo

For 167 years, nobody could prove the Riemann Hypothesis. On Friday, two separate proofs appeared within hours of each other: one produced with AI, and one, its backers stress, without any AI at all.

The first was the note on Anthropic mathematician Levent Alpöge’s website. It claims a proof of both the Riemann Hypothesis and the Generalized Riemann Hypothesis (GRH), found working with Claude and released only as Lean code. A few hours later, Ravi Vakil, the Stanford mathematician who is president of the American Mathematical Society, published a short and unusually urgent post on his blog. It announced that another mathematician, Alessandro Masullo, has a proposed proof of GRH.

“It is all human, with no AI”

Vakil’s post reads like it was written in a hurry, and he says so. He explains that he was at a meeting of the AMS Executive Committee and Board of Trustees and about to chair a session lasting several hours, but that the matter was too urgent to wait.

According to Vakil, Masullo’s proposed proof has been under close scrutiny for months by Brian Conrad, the Stanford number theorist widely known for his exacting standards. Conrad has given extensive feedback, Masullo has been working it into the draft, and Vakil says the proof is “so far looking good.” Vakil spells out how it differs from the day’s other claim: “It is all human, with no AI.”

The plan, Vakil writes, had been to bring Masullo to Stanford for a year to polish the write-up, run it past enough experts, and only then announce it. Instead, he decided it was best to release the current draft immediately. He asked the community to check it quickly, formalize it, and fill any small gaps that turn up. He added that he really hopes it works, that he has great faith in Conrad’s judgement, and that there is a long story to tell later.

Vakil’s post doesn’t mention Alpöge or Claude. But it went up a few hours after the AI-assisted claim surfaced, and it explicitly abandons a year-long plan for a careful release. That makes the reason for the rush hard to miss.

Masullo is not a well-known name outside specialist circles. He is a co-author with Conrad of lecture notes on the étale cohomology of number fields, which places him in the same corner of arithmetic geometry where Conrad works. Science journalist Kevin Hartnett wrote that he had been following the story for years and that Masullo “thought he had more time but didn’t.”

Two proofs, two very different kinds of evidence

If both claims hold up, the most famous open problem in mathematics will have been solved twice on the same day. The two releases could hardly be more different, though.

Alpöge’s claim comes with no human-readable proof. The note says a written-out version would run past 250,000 pages, so only the Lean formalization has been released. Its strength is that a machine can check it. Mathematician Elliot Glazer said early on that the formal statement in Alpöge’s code is GRH, and Scientific American has since reported that MIT mathematician Andrew Sutherland independently compiled the Lean code. A successful compile still has to be backed up by experts confirming that the definitions are right and that no hidden assumptions slipped in. But it is the kind of verification that can come quickly. What the Lean-only release doesn’t offer is understanding. As mathematicians reacting to the claim have noted, even its author describes parts of it as beyond his own expertise.

Masullo’s proof is the opposite. It is a traditional human argument, written to be read, and it has already had months of attention from one of the most careful checkers in number theory. But it hasn’t been formalized, it is still an unpolished draft, and its main endorsement so far is Conrad’s private review, which Vakil describes as encouraging but not final.

Put simply, one proof can be checked by a computer but not yet understood by anyone, and the other can be understood but hasn’t yet been fully checked.

“Not a stable equilibrium anymore”

The episode is the clearest example yet of how AI is changing the way mathematics gets announced.

The traditional norm, polishing a proof for months or years, circulating it quietly among experts and announcing only once confident, is exactly what Vakil and Masullo were following. An AI-assisted announcement upended it within a day. Glazer, who first flagged that Masullo’s proof had been under Conrad’s review for months, drew the lesson directly. He said that sitting on results until they are fully confirmed is no longer a stable equilibrium, and that people need to be more willing to post sketches.

It is the second time in three months that a rival’s announcement has forced a rushed release. In September, Alpöge and NYU’s Tristan Buckmaster saw OpenAI announce a Navier-Stokes proof hours after they published their own related work, which set off a bitter credit dispute. On Friday, Buckmaster posted an unforced Navier-Stokes blow-up proof with Alpöge and explained that “after last time, we couldn’t risk waiting.” Now the pressure has run the other way, with a human proof rushed out because an AI-assisted one appeared first. The Fields medalists who warned in September about AI speed pushing mathematics into “severe misalignment” could hardly have asked for a clearer example.

Who gets the credit?

If both proofs are correct, mathematics faces a priority question unlike any it has seen. Masullo’s work clearly predates Friday, since Conrad has been reviewing it for months. Alpöge’s timeline says his work with Claude took 41 days, from October 6 to November 16, while Masullo’s draft was already under review. The two arguments also appear to be independent. Alpöge’s adapts Stepanov’s method from function fields using iterated theta lifts, and nothing in Vakil’s post suggests Masullo took a similar route.

Credit in mathematics usually goes to whoever first publishes a correct, complete proof. Here, one proof was publicly released first but can’t yet be read, and the other has been in circulation privately for months but wasn’t public until Friday. The Clay Mathematics Institute’s rules for the $1 million Millennium Prize require a proof to be published and generally accepted for two years before any award, so the formal question won’t be settled for a long time. Simultaneous independent discoveries have a long history in mathematics, though, and if both proofs hold, shared recognition is a real possibility.

What happens next

The next few days will set the direction. For Alpöge’s proof, the questions are whether independent builds keep succeeding and whether experts agree that the Lean statements faithfully capture the Riemann Hypothesis with no extra axioms. For Masullo’s, the community now has the draft Vakil hoped people would read, and with Conrad’s months of review behind it, it starts from a position of credibility.

Vakil closed his post by saying how much he hopes the proof works. A day earlier, few mathematicians would have expected to see one serious proof of the Riemann Hypothesis in their lifetimes. Now there are two in the same day, one made with Claude and one without, and checking both has only just begun.

Posted in AI