TL;DR
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
A guest post by mathematician Thomas Hales, published by Terence Tao on Oct. 9, 2026, discusses Lean’s reliability and the rapid growth of AI-generated formal mathematics. The post points to large recent projects, but its supplied text does not establish that AI-generated proofs are error-free or that every announced result has been independently validated.
Mathematician Thomas Hales has published a report on the Lean theorem prover, arguing that its role in checking formal proofs deserves attention as AI-generated formalizations grow in scale. The guest post, published by Terence Tao on Oct. 9, 2026, lists recent projects involving major mathematical results and textbooks, while raising a central question: what does a proof checked by Lean establish, and what can still go wrong?
Hales describes a formal proof as one checked by computer against mathematical foundations and rules of logic. Lean is a proof assistant: it checks whether a supplied proof follows within its formal system. The system was introduced by Leo de Moura in 2013 while he was at Microsoft, and the post says Microsoft made the software open-source.
A key part of Lean’s ecosystem is mathlib, a shared library of formalized mathematics begun in 2017 by Mario Carneiro and Johannes Hölzl, using existing parts of Lean’s core library. Hales reports that mathlib contains nearly 300,000 theorems, more than 100,000 definitions and 2.5 million lines of code, with over 700 contributors. Those library results can be reused in later formal proofs.
The post identifies autoformalization—using AI to translate mathematical writing into formal code—as a practical development in 2026. It cites announcements including a formalization of the 24-dimensional sphere-packing result, Meta/Facebook Research’s ATLAS work on 26 textbooks, and projects concerning Fermat’s Last Theorem and a forced Navier-Stokes blowup result. These are examples reported by Hales, not independent evaluations supplied with the post. He also notes that the 24-dimensional project produced about 500,000 source lines of code before later pruning reduced it to about 200,000.
What Lean’s Proof Check Can Establish
Lean matters to mathematicians because it can check a proof’s formal steps against a specified logical framework, rather than relying only on a human reader’s judgment of a written argument. That can expose gaps in a formal derivation and make the exact dependencies of a result inspectable. Hales links this function to the reliability of mathematics in science and society.
AI changes the scale of the task, but not the meaning of the check. A language model can propose formal code; Lean’s kernel checks whether that code proves the stated formal proposition. A successful check is strong evidence that the encoded proof follows within Lean’s foundations, assuming the system and its trusted components behave as intended. It does not by itself show that the formal statement captures the intended informal theorem, that the translation preserved every assumption, or that the result has the broad interpretation readers may attach to it.
The distinction is practical as well as philosophical. Millions of generated lines may make formalization faster, yet reviewers still need to understand what was formalized, which definitions and assumptions were used, and how the code was checked. The projects Hales describes point to growing capacity; they do not alone settle how reliably AI handles mathematical meaning.
As an affiliate, we earn on qualifying purchases.
From Manual Proofs to AI Formalization
Formalization has long required substantial human effort. Hales gives the Kepler conjecture formalization as an example: he says it took about 20 human work-years and produced roughly 500,000 lines of proof scripts. Earlier formalized results cited in the post include the four-color theorem, the Feit–Thompson theorem, and sphere-packing results in several dimensions.
The recent milestones Hales lists suggest a shift from people translating proofs line by line toward AI systems generating large portions of formal code. The post describes a September 2025 prime number theorem project as “quasi-autoformalization,” because people still had to guide the system when it stalled. Later examples range from textbook material to individual headline results. Hales also notes that these projects used different proof assistants and language models; his report focuses on Lean.
The source is a guest post, not a technical audit of the named projects. Tao’s accompanying note says the post was converted from another file format using AI. That disclosure concerns the preparation of the blog post and should not be confused with validation of the proof projects discussed within it.
““For me, what matters is the consistency of math and its unparalleled reliability in support of science and civilization.””
— Thomas Hales, guest post on Terence Tao’s blog
As an affiliate, we earn on qualifying purchases.
Limits of the Reported AI Proofs
The supplied post discusses reliability but its provided text ends as Hales begins a section on Lean’s type-theoretic foundations. It does not provide a full technical account of the system’s trusted computing base, possible implementation risks, or independent audits of the AI-generated projects. It is also not clear from the source whether every announced formalization has been checked by outside researchers or how each project handled the match between informal statements and formal definitions.
Large code counts do not measure correctness, coverage or mathematical usefulness on their own. Nor does a successful Lean check establish that an informal paper has been translated without omitted assumptions. The announcements cited by Hales should therefore be understood as reported project milestones, with the precise scope and validation status dependent on each project’s documentation.
AI formalization tools for mathematics
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What Researchers Will Need to Verify
The next useful evidence will come from project-level documentation: the exact formal theorem statements, dependencies, checking procedures, and independent review where available. Researchers will also need to distinguish a proof assistant’s verification of code from an AI system’s ability to interpret a paper accurately. Hales’s post puts those issues on the table but does not resolve them in the supplied material.
Lean and mathlib development continues through contributions to the open-source ecosystem, while the autoformalization efforts cited by Hales point toward more large-scale projects. Whether those systems make routine formalization easier, as Urban predicted, will depend not only on generating code but on making the results understandable, reproducible and appropriately checked.
proof assistant for mathematicians
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Lean?
Lean is a proof assistant that checks formal mathematical proofs against a specified logical framework. Hales’s post focuses on its use by mathematicians and on the shared mathlib library.
Does a Lean check prove that an AI understood a paper correctly?
No. A successful check verifies that the formal code proves the proposition encoded in Lean. It does not alone show that the proposition accurately represents the intended informal result or that no assumptions were lost in translation.
What is autoformalization?
Autoformalization is the use of AI to translate mathematical material, such as a paper or textbook, into formal code for a proof assistant. Hales says it became a practical reality in 2026 and cites several announced projects.
How large is mathlib?
Hales reports that mathlib has nearly 300,000 theorems, over 100,000 definitions and about 2.5 million lines of code, contributed to by more than 700 people. These figures are from the guest post.
Have the AI-generated formalizations been independently validated?
The supplied report does not establish the independent validation status of every project it names. Readers should consult each project’s documentation for its formal statement, proof dependencies and review process.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
