16 Million Lines of Code Turn Modern Mathematics Infrastructure into a Navigable City
Axiom formalized the entire Stacks Project into 16 million lines of code while preserving its highly interconnected structure. Line count is only the surface; what is really worth seeing is whether the dependency relationships fall apart once scale increases.
Axiom has released a formalized version covering the entire Stacks Project, at 16 million lines. Thomas Wolf reposted the news and used it as a counterexample to vibe coding at scale. Vibe coding generally means having a model generate large amounts of code while humans rely on intuition and tests as a backstop. The tweet put it bluntly: “Most think vibecoding at scale is slop. Here’s a counterexample.”
The Stacks Project is a long-running collaborative reference project in algebraic geometry and algebraic stacks, and many people regard it as infrastructure for modern mathematics. Formalization generally means writing definitions, theorems, and proofs as code that a proof assistant can check. The difficulty is not only length: the material is highly interdependent, and if you remove one lemma, a whole surrounding area of definitions and proofs loses its context. Axiom says that this time it preserved this highly interconnected structure while also optimizing code quality. In the video, it is presented as a mathematical city, with theorems and definitions like buildings and dependencies like roads.
Sixteen million lines sounds like a typical scale narrative. Line count is certainly easy to spread, but it is only surface area. What is truly difficult in a formalization project is whether, once scale increases, the dependency relationships can still be navigated, the proofs can still be reproduced, and newcomers can still keep building on top. Spread the whole city out into disconnected folders, and no matter how many lines there are, it is just ruins. The signal Axiom is giving is that this thing has not fallen apart.
For people working on AI and writing code, this is also where the news is worth looking at. Large-scale code generation is often questioned as slop, and that question is reasonable. This case cannot prove that all vibe coding is reliable; it only shows that under specific tasks and specific constraints, scale and quality can hold at the same time. Judging output still requires looking at structure, maintainability, and verifiability, not just generation volume.
发布时间: 2026-11-05 03:30