Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseResearch Paper · For R&DThe Decoder→ original

Anthropic and Claude Claim Proof of the Riemann Hypothesis, Stanford Responds Without AI

The Riemann hypothesis has promised since 1859 to explain how prime numbers are distributed along the number line, which underpins encryption. Anthropic mathematician Levent Alpöge says that, together with Claude, he proved a generalized version of it and published code in Lean, a "digital judge" that checks every step. Hours later, Stanford showed an AI-free draft, but problems were found in that one too. The takeaway: AI results can be verified by a machine rather than taken on the lab's word.

AI-processed from The Decoder; edited by Hamidun News
Anthropic and Claude Claim Proof of the Riemann Hypothesis, Stanford Responds Without AI
Source: The Decoder. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceThe Decoder ↗
Editor-in-Chief
Key metric #1On paper, Claude's proof would run to more than 250,000 pages
Claude found the key ideaa sequence of length X needs √X transformation steps
Key metric #3Eight hours after Vakil's post, problems were found in Masullo's draft

Imagine an AI lab sends you a huge pull request of hundreds of thousands of lines and says: "This solves the most famous problem in mathematics." You can't take it on faith, and you can't read it all. That is roughly the situation the mathematical community found itself in on Friday, when Anthropic mathematician Levent Alpöge announced that he and Claude had proved a generalized Riemann hypothesis. For those building products on AI, this story is interesting not for its noise but for its method of verification: it shows how you can trust a model's result without trusting the model itself.

What the problem is and why it matters

The Riemann hypothesis has been open since 1859. It concerns prime numbers: 2, 3, 5, 7, 17 and so on, which are divisible only by one and by themselves. All other integers are built from them the way molecules are built from atoms: for example, 15 is 3 times 5. There are infinitely many primes, but along the number line they are scattered seemingly at random and become rarer: there are 25 between 0 and 100, but only 16 between 1000 and 1100.

Riemann proposed a tool, the zeta function, which works like a computing machine: you feed it coordinates of a point on a plane and get a value back. At some points the result is zero. These zeros, like instruments in an orchestra, explain the apparent chaos of the primes.

The hypothesis states that all the key zeros lie on one vertical line, that is, they "play" at the same volume. If this is proved, hundreds of papers that currently rely on the hypothesis as an assumption become solid theorems. The Clay Mathematics Institute has offered a million dollars for a solution since 2000.

Primes also underlie modern encryption, and the spacings between the zeros coincide with energy levels in atomic nuclei.

Alpöge claims his method also covers the generalized version (GRH), for a large family of related equations. He developed the idea from work by mathematician Stepanov, who needed an analogue of the derivative. For integers the ordinary derivative doesn't work, and the workaround was to link special transformations into a chain. Claude made the key observation: for a sequence of length X, exactly √X links are needed. According to Alpöge, by the end the model went so far ahead that he presents a result whose surrounding mathematics he has not fully mastered.

How to check 250,000 pages

This is where the engineering part begins. The proof is written in Lean. You can think of it as a compiler and test suite for mathematics: the program checks step by step that every conclusion follows from basic rules, with no errors and no unverified assumptions. On paper, Claude's reasoning would run to more than 250,000 pages, because the model spells out even elementary things like "a number not equal to one half is either greater or smaller than it." That is why Alpöge published only the Lean version, split into 14 data packages.

There remains a risk familiar to any developer: the tests pass but check the wrong thing. If the original problem was mistranslated into code at the very start, Lean will faithfully prove the wrong theorem. According to Scientific American, experts should be able to quickly verify whether Alpöge encoded the condition correctly. He himself expects the proof can be shortened considerably, and in the coming months he wants to write a version that a human can understand.

Stanford's response and the bad timing

A few hours later, Stanford professor and American Mathematical Society president Ravi Vakil posted a "hasty announcement" on his blog: an independent draft proof of GRH by mathematician Alessandro Masullo. "Everything here was done by humans, without AI," he wrote. The work had been quietly checked for months by number theorist Brian Conrad. Masullo had planned to spend another year refining the text, but Anthropic's announcement forced him to release the current version immediately.

About eight hours later, Vakil updated the post: several people had written to him and pointed to problems they had found using various AI models. He stressed that he is not asserting the strategy is correct, and that the proof will be accepted only once the community has resolved all objections.

There is also a business context: the announcement came the day after Anthropic's major IPO. Alpöge says the breakthrough happened late on Sunday evening, after which he wrote to CEO Dario Amodei. He calls the timing extremely unfortunate but rejects the idea that it was a publicity stunt: waiting, he says, was not an option.

What this means for practitioners

No proof has been recognized so far. But Claude has already solved other open problems in 2026; for example, in July Alpöge published a solution to a conjecture about the Jacobian. The main lesson for those deploying AI: the more complex the model's result, the more important formal verification performed by a machine becomes, rather than by a reviewer exhausted on page one hundred. If your product has AI producing something critical (code, a calculation, a legal formula), look for your own "Lean": a set of independent checks that do not depend on the model's reputation. And separately, verify that your tests describe the correct statement of the problem.

Story

Key Questions Answered

What the problem is and why it matters?

The Riemann hypothesis has been open since 1859. It concerns prime numbers: 2, 3, 5, 7, 17 and so on, which are divisible only by one and by themselves. All other integers are built from them the way molecules are built from atoms: for example, 15 is 3 times 5. There are infinitely many primes, but along the number line they are scattered seemingly at random and become rarer: there are 25 between 0 and 100, but only 16 between 1000 and 1100.

How to check 250,000 pages?

This is where the engineering part begins. The proof is written in Lean. You can think of it as a compiler and test suite for mathematics: the program checks step by step that every conclusion follows from basic rules, with no errors and no unverified assumptions.

What are the key details on Stanford's response and the bad timing?

A few hours later, Stanford professor and American Mathematical Society president Ravi Vakil posted a "hasty announcement" on his blog: an independent draft proof of GRH by mathematician Alessandro Masullo. "Everything here was done by humans, without AI," he wrote. The work had been quietly checked for months by number theorist Brian Conrad. Masullo had planned to spend another year refining the text, but Anthropic's announcement forced him to release the current version immediately.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…