Axiom Math has raised over $260m from investors including Menlo Ventures to build a self-improving superintelligent reasoner starting with an AI mathematician.
Axiom has formalized the classification of finite simple groups into over 50m lines of Lean, now larger than Linux.
Hong says the second seed round was easier because investors preempted her with term sheets.
Axiom formalizes math's largest proof for verifiable AI
Axiom Math is building an AI mathematician, and its first proof of concept is scale. It formalized the classification of finite simple groups, the largest human-written proof, into Lean, betting that machine-checkable verification is what AI needs before it can be trusted with software and science. Hong argues data and benchmarks decide which labs win, and that talent density defines a new lab.
Takeaways
Verification is the missing layer for AI. Hong says generated code is only trustworthy with a machine-checkable proof, and that this missing layer unlocks verified code generation and lets AI into large critical sectors.
Data and benchmarks define a model. Hong argues proprietary data and benchmarks decide what a model can do, so an AI company needs its own data to last. Axiom builds the full stack in house and outsources only pre-training.
Axiom bet against inference-time search. The team skipped Monte Carlo Tree Search at inference, a bet ByteDance later adopted. Transfer learning produced strong software verification scores and a perfect Putnam score.
Fundraising flipped from brutal to preempted. The seed round was very hard, closed on an idea and a deck Hong calls horrible. The next round was easy because she stopped treating one firm as the hill to die on, and firms preempted with term sheets.
Axiom plans on a yearly horizon. Hong sets goals for the year rather than the quarter, which suits research bets like formalizing the bounded gaps in primes. The bound held at 246 from 2014 until Julia Stalman reached 240.
Swarm is the next frontier. Hong expects the next wave in 2026 or 2027 and says swarm is central to it. Axiom could not have achieved what it has without swarm, and cutting its cost is the next problem.
Formalizing the largest human proof
Richard Brauer called for the hunt to classify all finite simple groups at the ICM in the 1950s.
The finished proof runs past 15,000 pages across many papers and books.
Axiom converted it into more than 50m lines of Lean, announced the day before the talk.
That makes it the largest open source software project, larger than Linux.
Gödel rules out absolute certainty, so verification reduces a proof to a small trusted kernel.
Talent density at a new lab
Talent density is the defining feature of a new lab. Axiom grew from fewer than 15 people in January to about 80 in October, with CTO Shubo, a DeepSpeech co-author, joining during the seed round.
Tap a name to go there · tap a line to see why they are linked
On this episode 4
Show 1
Hong's tactics start with managing her own energy, then naming feelings, which she was bad at before founding and which matters for gratitude and putting people at ease.
Her third is accepting that people work differently, which moved Axiom to a bottoms-up model rather than top-down command.
After half the office worked an all-nighter, she bought a sushi party boat instead of saying thank you.