Levent Alpöge is a mathematician at Anthropic. On Friday he put a short statement on his website, and its first sentence is the kind you read twice: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.” The method goes further, he writes, all the way to the Generalized Riemann Hypothesis. The Extended version, he expects, will follow “with more time.”
Riemann posed the conjecture in 1859. It says the primes follow a hidden order, and it is one of the seven Millennium Prize Problems, each carrying a million-dollar bounty from the Clay Mathematics Institute. Back in August, an unreleased Anthropic model pushed up the bound to which the hypothesis is known to hold. A proof is a different category.
The proof is written in Lean
Alpöge published the proof in Lean, a language in which a computer checks every step. “Claude and I have also formalized the arguments without assumptions,” he writes. According to Scientific American, MIT’s Andrew Sutherland compiled the result independently. “The world has changed,” he told the magazine. Later that day, Scientific American ran a headline saying independent mathematicians consider the proof almost certainly true. The open question is whether the hypothesis itself is stated correctly in Lean, and the magazine says that should not take experts long to settle.
Humans can’t really read it yet. Alpöge told the Wall Street Journal the argument rests on “an astounding amount of mathematics” that he is still digesting himself. An AI-written long version turned into “some massive, unreadable travesty of ‘slop’” of more than 250,000 pages. He plans a proper write-up once he can simplify it.
Done on Sunday, announced after the IPO
The pieces came together late Sunday night, Alpöge told the WSJ, a few days before Anthropic’s market debut. He called the timing “horrible” and messaged Dario Amodei to ask whether he could go public. On Friday morning, the day after the first trading session, he did. “I just could not live if I waited even longer.”
He has been here before. In July he posted a Claude-generated solution to the Jacobian conjecture, and in September he was part of the dispute over OpenAI’s Navier-Stokes proof, alongside Tristan Buckmaster. The WSJ notes this would be the second Millennium Prize Problem to fall to AI, after OpenAI’s in September.
A second proof, without AI
A few hours later, Stanford’s Ravi Vakil posted a “Hurried Announcement.” Alessandro Maria Masullo has a proposed proof of the Generalized Riemann Hypothesis, Brian Conrad has been checking it for months, and it is “all human, with no AI.” The plan had been to polish it for another year. About eight hours later Vakil added an update: several people had reported problems in the proof, found with the help of different AI models.
Lean makes checking fast, understanding is slower
Masullo’s manuscript will go through months of human refereeing. Alpöge’s proof gets checked line by line by a machine, and the experts mainly have to confirm that the statement at the top is the right one. If that holds, a language model working with one mathematician has settled the most famous open problem in number theory faster than the field can read it. What nobody has yet is an explanation of why the proof works. Most mathematicians will wait longer for that than for Lean’s checkmark.
Sources
- Levent Alpöge: Remark (PDF)
- Wall Street Journal: Anthropic Mathematician Says AI Has Solved the Most Famous Problem in Math
- Scientific American: Anthropic mathematician claims to prove Riemann hypothesis, most famous problem in all of math
- Scientific American: The AI Riemann hypothesis proof is almost certainly true, independent mathematicians attest
- Scientific American: Dueling claims of Riemann hypothesis proofs emerge
- Ravi Vakil: Hurried Announcement: Alessandro Masullo GRH