AI

A Riemann hypothesis proof made with Claude compiles, and no human has read it

Adrian Kessler
Add us on Google

Levent Alpöge, a mathematician at Anthropic, has published what he says is a complete proof of the Riemann hypothesis, the most famous unsolved problem in mathematics, and he says he found it working with Claude, the company’s AI model. There is no paper. He released only a formal proof written in Lean, a programming language in which a computer checks every logical step, and the experts who have run it say it goes through.

That reverses the usual order of mathematics. A proof normally convinces people first and gets checked by machines later, if ever. Here the machine check came first and the understanding has not caught up, not even for Alpöge, who writes that Claude “far outpaced” him near the finish line and that the argument rests on areas of number theory he has not mastered.

The hypothesis is about where the zeros of the Riemann zeta function sit, and those zeros control how evenly the prime numbers are spread along the number line. Thousands of published results simply assume it is true. Alpöge says his method also settles the stronger Generalized Riemann Hypothesis, and that the full argument is vast: a prose version he generated from it ran past 250,000 pages, by his account, which is why he released only the code. Terence Tao described it on his blog as an enormous Lean-only proof that, on early analysis, uses the full power of the Langlands program and modern arithmetic geometry.

The checking is happening in public, in the Lean community’s chat. The mathematician Elliot Glazer said the theorem in the files really is the Generalized Riemann Hypothesis, which matters more than it sounds, because a flawless formal proof of a subtly different statement proves nothing. TechCrunch reported that MIT mathematician Andrew Sutherland and others confirmed the code compiles. The number theorist Thomas Bloom said he could not decide which surprised him more, the claim itself or a coherent proof of that length.

A compiling proof is not the end of the process. Mathematicians still have to confirm that the definitions in Lean’s shared library match the classical objects, and that nothing in the archive slips in an extra assumption. Even if every line holds, a machine-verified proof that nobody can follow tells mathematics that the hypothesis is true while saying almost nothing about why it is true. Number theorists want the why, and that part does not exist yet. Alpöge himself calls the current argument long and full of choices he expects to be simplified.

The most famous prize attached to the problem is also out of reach for now. The Clay Mathematics Institute pays $1 million for each of its Millennium Prize problems, but only after a solution appears in a qualifying publication, survives at least two years of scrutiny and wins general acceptance among mathematicians. A Lean archive on a personal website meets none of those conditions today.

For everyone outside mathematics, daily life does not change. Online encryption does not depend on the hypothesis staying unproved, and most work that relies on it already treated it as true. What changes is the division of labour. In Alpöge’s telling, he chose the direction and the analogy to follow, and the model supplied machinery from fields he did not know, a pattern that turns the mathematician into the person who aims the search rather than the one who completes every step.

The same day, Stanford mathematician Ravi Vakil posted a separate, entirely human proof of the generalized hypothesis by the postdoctoral researcher Alessandro Masullo, which Brian Conrad had been reviewing for months. Vakil wrote that the plan had been to polish it for a year, but that it now seemed best to put out the current draft. About eight hours later he added that several people had reported problems in that draft, found with the help of AI models. The human proof was audited by AI. The AI-assisted proof was audited by a compiler.

The timing has its own story. Alpöge says he messaged Anthropic chief executive Dario Amodei late on Sunday night and published on Friday, the day after Anthropic’s shares began trading; he calls the timing “horrible” and rejects the cynical reading. Anthropic has not commented. In August the company said a research version of Claude had raised the share of zeta zeros known to lie on the critical line from 41.6% to 67.2%, and that it did not expect those techniques to lead to a proof of the hypothesis. Alpöge also posted, with NYU mathematician Tristan Buckmaster, an AI-assisted Lean proof that solutions to the Navier-Stokes fluid equations can blow up, the Millennium problem OpenAI claimed in September.

The next verdict will come from the Lean community, not from Anthropic or social media, as volunteers work through the archive and the definitions it rests on. Alpöge says he now intends to understand the argument himself, simplify it and write it up properly, and only that paper can start the two-year clock the Clay Institute requires.

Tags: , , , , ,

Add us on Google

Discussion

There are 0 comments.