Axiom (@axiommathai) header
Axiom (@axiommathai) profile picture

The Starting Point for Reasoning

Joined September 2025
Photo by Axiom (@axiommathai)Photo by Axiom (@axiommathai)Photo by Axiom (@axiommathai)Photo by Axiom (@axiommathai)Photo by Axiom (@axiommathai)Photo by Axiom (@axiommathai)Photo by Axiom (@axiommathai)Photo by Axiom (@axiommathai)Photo by Axiom (@axiommathai)Photo by Axiom (@axiommathai)
The largest math proof in human history is now computer-verified. Today, Axiom announces a complete Lean auto-formalization of the Classification of Finite Simple Groups, the "periodic table of symmetry." About 15,000 pages, built by over 100 mathematicians across generations.
425
1,302
656
7,164
3,754,786
Three research themes being productized that are worth watching: agents that use formal proofs to verify code (shoutout @axiommathai), agents that improve their own workflows from execution feedback, and GPU memory systems combining high-bandwidth memory with flash for more efficient long-context AI. Stoked to see what all of the researchers at @NeurIPSConf are up to!
1
1
4
347
In 1966, Mark Kac asked: "Can one hear the shape of a drum?" In graph theory, the question is whether a network's "frequencies" (the eigenvalues of its adjacency matrix) uniquely determine its shape. Almost always, the answer is no. Except for a rare gem. 🧵
8
9
80
5,322
The published proof relies on a general theorem about distance-regular graphs. When Axiom Prover formalized this result, it found a more direct shortcut, reaching straight into our finite group theory library to solve a problem in spectral graph theory.
1
1
8
788
From the eigenvalues alone, the Lean proof extracts the 126 vertices, degree 3, and girth 12, recognizing the incidence graph of a generalized hexagon of order 2. From there, it applies the 1985 Cohen-Tits geometric uniqueness theorem in our CFSG corpus.
1
1
10
638
The longest proof ever written by humans, about 15,000 pages from 500+ papers, is now fully verified by machine. Our AI swarm formalized the Classification of Finite Simple Groups and the 7,500-page Stacks Project, catching errors in the original proof along the way.
2
7
2
120
5,723
Thank you Alex!!!
Just stumbled onto this. A rare prognostication of mine that turned out to be right... :) (Wasn't that hard to call... I would *not* have guessed then that we'd also today have Deligne and Faltings!..)
1
11
1,926
Stacks Projectに続き、今度は4次元のMMPがLeanで形式化されたとのこと。やっぱり代数幾何は狙われているな。 基盤理論が形式化されれば、研究の効率が飛躍的に向上する可能性がある。 それが新しい定理やアイデアを発見する能力の向上にどこまでつながるかはわからないが。
Replying to @axiommathai
Today, Axiom Math is releasing a complete Lean formalization for one of the deepest results from that release: the Minimal Model Program in dimension up to 4. The solution spans 2 million lines of machine-checked code in a new reusable library.
2
13
1
65
4,810
On the night of May 29, 1832, 20-year-old Evariste Galois stayed up writing a frantic last letter to his friend Auguste Chevalier before walking into a fatal duel at dawn. In it, he made a bold claim about polynomial equations, with no proof. 🧵
3
4
113
13,758
Galois wrote that a key family of equations of degree p+1 can be simplified to degree p in only three special cases: p = 5, 7, and 11. In modern group theory: PSL₂(𝔽_p) has a proper subgroup of index p only for p = 5, 7, 11. He was fatally wounded the next day.
2
33
1,910
Nearly two centuries later, Axiom Prover has formalized Galois's claim (and extended it to every finite field) in Lean in under a day. Because Dickson's 1901 classification of PSL₂(q) subgroups was already in our CFSG corpus, 92% of the proof was ready to reuse.
1
2
37
1,685
Classification has been a century-old obsession for mathematicians. Here's a new milestone: termination of the 4-dimensional Minimal Model Program, released by @OpenAI, formally verified by @axiommathai. axiommath.ai/research/the-mi…
On Oct. 6, @OpenAI announced the resolution of hundreds of open problems in math, but many of their proofs were left unverified. When AI operates beyond the capacity of human peer review, how do we know what is actually true? The answer lies in formal verification.
3
3
23
2,650
At Axiom, we believe that formal verification is the way to deal with the imminent explosion in AI-driven scientific research. We'll soon have systems that not only use verification to confirm AI output, but that leverage verification as part of the development process itself.
On Oct. 6, @OpenAI announced the resolution of hundreds of open problems in math, but many of their proofs were left unverified. When AI operates beyond the capacity of human peer review, how do we know what is actually true? The answer lies in formal verification.
1
18
841
Replying to @axiommathai @OpenAI
When formalization capability gets above a threshold, 2 million lines of Lean code just emerge effortlessly out of Stacks, our reusable library. Open AI’s proof dump lacks many formalizations in algebraic geometry as they lack the formalization of this foundation. MMP is correct!
2
1
11
835
“Axiom is releasing a complete Lean formalization for one of the deepest results from that release: the Minimal Model Program in dimension up to 4. The solution spans 2 million lines of machine-checked code in a new reusable library.” We are now seeing 2M LOC as a corollary 💪
On Oct. 6, @OpenAI announced the resolution of hundreds of open problems in math, but many of their proofs were left unverified. When AI operates beyond the capacity of human peer review, how do we know what is actually true? The answer lies in formal verification.
5
35
1,684
Yes, the OpenAI result on the minimal model program for four-folds is indeed correct, verification courtesy of AxiomProver running on top of our Stacks library 😊
On Oct. 6, @OpenAI announced the resolution of hundreds of open problems in math, but many of their proofs were left unverified. When AI operates beyond the capacity of human peer review, how do we know what is actually true? The answer lies in formal verification.
1
3
21
1,873
On Oct. 6, @OpenAI announced the resolution of hundreds of open problems in math, but many of their proofs were left unverified. When AI operates beyond the capacity of human peer review, how do we know what is actually true? The answer lies in formal verification.
11
11
7
100
14,665
How did our agents build such a large codebase in a matter of days? They had the perfect starting point: our formalization of the Stacks Project, which provided the core APIs they needed to formally verify a theorem at the frontier of the field. Verified foundations compound.
1
12
1,002
As AI scales past the limits of human review, informal reasoning becomes a bottleneck. Formal verification turns AI output into mathematical certainty and reusable infrastructure. Github: github.com/AxiomMath/MMP Blog: axiommath.ai/research/the-mi…
1
15
957