Lean Creator Leo de Moura Says AI Makes Proofs Cheap. Here Is What the Checkers Show
A.I. / analysis
Lean Creator Leo de Moura Says AI Makes Proofs Cheap. Here Is What the Checkers Show
On a podcast released October 29, the AWS scientist who built Lean claimed AI-written proofs now run to 50 million lines. The claims check out only in part, and he names the two gaps himself.
Leo de Moura, creator of the Lean proof assistant, said on an episode released October 29 that AI has made formal verification cheap enough to change how code is checked. He also said the remaining risk is that people verify the wrong thing.
De Moura is Vice President and Distinguished Scientist at AWS and Chief Architect of the Lean Focused Research Organization. He spoke to Ben Lorica on The Data Exchange, in an episode titled "When AI Writes the Code, Who Proves It Works?" The page was updated on November 1.
The claims in the episode
De Moura said Mathlib, Lean's community maths library, "took 10 years, written by humans, to go from zero to 2.5 million" lines. He said proofs of up to 50 million lines have since been published, and that a proof of Fermat's Last Theorem written by Anthropic's AI runs to 13 million lines. Those figures are his statements on the podcast. The episode page cites no document for either,, and no such document turned up among the pages read here.
Mathlib's own statistics page lists about 140,094 definitions, 292,237 theorems and 772 contributors. It shows no line count and no date, so it neither confirms nor contradicts the 2.5 million figure.
He described two uses already in production. The SymCrypt team at Microsoft uses Aeneas, which verifies Rust code through Lean, for its cryptographic library. Kim Morrison ported the zlib compression library to Lean and proved that compressing then decompressing returns the original data.
He also said a developer he identified as Boris, Claude Code's lead developer, had Claude build Lean models that found real bugs, and merged the resulting pull requests. No repository or pull request was named on the page.
Why the trusted base is small
The argument for verification at this scale is that a reader trusts only the proof kernel, the small program that checks every step. De Moura said Lean has two verified checkers, with a third on the way, and that independent kernels exist in Rust, C#, Haskell, TeX, COBOL and Fortran.
The Lean Kernel Arena tracks those checkers. Its results page, marked "Round in progress", lists 33 entries, including the official checker and a parse-only baseline that only reads input. The test download holds 127 good tests, 80 bad tests and 3.5 MB of data, with groups named cedar, mathlib, std and reachability, plus 26 bugs/ tests and 31 corner-cases/ tests.
The table ranks entries by how many invalid proofs they wrongly accept, then by how many valid proofs they wrongly reject. The page gives no implementation languages, so the six-language list is the guest's claim, not something the arena confirms.
| Item | Figure | Where it comes from |
|---|---|---|
| Kernel Arena entries | 33 | Lean Kernel Arena results page |
| Good tests / bad tests | 127 / 80 | Lean Kernel Arena download |
| Mathlib theorems | 292,237 | Mathlib statistics page |
| Mathlib definitions | 140,094 | Mathlib statistics page |
| Longest published AI proof | 50 million lines | de Moura, on the podcast |
The site has covered what happens when a kernel is the weak point. Lean 4.35 added cross-checking with five proof-checking kernels, including one written in Lean by Claude, after an OpenAI model exploited kernel bugs to get a proof of False accepted.
The two gaps he names
De Moura did not claim verification removes risk. The episode summary lists two. The requirements gap is the distance between what you wrote down and what you want. The model gap is whether the Lean model matches the real code.
He said "testing is great for showing the presence of bugs, but not their absence." He also said "AI dramatically reduces the cost of verifying code" and that the question to ask a team is "show me your specifications, the properties, what you have verified, right?"
His view is that specifications and properties may become the artifacts engineers maintain, with code regenerated. That is a prediction about practice. Nothing on the page measures it.
The Navier-Stokes example
De Moura cited one case where the gap showed up. A human-curated problem statement was used for the Navier-Stokes problem, and OpenAI's informal and formal proofs had mismatches, reportedly over whether a fourth or fifth derivative was assumed. The site covered the dispute over OpenAI's Navier-Stokes data separately.
The example fits the requirements gap. A proof can be correct for a statement that is not the statement anyone meant, and the kernel will accept it.
What is unverified
The 13 million line and 50 million line figures rest on de Moura's word. The Boris anecdote has no linked pull request. The six-kernel language list is not on the arena page.
De Moura also said AWS's AgentCore and a project called Dogwood use Lean, and that Physlib, finance, chemistry and law have early autoformalization efforts. The episode page does not describe any of those in detail.
The arena also shows why a count of kernels is not a count of independent checks. Its ranking counts wrongly accepted invalid proofs first, so an entry that parses input and checks nothing sits in the table beside real checkers. A reader comparing 33 entries has to read the columns, not the row count.
The Data Exchange episode is a conversation, and Lorica's questions are not a test of the answers. The page summary records no one disputing de Moura's numbers and no outside voice who builds or audits proofs at that scale.
The Lean Kernel Arena round was still in progress on November 1, so its pass counts can change before any ranking is final.
Sources
More in A.I.
- 01Meta Prices Muse Ember at $0.10 per Million Input Tokens, Says Nothing on WeightsMeta's smallest model scores 36 on Artificial Analysis, level with Claude Haiku 5.5, but it trailed Haiku on one of the two coding tests the firm ran.
- 02Anthropic Sets $70 to $78 Range for Nasdaq IPO, Crunchbase ReportsAn amended filing reportedly adds a price range and a $26 billion private placement, and shows a nine-month net loss of $104 billion on $33.2 billion in revenue.
- 03Google's Gemini Help Page Says Human Reviewers See Drive and Gmail ExcerptsA GitHub post from a user named thisisme6598 claims Gemini reads files unprompted. Google's own page confirms review, but not that part.
- 04Anthropic Ships Claude Fable 5.2 at $8 per Million Input TokensArtificial Analysis ranks the model first of 74 on its index, but its run needed 275 million output tokens and 44 minutes per task.