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)
ReleaseInsight · On RadarGuardian→ original

Axiom Math's AI Writes 43.1 Million Lines of Lean Proof in 24 Days

The classification of finite simple groups is one of the largest proofs in mathematics, and many scientists worked on it for decades. Axiom Math's team used AI to write it in Lean in 24 days, producing 43.1 million lines of code in which a small kernel program checks every step. Along the way, the team found a flaw in the published argument for one key theorem. The takeaway for engineers: pair AI with an independent mechanical checker.

AI-processed from Guardian; edited by Hamidun News
Axiom Math's AI Writes 43.1 Million Lines of Lean Proof in 24 Days
Source: Guardian. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceGuardian ↗
Editor-in-Chief
Key metric #1Lean accepts or rejects every step, so an elegant but wrong derivation won't pass
Key metric #2For the flaw in the old argument, the team proposed an alternative argument
Key metric #3Humans still check that the formal statement means what was intended

Imagine a project in which several generations of mathematicians wrote one enormous proof, and now it has to be verified in full, down to the last step. That is exactly what the team at the startup Axiom Math did: with AI, it translated the proof of the classification of finite simple groups into the Lean language. The result is 43.

1 million lines of code, and the whole effort took 24 days. Ken Ono, a professor at the University of Virginia who is currently on leave and serving as chief mathematician at Axiom Math, described this in a column for the Guardian. For developers, the important point is not the philosophy but the working technique: how to trust a result produced by a neural network.

What Lean is and why 43.1 million lines matter

The classification of finite simple groups is one of the largest proofs in the history of mathematics. It was announced in 1983, many scientists worked on it for decades, and a major gap in the proof was closed only in 2004. Reading and rechecking an array of that size by hand is very hard.

Lean is a language in which a small kernel program checks every step of reasoning. You cannot talk it into anything: it either accepts a step or rejects it. So an elegant but wrong derivation simply won't pass in Lean. According to Ono, Axiom Math uses AI to formalize, check, and extend proofs. This work can expose gaps and makes proofs available for verification and reuse.

The programming analogy is simple: a neural network writes code, and a compiler and tests filter out the bad output. Here, the role of the tests is played by a formal checker that does not depend on the model.

What they found along the way

The team released the first complete formalization of the classification in Lean. AI helped produce those 43.1 million lines. In the process, they found a flaw in the published route to one of the key theorems, and the authors proposed an alternative argument. The final proof is checked by the Lean kernel. Ono stresses that the mathematics itself belongs to the generations who created it. Ready-made formalizations, in turn, help build and verify new results.

For practitioners, this is a useful lesson. Formalization works here as an audit: it found a hole in a long-published argument that had gone unnoticed in ordinary reading. If an error in your system is costly, such as in financial calculations, medical algorithms, or critical infrastructure, generation by a model alone is not enough. You need a separate mechanical verifier.

Context and honest limitations

Ono states the main limitation: the Lean kernel confirms that the derivation is correct, but it does not guarantee that the statement of the theorem says what was intended. People still check this. The same holds in development: a test can be green while checking the wrong behavior.

The column opens with a personal admission. Ono writes that last year he nearly quit mathematics. He refers to a report that on November 20, Levent Alpöge, together with Claude, proved the Riemann hypothesis, the 167-year-old problem about the distribution of prime numbers, and that the Lean and Mathlib communities have already certified the formal proof. He adds that people are only beginning to make sense of the argument itself. This is Ono's retelling of that report; the text contains no independent assessment of such claims.

Ono also mentions OpenAI's announcement about solving the Navier–Stokes problem. Two days after it, a scheduled lecture of his at Berkeley turned into a gathering of nearly two and a half hours. Students asked about authorship, transparency, hiring, and how mathematicians whose work came first are treated.

Ono cites the statement "A Severe Misalignment of AI in Mathematics," signed by Terence Tao and other Fields Medal laureates. It warns against hasty announcements, weak attribution, and the loss of mathematical understanding. Ono shares these concerns. In his view, universities, journals, employers, and funders must decide what they support and reward. Students and those who choose to work without AI should have a voice in those decisions. He puts his own position this way: hope is not a plan, and nostalgia is not mentorship.

In his own mentoring, Ono will ask what a young mathematician understood, came up with, or made possible. Access to a powerful system is not enough for that. And in the publications he is involved in, the reader should be able to see what was proved, what was taken on faith, and who authored the ideas.

Three practical conclusions follow. First, AI results in exact fields will increasingly come with a machine-checkable certificate. Second, formalization can serve as an audit of old results. Third, questions of attribution and transparency will have to be handled as seriously as the technical ones. Ono acknowledges that he may be wrong in his predictions and says he is still learning to be a mathematician under new conditions.

Story

Key Questions Answered

What Lean is and why 43.1 million lines matter?

The classification of finite simple groups is one of the largest proofs in the history of mathematics. It was announced in 1983, many scientists worked on it for decades, and a major gap in the proof was closed only in 2004. Reading and rechecking an array of that size by hand is very hard. Lean is a language in which a small kernel program checks every step of reasoning. You cannot talk it into anything: it either accepts a step or rejects it. So an elegant but wrong derivation simply won't pass in Lean.

What they found along the way?

The team released the first complete formalization of the classification in Lean. AI helped produce those 43.1 million lines. In the process, they found a flaw in the published route to one of the key theorems, and the authors proposed an alternative argument. The final proof is checked by the Lean kernel. Ono stresses that the mathematics itself belongs to the generations who created it. Ready-made formalizations, in turn, help build and verify new results. For practitioners, this is a useful lesson.

What are the key details on Context and honest limitations?

Ono states the main limitation: the Lean kernel confirms that the derivation is correct, but it does not guarantee that the statement of the theorem says what was intended. People still check this. The same holds in development: a test can be green while checking the wrong behavior. The column opens with a personal admission. Ono writes that last year he nearly quit mathematics.

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…