ResearchHot

An Anthropic Researcher Says Claude Proved the Riemann Hypothesis. Stanford Answered Within Hours

A researcher says he and Claude produced a proof of the hardest problem in mathematics, and the code compiles. Verification is a different job, and it has barely started.

Emily CarterEmily Carter
Heat: 1,860
Concentric spiral curves crossed by a vertical line with bright points pinned along it

Dek: A researcher says he and Claude produced a proof of the hardest problem in mathematics, and the code compiles. Verification is a different job, and it has barely started.

What Levent Alpöge actually claimed about the Riemann hypothesis

The sentence went up on a personal website on Friday 20 November, not on a company blog. "I believe I have, working with Claude, discovered a proof of the Riemann hypothesis," wrote Levent Alpöge, a mathematician employed by Anthropic.

Read the verbs. Believes. Working with. Those are his words, and Anthropic hasn't published an official announcement about the result, according to SiliconANGLE.

He also says the argument covers the Generalized Riemann Hypothesis, a broader version of the same statement, and that he expects to reach the Extended Riemann Hypothesis with more time.

The Riemann hypothesis, without the mystique

Bernhard Riemann posed the problem in 1859. It sits among the Millennium Prize problems, the seven that carry a $1 million award for a solution.

The subject is a function called the zeta function and the points where it equals zero. Riemann noticed that the interesting zeros seem to line up on a single vertical line, the one where the real part of the number is one half. His hypothesis says they all do.

So why do mathematicians treat one line in the complex plane as load-bearing? Because hundreds of theorems in number theory get proved with the hypothesis assumed true, and one valid proof would settle all of those assumptions at once, which is a rare gift in a field where individual results take years.

A false proof would leave that work standing on an unexamined premise.

Why the Claude result exists as a Lean proof

The artifact is strange, and the strangeness is informative.

Alpöge says the argument would run past 250,000 printed pages, because the model spelled out steps a human would skip. Rather than print a library, he released it as Lean code, split across 14 data packages.

Lean is an open-source programming language and proof assistant. You write a theorem and a derivation, and the software checks each deduction against the rules it encodes. If the program compiles, the formal statement follows from the axioms.

TechCrunch reports that several mathematicians, including MIT's Andrew Sutherland, confirmed the Lean solution compiles, and the piece was updated to say independent sources had checked it.

What a compiling Lean proof can't establish

This is the part that gets lost between headlines.

A compiler checks that the derivation follows from the statement you encoded. It doesn't check that the encoding matches the textbook definition of the Riemann hypothesis. Get the translation wrong at the start and Lean will happily verify a true theorem about the wrong object.

Nor does a compiling file mean a person understands the argument. A 250,000-page formal proof has no reader yet.

No human reader, anyway.

Somebody has to walk through it, and that job hasn't been scheduled.

Alessandro Masullo, Ravi Vakil and the all-human answer

Then the story turned, and it turned fast.

Hours after Alpöge's post, Ravi Vakil, a Stanford professor and president of the American Mathematical Society, published what he called a hurried announcement. Stanford postdoc Alessandro Masullo had an independent proof of the Generalized Riemann Hypothesis, Vakil wrote, and it's all human, with no AI.

The draft represents 12 years of work, and Stanford number theorist Brian Conrad had reviewed it for months before Vakil posted, which is the review timeline a serious proof gets and a rushed preprint rarely earns. Masullo was supposed to spend another year at Stanford polishing it before publication, and Vakil said Anthropic's announcement made releasing the current version the better option.

Eight hours later, reported problems in the Masullo draft

Roughly eight hours after that, Vakil updated his post.

Several people had reported problems in the proposed proof, he wrote, some of them identified with the help of different AI models. Every issue has to be resolved to the community's satisfaction before the proof is accepted. He made no claim about whether the underlying strategy works.

So the human proof isn't home either. That's the honest state of play, and it's more interesting than a contest.

How Alpöge says Claude helped find the argument

The Decoder's account of the method is worth a paragraph, because it shows where the model contributed.

Alpöge extended an idea from the mathematician Stepanov. The obstacle was that whole numbers can't be differentiated, which blocks the usual chain of transformations. Claude supplied the insight that the number of chained transformations should equal the square root of the sequence length.

He described the process as "incredible confusion with Claude", and said he messaged Anthropic CEO Dario Amodei on the Sunday night before the announcement. He also wrote that "the precise timing is horrible", since Anthropic had floated on Nasdaq the day before.

Alpöge and NYU mathematician Tristan Buckmaster released a separate proof related to the Navier-Stokes equations at the same time. Whatever else is true, it was a busy week.

What resolution of the Claude claim would require

Three things, and none of them is a press release.

First, a human exposition of Alpöge's argument that other mathematicians can read, argue with and check against the Lean file. Second, a formal statement in Lean that matches the standard definition, verified line by line by someone who knows the definition cold. Third, the community's verdict on the gaps Vakil flagged in Masullo's draft.

Until then, the accurate description is that a proof has been claimed and a Lean proof compiles. Both statements are checkable. Neither one settles the Riemann hypothesis.

Share This Story

Sources

Related AI News

A stack of agent nodes linked by a path to a server rack, with a keyhole and a breach mark on the rack
Research

An AI Agent Found a Zero-Day and Owned a University Server in 11 Minutes

An operator gave the agents one goal and no method. They found an undocumented flaw in medical imaging software and owned the server, after 148 failed actions somewhere else.

Heat: 1,520
A funnel carrying small token blocks from one model box to another, with a shield plate blocking part of the flow
Research

OpenAI Cut Reasoning Theft From 54% to 0.2%. The Rest Is an Industry Problem.

OpenAI published the before-and-after numbers for its latest safeguards and they look like a fix. The same post explains why the fix has to be repeated every week, by everyone.

Heat: 1,420
A green chip grid with one highlighted cell and a jagged signal line beside the words Blackwell fingerprints
Research

Blackwell fingerprints in six Chinese models: what a checkpoint can prove

Dek: PyTorch fills billions of empty weights with random numbers, and the GPU doing the filling leaves a mark in the finished file. Someone checked six Chinese checkpoints and found newer silicon than the rules allow.

Heat: 1,250
Radio tower sending signal arcs over a grid of connected devices
Research

Nvidia's open source 5G and 6G toolkits passed a million downloads

Nvidia spent a year giving away the software researchers use to prototype wireless networks. The download counter passed a million, and the more useful signal is what a handful of labs built with it.

Heat: 1,275