The $64M Bet on an AI That Has to Be Right | Carina Hong, CEO of Axiom - Gradient Dissent: Conversations on AI Recap

Podcast: Gradient Dissent: Conversations on AI

Published: 2026-02-05

Duration: 51 min

Summary

Carina Hong discusses Axiom's innovative approach to building a self-improving AI mathematician that combines generation and verification, resulting in impressive performance on competitive math exams. She shares insights from her transition from academia to entrepreneurship and the influence of rock and roll on her management style.

What Happened

In this episode, host Lucas B. Wald interviews Carina Hong, the CEO and founder of Axiom Math, a company dedicated to developing a reasoning engine that enhances mathematical reasoning through a unique combination of generation and verification. Axiom has garnered attention for its exceptional performance, including a notable score on the challenging Putnam exam, where their system achieved eight out of twelve correct answers, illustrating the capabilities of deterministic and probabilistic systems working in tandem. Carina explains that the aim is to create an AI mathematician that not only proves theorems but also generates conjectures, making it a versatile tool in the field of mathematics.

Carina reflects on her journey from being a star academic mathematician to leading a startup. She likens the excitement of running Axiom to the rebellious spirit of rock and roll, emphasizing the thrill of embarking on a new adventure. She believes that the future of mathematics will involve mathematicians adapting to new abstractions and leveraging AI to prove their intuitions. The conversation delves into the intricacies of Axiom's methodology, including the use of formal languages to improve sample efficiency and the importance of a knowledge base that supports both conjecturers and provers in the reasoning process.

Key Insights

Key Questions Answered

What is Axiom Math and how does it work?

Axiom's mission is to build a reasoning engine that is self-improving and combines generation and verification. They focus on creating an AI mathematician as a testing ground for this self-feedback loop, utilizing formal languages like Lean to improve the efficiency of problem-solving. The system includes a prover for theorem proving, a conjecturer for generating new ideas, and a knowledge base to support these functions, enabling a dynamic interaction between all components.

What notable achievements has Axiom Prover accomplished?

Axiom Prover has achieved remarkable results on competitive math exams, most notably the Putnam exam, where they scored eight out of twelve correct answers. This is particularly impressive given the exam's difficulty, with the median score often being zero for participants. Carina mentioned that their system's performance has been boosted by the collaboration of deterministic tooling and probabilistic systems, showcasing the potential of AI in math.

How does Carina Hong's background influence her leadership at Axiom?

Carina's journey from a star academic mathematician to a startup founder has profoundly shaped her leadership approach. She draws parallels between the thrill of managing a startup and the rebellious nature of rock and roll music, emphasizing a need for creativity and a non-traditional mindset in her management style. This influences how she fosters innovation and maintains a dynamic culture at Axiom.

What does Carina predict for the future of mathematics?

Carina predicts that mathematicians will need to adapt to new abstractions and work in tandem with AI tools. She envisions a scenario where human intuition guides the exploration of mathematical concepts, while AI systems serve as diligent assistants, proving theorems and exploring new conjectures. This collaboration could significantly alter the landscape of mathematical research and problem-solving.

How does Axiom's approach differ from traditional AI systems?

Axiom distinguishes itself by focusing specifically on the combination of generation and verification within its reasoning engine, which is often overlooked in current AI systems. By starting with mathematics, Axiom aims to create a more robust AI that can learn and improve over time through a self-feedback loop, leveraging formal languages and knowledge bases to enhance its capabilities beyond simple problem-solving.