

Meet AlphaEvolve: The Autonomous Agent That Discovers Algorithms Better Than Humans With Google DeepMind’s Pushmeet Kohli and Matej Balog
371 snips Jun 26, 2025
In this conversation, Pushmeet Kohli, VP at Google DeepMind, and researcher Matej Balog dive into AlphaEvolve, a groundbreaking autonomous coding agent. They discuss how this AI discovers new algorithms more efficiently than humans, tackling optimization of matrix multiplication and illustrating its revolutionary potential in solving complex problems. Kohli and Balog also explore the future of self-improving AI, the interplay between human scientists and AI, and the exciting implications for various fields beyond just computer science.
AI Snips
Chapters
Transcript
Episode notes
AlphaEvolve's Scientific Creativity
- AlphaEvolve is an AI agent that discovers new algorithms, pushing beyond human ability in scientific problems.
- It marks a shift from pattern regurgitation to technical creativity in AI, akin to AlphaGo's Move 37.
From AlphaTensor to AlphaEvolve
- AlphaTensor was DeepMind's first breakthrough improving matrix multiplication algorithms beyond human discovery.
- AlphaEvolve generalizes this approach, seeking broader algorithmic discovery across problems.
Why Matrix Multiplication Was Hard
- Matrix multiplication solutions remained undiscovered due to vast, non-intuitive search spaces.
- Experts were not complacent but couldn't find these intricate improvements without AI's expansive search.