Google Fellow John Platt: ERA AI scientist grew out of an attempt to automate Kaggle
Latent Space · youtube · 2026-09-23
Latent Space interviews Google Fellow John Platt (Platt scaling, SMO, an Academy Award) on ERA (Empirical Research Assistance), Google's AI research system.
Key points:
- ERA began as an "auto-Kaggle" project, combining LLMs with tree search and UCB to guide experiment iteration; the Gemini 2.0→2.5 leap made it work
- ERA helped crack a climate modeling problem that had stalled for years (measuring contrail warming) and supports FireSat wildfire detection
- The hard part is turning science into scoreable tasks — choosing the score matters more than the model
- Reward hacking and Goodhart's law: predictive accuracy isn't understanding, illustrated by a half-pixel labeling error that won a Kaggle competition
- Also: learning from Feynman, fusion and quantum outlooks, why he still recommends linear regression first, and the enduring value of scientific taste
More from AGI Musings
- Yarin Gal: LLM-written experiments that fail to replicate are no different from any others — yaringal · 2026-09-23
- Martin Casado: math and physics might just be apps of computer science — burny_tech · 2026-09-23
- Two personal AIs tried to schedule coffee: agent interop needs a protocol — signulll · 2026-09-23
- Researcher pushes back on "stochastic parrots": LLMs aren't autocomplete — aran_nayebi · 2026-09-23
- AI safety hypocrisy: call for a pause, then ship bigger models anyway — xiaohu · 2026-09-23
- AI costs fall 47% per quarter, 4x faster than DNA sequencing: Epoch AI — daveholtz · 2026-09-23