Anthropic mathematician claims he proved the Riemann hypothesis with Claude

Levent Alpöge, a mathematician at Anthropic, says he and Claude have solved the most famous open problem in mathematics: the 167-year-old Riemann hypothesis. There's no paper, just a four-page note and a computer proof written in Lean that no outsider has checked yet.

Research Nov 20, 2026 8 min read Source: Hacker News Startseite
Illustrative image, AI-generated

What's going on

On November 20, 2026, a four-page PDF titled "Remark" appeared on the personal website of Levent Alpöge. Alpöge is a mathematician who works at AI company Anthropic. The note opens with a bombshell:

"I believe I have, in collaboration with Claude, discovered a proof of the Riemann hypothesis."

The Riemann hypothesis is widely considered the most important unsolved problem in mathematics. It's one of the seven Millennium Prize Problems, each carrying a $1 million prize from the Clay Mathematics Institute. According to Scientific American and New Scientist, the statement comes from Alpöge himself. The site officechai points out, however, that the note carries no author name and that Alpöge didn't initially share it on social media or anywhere else. Still, it's written in the first person and hosted at alpo.ge/remark.pdf. The link quickly hit the Hacker News front page.

What the Riemann hypothesis is

German mathematician Bernhard Riemann proposed the conjecture in 1859. It's about prime numbers, numbers like 2, 3, 5, 7, or 11 that are divisible only by 1 and themselves. On the number line, they look randomly scattered. Riemann found a function, the zeta function, that encodes how the primes are distributed. The hypothesis says that all "nontrivial" zeros of this function lie on a single vertical line. Zeros are the points where the function equals zero, apart from a few easy-to-find exceptions. That line is called the critical line: in the complex plane, its real part is 1/2 everywhere.

If the hypothesis is true, the primes are distributed as regularly as they possibly can be. Thousands of theorems in number theory have so far only been proven on the assumption that it holds. A proof would secure all of them at once. Scientific American calls the conjecture the "North Star" of number theory. Progress has been so rare that many top researchers barely spent time on it, because no path forward was in sight.

The note goes further still. It claims the same method also proves the generalized Riemann hypothesis, which covers a whole family of related functions known as Dirichlet L-functions. With more time, the author says, he expects to reach the extended Riemann hypothesis as well.

A proof that exists only as code

What's most unusual is how the proof was released. There's no paper. Instead, the entire argument is said to be formalized in Lean, a programming language in which a computer checks every single logical step of a proof. According to the note, the formalization is complete and needs no additional assumptions. The files sit on the same website, split into 14 compressed archive parts.

A readable version is supposed to follow later. According to New Scientist, Alpöge writes that it would be "obviously inappropriate to publish a giant unreadable imposition of 'slop,'" meaning low-quality AI text. A footnote says the informal AI-generated drafts grew to more than 250,000 pages.

In principle, a complete Lean proof that compiles without errors and relies only on standard axioms would be extremely strong evidence. Axioms are the basic assumptions all of mathematics is built on. But the approach has weak spots: for example, you have to check that the final theorem really states the Riemann hypothesis and not something similar. officechai criticizes the note for not giving the size of the formalization, build instructions, or the axioms used. Kevin Buzzard of Imperial College London, one of the best-known Lean experts, told New Scientist:

"Let's see if it compiles first. It might take a while to find the experts who can explain the basic idea of the argument to us – assuming it all checks out, which it probably will, because Alpöge would look pretty embarrassed if the Lean code doesn't compile or is flawed in some way."

How the proof is supposed to work

The note includes a 20-step timeline that, according to the text, Claude put together. It runs from early October to mid-November 2026. The last step, the generalized Riemann hypothesis, is dated November 16.

The strategy comes from the world of function fields, number systems built from polynomials over finite fields. There, an analogue of the Riemann hypothesis was proven long ago, including with Stepanov's method. That method forces a specially constructed auxiliary function to vanish "very strongly" at certain points, which mathematicians call high multiplicity. That requires derivatives, and ordinary integers have no natural counterpart to them. That's one reason the approach has never worked for the classical hypothesis.

The note credits Claude with the way around this. Instead of derivatives, the argument uses a long chain of theta lifts: operations that move so-called automorphic forms (highly symmetric functions) back and forth between different symmetry groups. Each lift is supposed to add one unit of multiplicity. The chain is roughly as long as the square root of the size of the primes involved. The proof also draws on many tools of modern number theory, including relative trace formulas, Iwasawa theory, Bessel periods, and a weakened form of the Vandiver conjecture.

Claude's role and the author's caveats

The note is strikingly generous in its praise for the model. It describes the work as a joint project and says that toward the end Claude was far ahead of the author. Claude is credited, for example, with the length of the lift chain. On the higher-dimensional automorphic forms, Claude is treated as the real expert, since the author admits he hasn't mastered them himself.

The note is just as candid about its weaknesses. The timeline is rough, the argument may contain strange, suboptimal choices, and it's very long. No one has looked for a shorter proof yet. A footnote concedes that the descriptions of the parts where Claude is the expert were paraphrased from an AI summary and "may not make sense." Some steps contain open questions or intuitions rather than finished statements.

The acknowledgments list Princeton mathematicians Manjul Bhargava and Peter Sarnak, number theorist Maksym Radziwiłł, and Anthropic employees Jarred Sumner and Ralph Furman. officechai stresses that being thanked doesn't mean these people have checked the proof.

Odd timing

The author himself calls the timing "terrible" and admits there's an obvious cynical reading, which he says is wrong. According to New Scientist, Anthropic went public on November 19, one day before the release. Alpöge hints that this is why he waited. He says he spoke with Anthropic CEO Dario Amodei and that there was "a constraint that had to be respected." The note only uses the first name "Dario." The decision to go public now was his own, he says. Anthropic's leadership and a few trusted mathematicians signed off on the approach.

On the same day, according to officechai, Alpöge and NYU mathematician Tristan Buckmaster also released a purported proof that the Navier-Stokes equations of fluid dynamics can "blow up" even without an external force. That's another Millennium Problem. If both claims hold, two Millennium Problems would have fallen in a single day.

The AI race for the primes

This didn't come out of nowhere. Six weeks ago, OpenAI released 722 AI-generated manuscripts at once, including one titled "The Quasi-Riemann Hypothesis." It narrowed the region where zeros off the critical line could hide, but came nowhere near proving the hypothesis. Rutgers mathematician Alex Kontorovich said at the time that a human would have won the Fields Medal, math's highest honor, on the spot for that result.

Anthropic presented its own attempt in August. According to TechSpot, an unreleased research version of Claude fell short of the full hypothesis. Instead, it showed that at least 67.2 percent of the zeros lie on the critical line, up from 41.6 percent. Along the way, Claude discarded 650 ideas, and 60 Claude agents working in parallel checked approaches and results. The human researchers reportedly gave hardly any mathematical hints, mostly encouragement like "believe in yourself" and "keep going." Oxford's James Maynard praised a "genuinely new idea" at the time, but stressed that even on an optimistic view, none of these approaches led to the actual hypothesis.

Alpöge has already made a name for himself with AI results. In July, he used the model "Claude Fable" to disprove the Jacobian conjecture with a counterexample just 216 characters long. In October, he shared a refutation of the Exponential Time Hypothesis from complexity theory that Claude had found on its own.

Our take

If the proof holds up, it would be a historic moment for mathematics, and proof that AI systems can solve problems generations of humans have failed at. Right now, though, it's just a claim: an unsigned note on a personal website, with no readable proof and no independent review. The history of the Riemann hypothesis is littered with failed proofs. What argues for taking it seriously is Alpöge's track record over the past few months and his choice of Lean. Whether the code compiles can be settled fairly quickly. The harder question is whether the formal theorem really means the Riemann hypothesis.

Whatever the outcome, the case exposes a fundamental problem. Terence Tao has long warned about AI proofs as "sealed black boxes": they tell you that something is true, but not why. A proof only a computer understands also runs counter to the call by 25 Fields Medalists for slower, carefully written-up publications. And the fact that the note dropped a day after Anthropic's IPO leaves a bad taste, even though the author explicitly rejects that reading.

Sources

#anthropic#claude#mathematics#riemann hypothesis#lean
Read the original source

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.