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&DLatent Space (swyx)→ original

Anthropic Mathematician Claims Riemann Hypothesis Proof with Claude: 1.3 GB of Code

Anthropic mathematician Levent Alpöz says he proved the Riemann hypothesis with Claude, and published 1.3 GB of Lean code instead of a paper. Rendered as ordinary text, the code would run to about 250,000 pages, far more than any person could read. Lean checks each step itself, and Elliot Glaser put the chance of a bug in it below 1%. The takeaway for developers: when AI produces more than anyone can read, automated verification is what saves you.

AI-processed from Latent Space (swyx); edited by Hamidun News
Anthropic Mathematician Claims Riemann Hypothesis Proof with Claude: 1.3 GB of Code
Source: Latent Space (swyx). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceLatent Space (swyx) ↗
Editor-in-Chief
Key metric #1Work began on October 6; the last commit, "GRH", appeared on November 16
Key metric #2Alessandro Massullo posted his 940-page proof, written without AI
Key metric #3OpenAI's Navier–Stokes claim was criticized because its Lean code did not match the text

Anthropic mathematician Levent Alpöz says that, together with Claude, he proved the Riemann hypothesis, the most famous unsolved problem in mathematics. Instead of a paper, he published 1.3 GB of Lean code. Lean is a language in which a proof is written as a program, and a compiler checks it the way tests check ordinary code. Rendered as human mathematical prose, the code would come to about 250,000 pages, so no one is going to read it in full. For people building products on AI, this story matters more than the mathematics: it shows how to trust a result that a human physically cannot read.

What was claimed

Alpöz says his method also covers a generalized version of the hypothesis, abbreviated GRH (the generalized Riemann hypothesis). He posted a short note and the Lean code, and did not write a full paper.

According to the timeline, the work started on October 6. By November 15, 8 of 20 steps were done, and the final commit from November 16 is called "GRH". Alpöz acknowledges that the hardest parts, concerning automorphic forms on groups of high rank, were closed by Claude. He has little experience in this area himself. In his words, Claude "flew and danced" toward the finish line.

Which exact model was used, and how much a human steered the work, has not been disclosed. This is the main gap in the story.

Almost at the same time, within a few hours, Alessandro Massullo announced his own proof. It is entirely human-produced: a 940-page draft he reportedly worked on for about 12 years and which, according to reports, went through closed manual review over that whole period.

How it is checked

Lean works like a strict test runner. If the code compiles and adds no extra axioms (unproven assumptions), the proof is valid even if no one has read it end to end. Elliot Glaser confirmed that the formalized statement really corresponds to GRH, and put the chance of an error in Lean itself well below 1%. At the time of publication the code was still compiling, which took several hours.

But automated checking does not settle every question. The main risk is not a bug in Lean but a mismatch: whether the code proves exactly what is claimed in words. That is what happened with OpenAI's September claim on the Navier–Stokes equations, where critics argued that the Lean proof did not match the result as described in ordinary language. For this reason, both "AI results" on the millennium problems are still considered unverified, and confirming them will take years.

Human versus machine

The contrast is telling. The author of the Latent Space newsletter puts it this way: Claude's solution is probably correct but incomprehensible, while Massullo's solution is probably technically flawed but understandable.

This gives rise to the debate over "slop": a product that took less effort to make than the reader needs to understand it. Alpöz deliberately chose not to publish a huge unreadable text and limited himself to Lean. Whether the result should have been held back for a decade just because no one could get through it remains an open question.

There is a second lesson, about managing AI. We usually work on a "strong leads weak" model: we know how to solve the problem and guide the model. Here, according to Alpöz, it was the reverse: Claude helped at the start and then pulled far ahead in the hardest part. We have to learn to manage systems that are smarter than us in a specific domain.

Expert reaction is mixed. Some argue the result says more about how poorly we understand the objective difficulty of problems than about AI's general superiority. Others compare the approach to olympiad-style brute-force search, which humans lack the patience for. Omar Khattab points to the contrast: the model solves millennium problems yet is "so dumb" at a thousand other things.

What practitioners should do

If a task admits formal verification, such as tests, types, Lean, or contracts, it can be trusted more than a lengthy report from a model. But the verification has to match the task's wording: auditing the specification matters more than auditing the output. Since none of the claims has been conclusively confirmed and the details of Claude's work are unknown, it is premature to speak of a "solved Riemann hypothesis".

Story

Key Questions Answered

What was claimed?

Alpöz says his method also covers a generalized version of the hypothesis, abbreviated GRH (the generalized Riemann hypothesis). He posted a short note and the Lean code, and did not write a full paper. According to the timeline, the work started on October 6. By November 15, 8 of 20 steps were done, and the final commit from November 16 is called "GRH". Alpöz acknowledges that the hardest parts, concerning automorphic forms on groups of high rank, were closed by Claude.

How it is checked?

Lean works like a strict test runner. If the code compiles and adds no extra axioms (unproven assumptions), the proof is valid even if no one has read it end to end. Elliot Glaser confirmed that the formalized statement really corresponds to GRH, and put the chance of an error in Lean itself well below 1%. At the time of publication the code was still compiling, which took several hours. But automated checking does not settle every question.

What are the key details on Human versus machine?

The contrast is telling. The author of the Latent Space newsletter puts it this way: Claude's solution is probably correct but incomprehensible, while Massullo's solution is probably technically flawed but understandable. This gives rise to the debate over "slop": a product that took less effort to make than the reader needs to understand it. Alpöz deliberately chose not to publish a huge unreadable text and limited himself to Lean.

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…