The Cake, Not the Cherry: Inside OpenAI's Reinforcement Learning Breakthrough in Mathematical Reasoning

OpenAI's breakthrough in solving the Erdős conjecture through reinforcement learning marks a fundamental paradigm shift: test-time computation and extended exploration, not just model scaling, unlock genuine reasoning capabilities. This conversation with Dan Roberts reveals how RL enables AI systems to tackle problems once thought unsolvable—challenging conventional wisdom about where AI's true power comes from.