She is teaching machines to prove their own math - then check their own work. The thesis fits in three words.
Most AI labs are trying to make the machine sound right. Carina Hong is building one that has to be right - and show its work in a language that cannot be fooled. Axiom Math, the company she founded in March 2025, is building what she calls an AI mathematician: a system that generates proofs, verifies them in formal proof languages like Lean, and improves itself by checking its own reasoning.
The pitch is deceptively simple. "Math is the perfect sandbox for building superintelligence," she says. Code can hallucinate. Chatbots can bluff. A formal proof either holds or it doesn't. Pick the one domain where a machine cannot talk its way past a wrong answer, and you have a place to grow real reasoning. That is the bet, and investors put a reported $1.6 billion valuation behind it inside of roughly nine months.
She did not arrive here by waiting her turn. Hong finished dual MIT degrees in mathematics and physics in three years, with more than twenty graduate-level courses and nine peer-reviewed papers along the way. She won the Morgan Prize - the top research honor for an undergraduate mathematician in North America - and the Schafer Prize for undergraduate women in math. Then a Rhodes Scholarship to Oxford. Then a joint law and mathematics PhD track at Stanford. She left all of it.
The story she tells is not a boardroom story. The choice to leave Stanford and start Axiom landed during a morning run. Her first hire, former Meta AI researcher Shubho Sengupta, she met by chance at a Stanford coffee shop over weekend matcha lattes. The company began on a plastic folding table and a friend's spare couch. From that, she recruited researchers out of Meta FAIR, Meta's GenAI team, and Google Brain/DeepMind - people who had spent careers at the frontier and chose to follow a 24-year-old into proofs.
Hong points to advice from AMD chief executive Lisa Su - run toward the hardest problems - as part of why she jumped. She took it about as literally as a person can. The hardest problem she could find was not a product. It was whether a machine can do mathematics that a mathematician would respect.
Hong frames this as a convergence moment. Three things arrived at once: large language models that can reason at length, formal proof languages like Lean that turn mathematics into something a computer can check line by line, and a leap in code generation. Put them together and you can build a system that proposes, proves, and verifies - a loop that improves without a human grading every step.
"The part that is actually quite beautiful," she says, "is that it's not about erasing hallucination or catching mistakes." It is about building rigor in before the errors happen. The applications she names go well past math class: hardware and software verification, quantitative finance, cryptography, anywhere a wrong answer is expensive and a provably correct one is worth paying for.
She grew up in Guangzhou and went to South China Normal University Affiliated High School, where she was one of only four girls on the provincial math Olympiad team. A first-generation college student, she taught herself English well enough to read advanced math texts - and then read them.
By the time she reached MIT, the pace was set. Twenty-plus graduate courses. Nine papers across number theory, combinatorics, probability, and theoretical computer science. Distinction on both Oxford dissertations. She has served as a referee for more than ten academic journals and spoken at TEDx and international conferences. The credentials are not the point, though. The pattern is: find the hardest available thing, then do it faster than expected.
She has described research mathematics as "pain and suffering" - and, in the same breath, the most compelling thing she has ever done. That tension is the whole personality in one line.
Propose a proof. Verify it in a formal language like Lean. Let the system learn from its own checked reasoning. Rigor goes in before the error, not after. A loop that gets better without a human grading every line.
An emerging math-AI category that includes Harmonic (co-founded by Robinhood's Vlad Tenev) and Logical Intelligence. Axiom's early credibility moment: taking on the Putnam, one of the hardest college math exams.
The career-defining decision to leave Stanford happened mid-stride, on a morning run.
Her first hire came from a chance encounter over matcha lattes at a campus coffee shop.
Axiom started on a plastic folding table and a friend's spare couch.
She held a JD and a math PhD track at Stanford at the same time - then walked away from both.
Her recurring word for math is "sandbox" - the one place an AI can't bluff its way to an answer.
Sources: Rhodes Trust · MIT Mathematics · VnExpress · Upstarts Media · TWIML · B Capital · IDNFinancials · YouTube.
Figures including valuation and funding are as publicly reported and may change.