AlphaFold Proves Calibration Beats Interpretability
DeepMind's Pushmeet Kohli explains why exact uncertainty calibration like AlphaFold's pLDDT matters far more than human interpretability.
The Bitter Lesson of Biology: Why Scaling AI Isn't Enough — Pushmeet K, DeepMind & Sal C, Biohub
In this panel discussion moderated by Brandon Anderson, Google DeepMind's Pushmeet Kohli and CZ Biohub's Sal Candido discuss whether the 'Bitter Lesson' of AI scaling applies directly to biological data and modeling. They explore the real limitations of AlphaFold, the nuance between artisanal architecture inductive biases and pure compute scaling, the true nature of model interpretability versus uncertainty calibration, and why solving translational medicine requires looking for 10x paradigm shifts rather than incremental 10% gains.
Key takeaways
- DeepMind's Pushmeet Kohli argues AlphaFold 2 cannot be interpretable to people because biological mechanisms exceed the cognitive capacity of the human brain. Read more →
- Sal Candido warns that scaling laws do not appear automatically in biology; discovering the exact regimes where data volume predictably drives performance is an open empirical problem. Read more →
- DeepMind designed AlphaFold 2 by hardcoding biophysical rules into the network rather than relying on pure compute scaling. Read more →
- Machine learning models already operate across every stage of target validation and screening, yet clinical pipelines remain stalled without mechanistic biology models. Read more →
4 articles from this episode
Scaling Laws in Biology Require New Assays, Not Bigger Models
DeepMind's Pushmeet Kohli and CZ Biohub's Sal Candido explain why biological scaling laws break down without new assays and better data.
AlphaFold and the Limits of Inductive Bias in AI Biology
Pushmeet Kohli and Sal Candido debate whether hardcoded scientific rules or raw scale will win the future of biological AI models.
Why 10x Leaps Beat 10% Gains in Biotech
Google DeepMind's Pushmeet Kohli and CZ Biohub's Sal Candido explain why drug discovery requires first-principles leaps instead of small gains.
More Latent Space episodes
- AI Scientists Are Here: Autonomous Labs & Synthesis Superintelligence — Periodic LabsOct 11 · 5 articles
- The Future of Claude Code: Mods, Mutable Software, & Multiplayer Agents — Thariq Shihipar, AnthropicOct 4 · 4 articles
- Which GPU Clouds Are Actually Good? | ClusterMAX 3.0Oct 4 · 6 articles
- Recursive Language Models — Alex Zhang, MIT PhDOct 4 · 6 articles