No Priors Ep. 120 | With Google DeepMind’s Pushmeet Kohli and Matej Balog

In this episode of No Priors, Pushmeet Kohli and Matei Balag from DeepMind discuss AlphaEvolve, an AI coding agent that uses Gemini models and evolutionary search to discover new algorithms. They delve into the origins of this project, tracing its lineage from AlphaGo and AlphaTensor, and explain how AlphaEvolve differs by generalizing the algorithm discovery process. The conversation covers the mechanics of how AlphaEvolve evolves code, particularly in optimizing data center scheduling, and addresses the role and limitations of automated evaluators. They also explore the potential for self-improving AI and the implications of AI in accelerating scientific discovery across various fields, emphasizing the collaborative role of AI with human scientists and mathematicians. Finally, they touch on the accessibility of AlphaEvolve and potential future applications within Google's infrastructure.