AI
Axiom Math formalizes sprawling math library to give AI a verified foundation
The startup has translated the Stacks Project—a monumental reference work in mathematics—into computer-verified code, aiming to prevent AI from propagating faulty reasoning.
TechCrunch — AI
Artificial intelligence systems are churning out claimed mathematical breakthroughs at an accelerating pace, but the rush carries risks: some automated verification systems can themselves be deceived, and the sheer volume of unvetted output has raised concerns about AI slop infiltrating the field.
AxiomMath, a $1.6 billion startup founded by Stanford dropout Carina Hong, is tackling the problem from a different angle. Rather than trying to slow the pace of AI-driven discovery, the company is building a verified mathematical bedrock that AI agents can lean on—and trust.
Formalizing the Stacks Project#
On Wednesday, Axiom announced it had formally verified the Stacks Project, one of history's most comprehensive mathematical references. Started in 2005 as an open-source resource by Columbia mathematician Johan de Jong, Stacks brings together results spanning elementary algebra through cutting-edge work in geometry and number theory. Mathematicians have long treated it as exceptionally authoritative; Columbia's Peter Woit has said nothing rivals it except the 20th-century Bourbaki books that rebuilt mathematics from first principles.
The company's AI agents converted roughly 7,654 pages of Stacks into approximately 16 million lines of formally verified code written in Lean, a proof-checking language. Each theorem is then scrutinized line by line by a proof assistant program to confirm logical soundness—the same technique engineers use to validate chip designs and researchers employ to catch bugs in critical software.
The Challenge of Interconnection#
Formalizing Stacks presented a singular difficulty: the entire library is woven together through dependencies. Each theorem and definition builds on others, creating a fragile web where a single error could propagate across the whole structure. "Things that are big and things that are delicate in life just don't go together," Hong told TechCrunch, emphasizing the need to keep all 16 million lines coherent and free of contradictions.
Axiom solved this by deploying AI agents to work in parallel with loose coordination. Rather than imposing strict hierarchical control, the team allowed agents to identify and tackle tasks independently. Simultaneously, Axiom invested in training its own mathematical reasoning models—a step that Swaminathan noted most other verification startups had skipped—enabling agents to develop intuition for what makes good mathematical work.

