Terrance Tao says verification is still some way off
Levent Alpöge, an Anthropic employee and mathematician, said on Friday that he used a non-public version of the company’s Claude chatbot over about six weeks to create a detailed proof of the Riemann hypothesis. The claim came a day after Anthropic made its US stock market debut.
Alpöge put forward the complete proof in Lean, a computer language used for confirming maths proofs, and the document runs to over 250,000 lines. He described the result as “one dawn across science of something far greater.” BBC News said that if the proof is found to be correct, it would be among the most significant in history.
The Riemann hypothesis is a nearly 170-year-old maths problem named for the mathematician Bernhard Riemann, and it puts forward that prime numbers have a determined order along the number line. Prime numbers are the bedrock of higher maths, and the problem’s complexity is a large part of why it is widely considered the single most famous problem in the field.
The proof has not yet been verified. Terrance Tao, a prominent mathematician, said on Friday that checking the submission is still some way off. “I imagine the digestion process will take some time, though in this case at least we have a human author to communicate with to assist with this task,” Tao said. The contrast Tao drew was with major maths proofs made earlier this year using AI, for which no such human was obviously part of the work.
Alpöge said he messaged Anthropic CEO Dario Amodei on Sunday, asking him directly about how he could announce the discovery. “The precise timing is horrible,” Alpöge said, referring to Anthropic’s public stock listing. “If I could have, I’d have finished this all much earlier.” Alpöge also addressed what he called an obvious cynical view of the timing: “I am clearly unilaterally responsible for feeling I just could not live if I waited even longer. There is an obvious cynical view, to do with the timing, but it is false,” he added.
The claim comes as AI firms like OpenAI have in recent months announced scores of solutions to major higher maths problems that have eluded scholars for decades. In September, OpenAI said employees using a non-public version of its AI tools solved part of a notoriously difficult problem in just 88 hours. That problem was part of the Navier-Stokes equations, which concern how fluids move, and more than 10,000 AI agents — computer-programmed bots that operate somewhat autonomously — were put to the task as an AI training exercise.
OpenAI’s released proof on Navier-Stokes has yet to be confirmed as accurate, and it set off criticism within academic and higher maths communities and organisations about the use of AI to solve maths problems. A month later, OpenAI released what it claimed were solutions, proofs or major progress on 372 more outstanding higher maths problems, claims that will take months, if not longer, for experts to review and prove their validity.
A group of maths scholars said AI companies boasting of major solutions to show how sophisticated their commercial products had become was “detrimental to the science of mathematics.” In a group letter, the scholars said: “We are witnessing a general threat to intellectual work, with misalignment between the outcome of the use of AI and its initial purpose.”
Anthropic has also claimed to have made other progress and solutions in maths and physics. Alpöge’s submission is a Lean document of more than 250,000 lines, and Terrance Tao said verification of it is still some way off.