DeepMind's Pushmeet Kohli on why AlphaFold didn't solve protein folding
Latent Space · rss · 2026-10-10
Latent Space hosts Pushmeet Kohli (Google DeepMind) and Sal Candido (Biohub) on AI-driven biology. Key points:
- Bitter Lesson for data: scaling laws aren't everywhere — the real work is finding regimes where more compute and data reliably help; without data containing the right information, no model gets there.
- Data quality paradox: training protein language models on low-quality metagenomic sequences improves real protein design, but the field should ask what data a problem actually needs rather than scaling what's easy to generate.
- AlphaFold's limits: heavy handcrafted architecture; static structure prediction misses protein dynamics and disorder — folding is far from solved, and a virtual cell needs fundamentally different datasets.
- Interpretability: trustworthiness and uncertainty calibration matter more than full interpretability; frontier models may one day interpret other AI systems better than humans.
- Drug discovery: 10x–100x acceleration is plausible; curing all disease requires 10x breakthroughs, not 10% improvements.
More from Research
- New kissing number lower bounds for dimensions 19-31, with C++ exact verification — felpix_ · 2026-10-10
- Mathematicians push back on 'solved problems = progress' as AI tests the metric — _onionesque · 2026-10-10
- BudgetPix: Google and UIUC's pixel diffusion model adapts compute per image, cutting tokens to 10% — CSProfKGD · 2026-10-10
- Anthropic Launches Conceptual Reasoning Fellows Program for Cog Sci PhDs — AndrewLampinen · 2026-10-10
- COLM Best Paper: Int-Bench Shows LLM Assistants Intervene Too Early and Too Often — maxhkw · 2026-10-10
- Demis Hassabis: Improving medicine is the most important thing AI can do, via new CZI partnership — DeryaTR_ · 2026-10-10