Search
Science

Anthropic mathematician says he and Claude have proved the Riemann Hypothesis

The computer-checked claim arrived alongside a rival, all-human proof of a broader version from a Stanford postdoc. Neither result has been confirmed.

A night-time study with a chalkboard of curves, handwritten notes and a glowing laptop beneath a window of scattered lights.
Illustration: ViewTime

Key Takeaways

  • Levent Alpöge says his proof, found with Claude, compiles in the Lean proof-checking language but has no human-readable explanation yet.
  • Stanford's Ravi Vakil released an all-human proof of the Generalized Riemann Hypothesis by postdoc Alessandro Masullo, the product of 12 years of work.
  • Both claims surfaced the same week Anthropic went public on Nasdaq.

Levent Alpöge, a mathematician who works at Anthropic, says he has proved the Riemann Hypothesis with help from the company's Claude AI model. The 167-year-old conjecture about prime numbers is widely regarded as the most famous unsolved problem in mathematics.

The claim has not been confirmed. Two things set it apart from the usual flood of announced proofs. Alpöge says the argument compiles in Lean, an open-source language that checks each logical step of a proof by computer. By his own account, though, there is not yet an explanation that a human mathematician can actually follow.

"I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis," he wrote in a paper released this week.

The announcement landed in an awkward week for his employer. Anthropic had just listed on the Nasdaq and become the first frontier AI lab to trade publicly. "The precise timing is horrible," Alpöge wrote in an essay about the result. He said he had messaged Anthropic CEO Dario Amodei on Sunday night to ask how to go public with it.

A proof from a company researcher, produced with the company's product and released during its stock market debut, will draw skepticism no matter how it was made. The Lean file is the strongest answer to that. Anyone can run it, and its verdict does not depend on who employs the author.

What the hypothesis says and why it matters

Bernhard Riemann proposed the hypothesis in 1859. It concerns the zeta function, a function whose behavior is closely tied to how prime numbers are spread along the number line. The Clay Mathematics Institute, which counts the problem among its seven Millennium Prize Problems with a $1 million award for each, describes the claim this way: every "interesting" solution of the equation ζ(s) = 0 sits on one particular vertical line. According to the institute, the first 10 trillion such solutions have been checked and all of them behave. A proof covering every case would explain a great deal about why primes are distributed the way they are.

The problem's difficulty has driven many specialists away from attacking it directly. Alpöge's ambitions go further still. He believes his method could extend to the Generalized and Extended Riemann Hypotheses, broader statements that much of modern number theory takes for granted. "An astounding amount of mathematics is present in the argument," he wrote, adding that it required far more machinery than he expected when he was a student.

Early reactions from mathematicians

Experts have started working through the argument in public on Zulip, a chat platform popular with mathematicians. Chris Birbeck, a number theorist at the University of East Anglia, wrote there that a key bound appears to rest on an automorphy-lifting argument. He said it draws on Galois representations attached to certain automorphic forms, base change, Clozel's purity theorem and results of Deligne and Ribet, among other tools.

For non-specialists, the takeaway is that the proof reportedly does not rely on a clever trick. It appears to pull in some of the heaviest machinery of 20th- and 21st-century number theory, the kind of tools associated with the Langlands program. That would be a surprising route to a problem usually approached through analysis. If the argument holds, it could be as interesting for its method as for its conclusion.

TechCrunch reported that most mathematicians in the discussion appear confident the formalized Lean result will hold. The outlet also said it will be a long time before the community feels it understands the proof.

That gap is worth spelling out. A Lean check is very strong evidence that each step follows from the one before. It does not by itself show that the formal statement at the top matches the hypothesis mathematicians mean, and it says nothing about why the result is true. Experts will have to audit the first point and decode the second. Until they do, the proof is closer to a verified answer key than to a piece of mathematics anyone has learned from.

Alpöge seems aware of this. He wrote that he felt some guilt about releasing the proof without a truly human-intelligible account, but said that after a lifelong fixation on the problem he "could not live" with waiting. He has committed to writing that explanation.

His essay also captures the strangeness of collaborating with a machine. He called the process a "stream-of-consciousness jaunt" through deep confusion with Claude, and wrote that the model "flew, and danced, and sparkled near that finish line." Alongside that euphoria he expressed sadness. "It is natural to feel that this is an end," he wrote, calling the problem a principal guiding light in his life. He said he hopes the argument offers a glimpse of "the miracles that lie ahead."

A human proof arrives the same week

Alpöge is not the only one with a claim. Ravi Vakil, a Stanford mathematics professor, posted a proof of the Generalized Riemann Hypothesis written by Stanford postdoctoral researcher Alessandro Masullo. Vakil stressed that "it is all human, with no AI."

The original plan, Vakil wrote, had been to bring Masullo to Stanford for a year to polish the writing and circulate it among enough experts before announcing it. "Now it seems best to put out the current draft," he wrote.

Masullo said the proof represents 12 years of work. In a short introduction, he addressed mathematicians discouraged by the pace and tone of the current climate. "We are still here. We are very much alive. We are not going anywhere," he wrote.

Because the generalized version includes the original hypothesis as a special case, the two claims would be independent routes to the same summit if both survive review. That would be a remarkable outcome for a problem that resisted everyone for more than a century and a half. It also leaves the field weighing two unusual documents at once: a draft its own backers had planned to spend a year refining and vetting, and a machine-checked proof that no one can yet explain.

The AI race on math's hardest problems

The Riemann claim follows a busy year for AI-assisted mathematics. Earlier this year, an OpenAI system produced a counterexample to the planar unit distance problem, overturning a conjecture of Paul Erdős that had stood for decades. OpenAI later announced an apparent solution to the Navier-Stokes problem, another Millennium Prize question, which asks whether the equations of fluid flow can blow up into infinite values. That result has not been settled, and Tom's Hardware noted that it was partly overshadowed by a plagiarism controversy. OpenAI has since released hundreds more mathematical results.

Alpöge and New York University mathematician Tristan Buckmaster have also released a proof connected to the Navier-Stokes equations. According to TechCrunch, that problem helped start this summer's contest between AI labs and mathematicians.

The volume is now its own problem. In a Live Science opinion piece, Kit Yates, a professor of mathematical biology at the University of Bath, argued that AI can produce arguments faster than the field can verify or absorb them. He called on journals to require authors to explain what is new, why it matters and what role AI played. He also urged universities and funders to stop treating publication counts as a stand-in for quality.

The Riemann claim tests his argument well. If the Lean check holds, correctness may turn out to be the easy part. The scarce work is understanding: human experts reading the argument, confirming that the formal statement is the right one, and turning machine-checked steps into ideas other mathematicians can use. That last step is what has historically made a famous proof matter beyond the headline.

Neither claim is close to a prize. Clay's rules require a solution to be published and to win acceptance in the mathematical community over a waiting period, a process measured in years rather than weeks. The first real milestone is smaller and more telling: whether Alpöge, or anyone else, can explain in plain mathematical terms why his proof works.

Sources

Hannah Bell

Hannah Bell writes about physics, experimental discoveries, and the ideas used to describe the physical world. Her coverage explores how researchers test theories, interpret measurements, and investigate unresolved questions. She makes unfamiliar concepts approachable while distinguishing what an experiment demonstrates from the possibilities it may open for future research.

View all stories