Scaling Past Informal AI – Carina Hong, Axiom Math (www.latent.space)

In a remarkable achievement, the seven-month-old startup Axiom Math scored 12 out of 12 on the notoriously difficult Putnam mathematics exam, outperforming leading AI systems and even top human undergraduates. Axiom’s CEO, Carina Hong, emphasizes that while coding prowess is essential for AI's advancement, it is not sufficient for achieving artificial general intelligence (AGI). She argues that “Verified AI,” which includes formal mathematical proofs, is crucial for scaling and compounding AI capabilities—a perspective illustrated by the historical mathematician Srinivasa Ramanujan's work on formalizing his intuitions. Axiom's innovative approach involves using Lean, a tool for formal verification, enabling their systems to generate proofs that enhance reinforcement learning and training efficiency. While current large language models struggle with such formalization, Axiom has achieved an impressive 99% success rate on the Verina code generation benchmark, significantly surpassing prior efforts from competitors like OpenAI. By building a foundation on verified proofs, Axiom seeks to overcome the informal bottlenecks that hinder AI’s progress and believes that such verification is the path to a feasible AGI future, making it a pivotal point of focus for the AI and machine learning community.