A mathematician who works for Anthropic says he has used the company's Claude chatbot to prove the Riemann hypothesis, one of the best-known unsolved problems in mathematics. The claim has not been verified, and the Clay Mathematics Institute still listed the problem as unsolved hours after the announcement, according to NPR.
Levent Alpöge, a member of Anthropic's technical staff, made the announcement Friday, a day after the company's stock market debut. If the proof holds up, the BBC reported, it would rank among the most significant in history.
Alpöge posted a four-page explanation on his website, NPR reported. The BBC said he worked with a non-public version of Claude over about six weeks to build a detailed proof. He wrote that Claude "far outpaced" him near the end of the work, according to NPR.
The hypothesis was formulated in 1859 and concerns how prime numbers, which are divisible only by one and themselves, are distributed as numbers grow toward infinity. The BBC described it as nearly 170-year-old and widely considered the most famous problem in the field. Prime numbers underpin higher mathematics, the BBC noted, so progress on them can have broad effects.
The proof was written in Lean, a programming language used to check mathematical proofs by computer. The two outlets describe its size differently. The BBC said the full text runs to more than 250,000 pages, while NPR said it came as 17 million lines of Lean code and also cited the figure of more than 250,000 pages.
Alpöge said he found the proof late Sunday night and then messaged Anthropic chief executive Dario Amodei to ask how he could announce it. The company was days from its public listing. NPR reported that the offering raised a record $81 billion at a $1.5 trillion valuation. Anthropic did not respond to a request for comment from The Associated Press, NPR said.
Alpöge acknowledged that the timing looks bad. He told readers that a cynical reading tied to the listing is mistaken and that he felt he could not wait any longer, the BBC reported.
For now, nothing about the problem's official status has changed. The Clay Mathematics Institute named the hypothesis a Millennium Prize Problem in 2000 and attached a $1 million reward to a solution. Institute president Martin Bridson said in a statement that the announcement was of “historic significance,” but that the path to real understanding would be long and difficult, according to NPR.
Checking a proof of this size is a slow job. Alex Kontorovich, a mathematics professor at Rutgers University, told NPR that a computer accepting the Lean code is a strong sign, but that he still has to do the work of understanding it himself. Prominent mathematician Terence Tao said verification is still some way off, adding that checking would take time but that a human author is available to help, the BBC reported.
The BBC said expert review of OpenAI's 372 claimed solutions will take months, if not longer.
NPR described this as the second time in recent months that an AI company, or someone linked to one, has claimed to crack a famed unsolved problem. In September, OpenAI said employees using a non-public version of its tools solved part of a problem about the Navier-Stokes equations, which describe how fluids move, in 88 hours. The BBC said more than 10,000 AI agents were put to the task and that the result has yet to be confirmed. A month later, the BBC reported, OpenAI released what it described as solutions or major progress on 372 more problems.
Those announcements drew criticism. A group of mathematics scholars wrote in a letter, quoted by the BBC, that AI companies boasting of major solutions to show off their products was harmful to the science of mathematics. The BBC also said Anthropic has claimed progress in mathematics and physics before.
Amodei has described an optimistic future in which AI drives advances in math and science, NPR reported, while also warning about the risks the technology poses.