AI

Riemann hypothesis: Claude proof meets an all-human rival

Key takeaways

  • Alpöge says Claude helped him prove the Generalized Riemann Hypothesis, released as Lean code split into 14 data packages
  • Stanford's Ravi Vakil posted an all-human draft by Alessandro Masullo within hours, and problems were then reported with the help of AI models
  • A proof would turn hundreds of papers that assume the hypothesis into established theorems, though community acceptance is still required

Two claimed proofs of the Generalized Riemann Hypothesis appeared within hours of each other: one built by an Anthropic mathematician working with Claude, the other written entirely by a human at Stanford. Neither has been accepted, and the question for readers is what exactly is being claimed, why it matters, and how mathematicians decide who is right.

What the Riemann hypothesis actually is

Prime numbers are whole numbers divisible only by one and themselves: 2, 3, 5, 7, 17. They are the atoms of arithmetic, because every other whole number is a unique product of primes (15 is 3 times 5). Euclid proved around 300 BC that there are infinitely many, but their positions on the number line look random. They thin out as numbers grow: 25 primes lie between 0 and 100, but only 16 between 1,000 and 1,100.

Carl Friedrich Gauss found a formula that roughly estimates how many primes fall in a given range. Real primes keep drifting from that estimate. Bernhard Riemann, a student of Gauss, set out in 1859 to calculate the error exactly. His tool was the Riemann zeta function, which takes the coordinates of a point on a two-dimensional map and returns a value. At certain points the result is exactly zero. Those are the zeros.

Riemann's conjecture is that all the key ("non-trivial") zeros sit on one vertical line, at the horizontal value of one-half. David Hilbert listed it among the most important open problems in 1900, and since 2000 it has carried a $1 million prize as one of the Clay Mathematics Institute's seven Millennium Problems. The Generalized Riemann Hypothesis (GRH) extends the same ordering principle to a large family of related equations. Both new claims target GRH.

How the zeros explain the primes

Think of an orchestra. Rutgers University's Alex Kontorovich puts it this way: "The Riemann hypothesis says all these different instruments in the orchestra of the primes play at exactly the same volume." Each zero acts like an instrument whose vertical position sets the pitch and whose horizontal position sets the volume. Together they correct Gauss's estimation error, much as a musical note can be broken into overtones. If every zero sits on the one-half line, the volumes match and apparent randomness becomes strict harmony.

How the two proofs were built

Levent Alpöge, a mathematician at Anthropic, said he worked with Claude over several months. He set the direction and combined two approaches on a simplified toy model. Once they moved to the real problem, he says, the AI took over and left him far behind near the finish line.

The pair extended an idea from the mathematician Stepanov, who solved a similar question in another mathematical system. That method relies on something like the derivatives of high-school calculus, which normally fails on whole numbers because individual values cannot be differentiated like a continuous line. Alpöge and Claude got round this by chaining special mathematical transformations together to mimic a derivative. The model supplied the key insight: for a number sequence of length X, you need to chain exactly as many transformations as the square root of X.

The rival draft is by Alessandro Masullo. Stanford's Ravi Vakil, president of the American Mathematical Society, posted it in a "hurried announcement", writing "It is all human, with no AI." Brian Conrad, a Stanford number theorist, had reviewed it for months, and Masullo had planned a year at Stanford refining it before going public. Anthropic's announcement forced an early release.

How a proof gets checked, and where this one stands

The two claims arrive with different kinds of backing. Masullo's draft has months of vetting by Conrad. Alpöge's has full computer verification in Lean, a programming language that works like a digital referee, checking whether every deduction follows from the basic rules of mathematics without errors or unproven assumptions. Printed, Claude's argument would run over 250,000 pages, because it spells out even trivial steps, such as a number other than one-half being larger or smaller than one-half. Alpöge released it only in Lean, split into 14 data packages.

Lean has one blind spot. If the original question was translated into code incorrectly, the proof could be airtight for the wrong problem. Scientific American reports that checking Alpöge's starting statement should not take experts long.

Vakil's own post carries a warning. About eight hours after publishing, he updated it to say several people had reported problems in Masullo's proof, found with the help of different AI models. He added that all issues must be resolved "to the satisfaction of the community" before acceptance and made no claim about correctness. He urged mathematicians to review the draft, translate it into systems like Lean and close small gaps.

Who cares, and why

If either proof holds, number theorists gain precise control over the distribution of primes along the whole number line. Kontorovich says hundreds of papers assume the Riemann hypothesis is true, and all of them would become established theorems overnight. Because both approaches cover GRH, the effect would reach fields from simple equations to high-dimensional geometry.

The links run further. Primes underpin modern digital encryption. In 1972 the physicist Freeman Dyson found that the spacing of Riemann's zeros matches energy levels inside atomic nuclei, and the same formula recurs in random particle motion, chaos theory and black hole calculations.

For the US, the story involves Anthropic, Stanford and the AMS directly.

What the limitations are

Alpöge is candid. He says Claude was so far ahead in some complex subfields that he is presenting a result whose surrounding mathematics he has not fully mastered. He expects the huge proof can be shortened and plans a human-readable version over the coming months.

The timing is awkward. The announcement came a day after Anthropic's major IPO on Thursday. Alpöge says he made the breakthrough late Sunday night and contacted CEO Dario Amodei for clearance. He calls the timing "extremely unfortunate" but rejects any suggestion of a PR stunt, saying waiting any longer was not an option.

It also follows a run of claims this year. In July Alpöge published a Claude-generated solution to the Jacobian conjecture. September brought a dispute over the Navier-Stokes equations, when NYU's Tristan Buckmaster suggested OpenAI had used prior work by him and Alpöge. In October OpenAI released results on prime numbers, and Alpöge credits their problem selection for giving him key ideas.

What this means in practice

Treat both as claims under review. A Lean-checked proof is a stronger form of evidence than a long argument read by humans, but the community still has to confirm that the problem was stated correctly. The all-human draft already has reported problems. The useful signal over the coming weeks will be which objections stick and whether the Lean statement survives expert scrutiny.

Where this is heading

Alpöge expects to extend his method to the broader Extended Riemann Hypothesis given more time, and Vakil wants Masullo's draft translated into Lean. Watch for the shortened, human-readable version Alpöge has promised and for the community's verdict on both drafts.

Frequently asked questions

What is the Riemann hypothesis?

It is a conjecture from 1859 that all the key zeros of the Riemann zeta function lie on a single vertical line at the value one-half. It would give precise control over how prime numbers are distributed.

Has the Riemann hypothesis been proved?

Not accepted yet. Anthropic's Levent Alpöge and Stanford's Alessandro Masullo each published a claimed proof of the Generalized Riemann Hypothesis, but the community has not signed off on either.

What is Lean and why does it matter here?

Lean is a programming language that acts as a digital referee, checking step by step that every deduction follows from maths' basic rules. Alpöge released the proof only in Lean, because the printed version would run over 250,000 pages.

What is the Generalized Riemann Hypothesis?

It extends the ordering principle of the Riemann hypothesis to a large family of related equations. Both the Anthropic and Stanford claims target this broader version.

More from Future Technology HQ