i
News
News · 2026-11-22

Claude offers a Riemann hypothesis proof; Stanford presents a human version

Dataist Dataist

A few hours after Anthropic said it had proved the Riemann hypothesis with Claude, Stanford presented a competing draft written without AI. The mathematician Alessandro Mazzullo prepared the independent draft, but mathematicians soon reported problems in it. Both results still need to be checked: confirming the hypothesis would affect hundreds of mathematical theories and strengthen AI’s role in scientific research.

Cover: Claude offers a Riemann hypothesis proof; Stanford presents a human version

How Claude reached the proof

Anthropic mathematician Levent Alpöge described the result on Friday and posted a paper and code on his website. He said he had started working on the problem as a student, seeing it as a lifetime goal and expecting a solution to take decades.

For the past few months, he had worked with Claude, setting the direction and combining two approaches on a simplified model of the problem. When the researchers moved to the original problem, Alpöge said, the model did most of the work, while he fell behind it toward the end of the proof.

The method builds on an idea by mathematician Stepanov, who solved a similar problem in another mathematical system. His approach uses a tool that resembles the derivative taught in introductory calculus. Normally, it cannot be applied directly to integers: individual values cannot be differentiated in the same way as a continuous curve.

Alpöge and Claude got around this limitation by linking special mathematical transformations into a sequence that reproduces the effect of a derivative. Alpöge said the model supplied the key idea: for a sequence of numbers of length X, perform exactly the square root of X transformations.

The claimed result concerns the generalized Riemann hypothesis (GRH), which extends the original problem about prime numbers to a broader set of related equations. Alpöge believes that with a little more time he can also prove an even more general, extended Riemann hypothesis.

Stanford publishes an independent draft

A few hours after Anthropic’s announcement, Stanford University mathematics professor and American Mathematical Society president Ravi Vakil wrote in a blog post about a separate effort: an independent draft proof of GRH prepared by mathematician Alessandro Mazzullo. Vakil stressed that the work was done without AI.

Vakil said Brian Conrad, a prominent Stanford number theorist, had spent several months reviewing Mazzullo’s work and offering detailed comments, which are now being incorporated into the text. Mazzullo had originally planned to spend a year at Stanford refining the draft before publishing it. Anthropic’s announcement prompted the researchers to release the current version sooner.

About eight hours later, Vakil updated the post: he had been told that problems in the proposed proof had been found with the help of different AI models. He said several mathematicians had been the first to point out some of them. Vakil added that the proof would not be accepted until the community was satisfied that all the problems had been resolved, and specifically declined to say that the proof strategy was sound.

What is being checked

Conrad reviewed the Stanford work over several months. Alpöge and Claude’s proof is presented in Lean, a language that lets computers check mathematical arguments step by step. Alpöge says Claude’s reasoning would fill more than 250,000 pages if printed: the model spells out even basic rules in detail. For that reason, the proof has been published only in Lean, split into 14 data packages.

Lean checks whether each logical step follows from the mathematical rules and whether the reasoning contains errors or unproved assumptions. But an error is still possible if the original problem was translated into code incorrectly. Scientific American reports that experts should not need long to check whether Alpöge encoded the statement correctly. Vakil urged mathematicians to examine Mazzullo’s draft quickly, translate it into systems such as Lean and close any small gaps.

Alpöge acknowledges that Claude moved far ahead in some difficult areas of mathematics. He is presenting the result even though he has not yet fully mastered all the mathematics involved. The mathematician expects to shorten the proof quickly and prepare a version that people can follow in the coming months.

How the hypothesis relates to prime numbers

The German mathematician Bernhard Riemann proposed the Riemann hypothesis in 1859. In 1900, David Hilbert included it among the most important problems in mathematics. Since 2000, solving it has carried a $1 million prize: the hypothesis is one of the Clay Mathematics Institute’s seven “Millennium Problems.”

The problem concerns the distribution of prime numbers—such as 2, 3, 5, 7 and 17—which are divisible without a remainder only by 1 and themselves. Every other integer can be represented uniquely as a product of primes: for example, 15 is 3 multiplied by 5. Stanford mathematician Brian Conrad compares the role of prime numbers in number theory to the role of forces in physics.

Around 300 BC, Euclid proved that there are infinitely many prime numbers. Yet on the number line they seem to appear at random, becoming less frequent as numbers grow. There are 25 primes between 0 and 100, but only 16 between 1,000 and 1,100.

In the late 18th century, Carl Friedrich Gauss studied long tables as a teenager and found a formula that approximately estimates the number of primes in a given range. But the actual counts continually deviate from that estimate.

Riemann, who studied under Gauss in Göttingen, sought a way to calculate those deviations exactly and determine the position of every prime. To do so, he used the Riemann zeta function, a kind of computational mechanism that takes the coordinates of a point on a two-dimensional plane with horizontal and vertical axes. The function returns a value; some points produce zero. These are called the function’s zeros.

Riemann proposed that these zeros explain the apparent disorder in the distribution of prime numbers and correct Gauss’s estimate. By analogy with breaking a musical note into individual sound waves, each zero can be thought of as an instrument in an orchestra: its vertical position determines the pitch, and its horizontal position the volume.

It is impossible to calculate an infinite number of points one by one. So in 1859 Riemann proposed that all the important, or nontrivial, zeros lie on a single vertical line whose horizontal coordinate is one-half. According to Alex Kontorovich of Rutgers University, the hypothesis means that every instrument in the “orchestra” of prime numbers plays at the same volume: what looks random follows a strict order.

What would change if the proof is confirmed

If either proof withstands scrutiny, number theorists will gain an exact description of how prime numbers are distributed across the entire number line. Kontorovich says hundreds of research papers assume the Riemann hypothesis is true. Once it is proved, their results will become established theorems. Since Alpöge and Claude’s work and Mazzullo’s work concern the generalized hypothesis, the consequences would reach from simple equations to high-dimensional geometry.

The connections to prime numbers extend beyond mathematics. They underpin modern digital cryptography. In 1972, physicist Freeman Dyson was having tea with a mathematician when he stumbled on the fact that the distances between zeros of the Riemann function match the energy levels inside atomic nuclei. The same formula appears in problems involving the random motion of particles, chaos theory and calculations related to black holes.

Before the latest announcements, mathematicians seemed to have reached an impasse. With the exception of little-known efforts such as the work by Mazzullo and Conrad, leading specialists had rarely worked on the problem directly: they lacked a starting point. Two years ago, James Maynard of Oxford and Larry Guth of MIT showed that only a small number of points could stray far from the Riemann line. It was the biggest progress in decades, but Maynard himself and his colleague Andrew Granville of the University of Montreal saw it as a small step and a temporary measure.

At a seminar at Harvard University in October 2024, MIT number theorist Andrew Sutherland spoke about the risk posed by AI capable of proving the Riemann hypothesis: in his view, whether mathematicians would still have work would be the least of the possible problems.

Announcements amid the events of 2026

Anthropic’s announcement came during an important week for its business, a day after the company’s major initial public offering on Thursday. Alpöge said he got the result late Sunday evening, then asked Anthropic CEO Dario Amodei for permission. He called the timing, immediately after the company went public, extremely unfortunate, but rejected the suggestion that the announcement was a publicity stunt: he said publication could not be delayed.

The news followed several mathematical results involving Claude in 2026. In July, Alpöge published a solution to the well-known Jacobian conjecture created by Claude. In September, a dispute broke out over the Navier–Stokes equations, which describe the motion of liquids and gases: New York University mathematician Tristan Buckmaster suggested that OpenAI had used his earlier work and Alpöge’s. In October, OpenAI presented an important initial result on prime numbers. Alpöge praised the OpenAI researchers’ choice of problem and said their work gave him key ideas for his own proof.

Over the course of a year, models went from unreliable tools to systems capable of overcoming serious mathematical obstacles. This is changing how the mathematical community thinks about its own work. Alpöge says he feels both sadness at losing an objective that had guided mathematicians for 150 years and optimism about future scientific discoveries.

Dataist’s take: Both announcements remain subject to review. Even a proof formalized in Lean depends on whether the original problem was translated into code correctly, while problems have already been found in Mazzullo’s draft. If a result is confirmed, its consequences will extend well beyond number theory, from cryptography to physics. For now, though, the publications show above all that mathematical ideas developed with AI and proofs produced entirely by humans must both withstand scrutiny from the community.

Daily AI news

Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.

Only what matters — every day

Follow on X