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 RadarIEEE Spectrum AI→ original

Axiom Math's AI agents turn a 15,000-page proof into 43 million lines of code

The proof of the classification of finite simple groups runs to 15,000 pages, and no mathematician holds all of it in their head. Axiom Math's swarm of AI agents translated it into machine-readable code, where every step is checked automatically, much like a software build. The result is 43 million lines, roughly the size of the Linux kernel. The takeaway for developers: a giant task gets cut into verifiable milestones, and here there were about 100.

AI-processed from IEEE Spectrum AI; edited by Hamidun News
Axiom Math's AI agents turn a 15,000-page proof into 43 million lines of code
Source: IEEE Spectrum AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceIEEE Spectrum AI ↗
Editor-in-Chief
Key metric #1Nine days later, Axiom formalized the Stacks project the same way: 7,500 pages, 16 million lines
Key metric #2The verification found and fixed several errors in the human-written proof
The Stacks library is openany person or bot can build on it

Mathematical proofs are usually checked the old way: colleagues read the text and look for errors. But what if the text runs to 15,000 pages? That is the size of the classification of finite simple groups (CFSG), which mathematicians call the "enormous theorem."

It is the longest proof ever written by humans. On October 25, startup Axiom Math announced that its swarm of AI agents had finished an automated formal verification of this proof. The agents translated the text into machine-readable code, where every step must follow strictly from the previous ones.

The result is about 43 million lines of code, roughly the size of the Linux kernel. Imagine a monorepo in which no test can be run until the last module is written: that is what this task looked like. For developers, the story is interesting because of how a multi-agent system was made to work on a project of this scale and produce a result that can be verified.

What was checked and why it was so hard

Finite simple groups are sets of elements with an operation such as addition or multiplication, and they cannot be broken down into simpler parts. A real-world example is all the possible permutations of a deck of cards. In the nearly two centuries since Évariste Galois, mathematicians have sorted these groups into 18 families plus 26 "loners" that belong to no family. The CFSG proof confirms that this list is complete.

But the proof is pieced together from more than 500 journal articles by roughly a hundred authors, written over half a century. No single person holds all of it in their head. Formal verification requires everything to be translated into machine form first: until all of it is ready, even a single fragment cannot be reliably checked.

Project lead Nils Voss described how the team got around this. First, they processed the entire literature and built a map of the links between statements in the informal proof. Then they split the main theorem into about 100 milestones: intermediate results that can be verified independently. These served as landmarks so the work would not fall apart.

Stacks: not a long chain but a wide network

Nine days later, Axiom reported that it had automatically formalized the Stacks project in the same way. Stacks is an open textbook and reference on algebraic geometry, maintained since 2005 by Columbia University mathematician Aise Johan de Jong. If CFSG is one gigantic theorem, Stacks is thousands of interconnected definitions, constructions and lemmas. The project is 7,500 pages long, and its formal version is 16 million lines, noticeably smaller than CFSG's.

The difficulty here is different: thousands of definitions must mean the same thing everywhere and must not drift from the source. Project lead Vasily Ilyin says the team built a hierarchy of AI agents with distinct responsibilities that work in parallel. Some make sure definitions are consistent across the whole library. Others check fidelity to the original: multi-agent audit, cross-checking against sources and strict tests. Engineers will recognize the pattern of implementers, reviewers and automated tests, only here the cost of an error is higher.

Now any person or bot can access the formal Stacks library and build new formal mathematics on top of it. Kim Ono, a mathematician at Axiom, describes the two results this way: they show two directions in which formalization should scale, one very long proof and one very wide body of connected theory.

Assessments and caveats

Colva Roney-Dougal of the University of St Andrews was not involved in the work. She admits she had thought formalizing CFSG was out of reach because of the sheer volume of computation, and she was surprised. She was especially pleased that the formal proof found and fixed several errors in the original, human-written one. In her view, the result matters for formalizing any modern research in group theory.

There is a downside too: the verification ran automatically, so few mathematicians will be able to extract knowledge from it for the community. The new generation, in Roney-Dougal's words, could "miss the chance to learn something along the way."

It is worth remembering that the figures and assessments in the IEEE Spectrum piece mostly come from Axiom's own staff, and the article contains no independent review of the code's quality.

Practical takeaways for those building agent systems:

  • Split a huge task in advance into independently verifiable milestones.
  • Consistency calls for separate auditor agents and cross-checking against sources, not a single universal executor.
  • The open Stacks library provides a testing ground for verifiable AI output.

Ono and Voss see formal mathematics as a way to confirm theorems, preserve knowledge as experts become fewer, and create a shared foundation for new discoveries. For people, they say, formalization will provide a launchpad to use and understand mathematics in new ways.

Story

Key Questions Answered

What was checked and why it was so hard?

Finite simple groups are sets of elements with an operation such as addition or multiplication, and they cannot be broken down into simpler parts. A real-world example is all the possible permutations of a deck of cards. In the nearly two centuries since Évariste Galois, mathematicians have sorted these groups into 18 families plus 26 "loners" that belong to no family. The CFSG proof confirms that this list is complete.

What are the key details on Stacks: not a long chain but a wide network?

Nine days later, Axiom reported that it had automatically formalized the Stacks project in the same way. Stacks is an open textbook and reference on algebraic geometry, maintained since 2005 by Columbia University mathematician Aise Johan de Jong. If CFSG is one gigantic theorem, Stacks is thousands of interconnected definitions, constructions and lemmas. The project is 7,500 pages long, and its formal version is 16 million lines, noticeably smaller than CFSG's.

What are the key details on Assessments and caveats?

Colva Roney-Dougal of the University of St Andrews was not involved in the work. She admits she had thought formalizing CFSG was out of reach because of the sheer volume of computation, and she was surprised. She was especially pleased that the formal proof found and fixed several errors in the original, human-written one. In her view, the result matters for formalizing any modern research in group theory.

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…