DeepMind's Pushmeet Kohli: Why AlphaFold Didn't Actually Solve Protein Folding
Latent Space · youtube · 2026-10-10
In a Latent Space panel, Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido argue AlphaFold was only the beginning: scaling compute and data alone won't solve biology. Key points:
- The Bitter Lesson has limits in biology — finding the right scaling law in biological data matters more than blindly growing models; researchers too often optimize for available data rather than the most important scientific problems.
- AlphaFold relied heavily on handcrafted architecture and scientific intuition, and static structure prediction misses protein dynamics and disorder.
- Protein language models may already contain unlocked scientific knowledge; low-quality metagenomic data can still improve them; trustworthiness and uncertainty calibration matter more than full interpretability.
- Moving from individual proteins to virtual cells requires fundamentally different datasets, like cryo-EM micrographs.
- Curing all disease demands 10x breakthroughs, not 10% gains; AI could deliver 10–100x acceleration in drug discovery, and future frontier models may interpret other AI systems better than humans.
More from AGI Musings
- Decomposed fertility data shows mothers still average 2.1+ kids; the crisis is childlessness — Pascallisch · 2026-10-10
- Should AI safety outsource comms and translation to outsiders? — DavidSKrueger · 2026-10-10
- Google Fellow Milanfar argues AI still needs engineers in keynote speech — docmilanfar · 2026-10-10
- Bengio boosts AI slowdown call: Claude now leads 26% of Anthropic R&D, RSI red line 'has become the plan' — Yoshua_Bengio · 2026-10-10
- AI data labeling gold rush: Mercor founders, 22, become youngest self-made billionaires as data goes stale at 70% accuracy — FinanceYF5 · 2026-10-10
- "Made by humans" could become the new "Made in America" badge — petesena · 2026-10-10