DeepMind's Pushmeet Kohli on AI's Scientific Revolution

DeepMind’s Pushmeet Kohli explains how AlphaEvolve uses LLMs + evolutionary search to discover new algorithms and prove math results that stumped researchers for decades.

• AlphaEvolve improved a 50-year-old matrix multiplication record: 4x4 matrices now require 48 multiplications, down from Strassen’s 49.
• Unlike predecessor FunSearch, AlphaEvolve searches entire algorithms (large code bodies), not just small function completions, and needs far fewer evaluations.
• AlphaEvolve generated interpretable code for Google data center job scheduling that outperformed expert-designed heuristics — engineers can read and debug it.
• Co-Scientist uses multiple Gemini agents playing distinct roles (hypothesis, critique, ranking, editing) in shared memory; performance improves over days of compute, not hours.
• AlphaFold 2 cut protein structure prediction from 1–5 years and ~$1M per protein to seconds, but clinical drug validation remains the unresolved bottleneck.
• Kohli argues calibrated uncertainty — knowing when a prediction is unreliable — is what made AlphaFold trusted in practice, and current LLMs lack this.
• Kohli believes AI-accelerated scientific discovery is already underway, with validation (digital-to-real-world gap) and accessibility as the two main remaining bottlenecks.

2025-07-15 · Watch on YouTube