AI Momentum 98.5 +1.5%GitHub Pulse 5503 +6.5%Capital Flow $204.0B +19039.3%Hiring Heatmap 0.0 -100.0%
Advertisement
Models & Research

An Anthropic researcher says Claude helped prove math's most famous unsolved problem. Software has checked it, but mathematicians have yet to review it

A day after Anthropic's record stock market debut, a mathematician at the company said he and the company's Claude AI had proved the Riemann hypothesis. It is a question about prime numbers that has gone unanswered since 1859. It is also one of seven Millennium Prize Problems, each carrying a $1 million award.

If the claim holds, it would be the strongest evidence yet that AI can do original work at the very top of a field, not just speed up routine tasks. That matters to anyone pricing AI labs on the promise of automated research. But the result arrived in an unusual form, and the evidence so far proves something narrower than the headlines suggest.

Levent Alpöge posted a four-page note on Friday, November 20. "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis," he wrote. He says the same method also proves a broader version, known as the generalized Riemann hypothesis. He credited Claude with the final steps, writing that it "far outpaced me near the finish line."

What has actually been checked

Alpöge did not publish a conventional paper. He released the argument only in Lean, a programming language whose software checks every logical step of a proof. The release runs to about 17 million lines of code. He says his attempts at a written version ran past 250,000 pages.

Outside checking moved fast. Within hours on Friday, MIT mathematician Andrew Sutherland and Carnegie Mellon's Simon DeDeo had each run the code through Lean's checker themselves. Sutherland told Scientific American the chance of anything being "fishy" was "approximately zero."

One worry remained: whether the code proved the real Riemann hypothesis or a subtly redefined version. On Saturday the Lean and Mathlib communities, the groups behind Lean and its shared math library, addressed that point. In a joint statement, they said separate tools confirmed the proved theorem matches the library's standard statement, and several independent checkers accepted the proof. They certified that "both the statement and the proof" are correct.

That is a strong result for a machine check. It is not the same as mathematicians accepting the proof. No human-readable account of the argument has been shared, and Alpöge says he has not yet mastered all of the surrounding mathematics. Oxford's James Maynard, a Fields Medal winner, told IBM Think it would be "clearly a historic milestone" if it checks out, but that understanding it will take substantial effort.

Others were sharper. Number theorist Maksym Radziwiłł told Quanta Magazine that the four-page note contains nonsensical statements and that code alone does not show why the result is true. Stanford's Brian Conrad told The New York Times that releasing a proof its author does not fully understand was "frankly completely outrageous," The Next Web reported.

Anthropic has not formally claimed it

For a company two days into public trading, the timing invites questions. Anthropic issued no formal statement on Friday and did not immediately respond to AFP's request for comment. None of the reports we reviewed say which version of Claude was used, how much computing power the work consumed, or how long it ran. Alpöge said he consulted chief executive Dario Amodei before announcing, and rejected the idea that the timing was cynical.

The market did not treat the claim as a windfall. Anthropic's shares fell about 8.5% on Friday to close near $89, amid a broad sell-off in AI-linked stocks. Its low for the day came within five cents of the $83 offering price, FinanceFeeds reported. Market coverage blamed the wider sell-off, not the math claim.

The result also lands in a credit race with OpenAI. About two months ago, OpenAI said one of its models had solved another Millennium problem, on the equations of fluid motion. Mathematicians say that proof appears correct but has not been fully vetted. Alpöge was at the center of a dispute over credit for related fluid work he did with NYU's Tristan Buckmaster; OpenAI denied wrongdoing.

There is a further twist. The same day, Stanford's Ravi Vakil, president of the American Mathematical Society, posted a 940-page, human-written proof of the generalized hypothesis by Stanford postdoctoral researcher Alessandro Masullo. About eight hours later, Vakil said several people had reported problems in it, found with AI models, The Next Web reported. His earlier attempt did not hold up, and many number theorists are withholding judgment, Quanta reported.

What buyers and investors should take from it

First, the claim is real enough to take seriously and not yet settled. Before it considers a prize, the Clay Mathematics Institute requires publication in a qualifying journal, two years of scrutiny and general acceptance by mathematicians. Alpöge says he plans to write the proof up properly, and expects the argument to be greatly simplified.

Second, it shows where trust in AI work is heading. Here, people did not read the output. They ran independent checking software on it, and that proved fast and persuasive. Businesses using AI for code, contracts or compliance face the same gap: results arriving faster than people can understand them. Automated checking is becoming the practical way to trust them.

Third, the lab-to-lab comparison remains murky. Both Anthropic and OpenAI now have Millennium-level claims linked to their models. Neither has published enough about the cost and human effort involved to let a customer judge what its products can do. Until the companies publish those details, these results say more about where AI research is heading than about what a business can buy today.

Written by AI from the public sources linked below.