John Platt on Google's ERA: MCTS-style Gemini tree search that auto-solves scoreable science problems

Latent Space · rss · 2026-09-23

Latent Space interviews John Platt — Academy Award winner, inventor of textbook ML algorithms, Erdős–Bacon number 6 — on Google's Empirical Research Assistance (ERA) and AI for Science.

ERA: Born from an "auto-Kaggle" idea, ERA reduces science problems to "scoreable tasks" and automates finding maximizers. Conceptually simple: Gemini maintains a tree of experiment notebooks, MCTS-style — UCB picks the most promising branch (optimistic, not greedy — even a fifth-place notebook gets mutated), proposing 10 mutations per round with history shared across branches. Performance stepped-change between Gemini 2.0 and 2.5, going from non-functional to excellent; it has already yielded at least 10 papers. On avoiding self-deception: "It's a power tool. It can slice your fingers off" — the Kaggle contrail contest was won via a half-pixel label quirk, pure Goodhart. His starter advice: "Always just fit linear regression. Just do it."

Climate: Contrails account for 1% of human-induced warming (1g of exhaust → 10kg of ice crystals acting as an infrared blanket); the fix is dropping flight levels, but counterfactual accounting stumped the team for 2+ years until ERA cracked it. Also FireSat, satellites spotting fires while they're room-sized.

Advice: Deep domain expertise above all; sometimes "hike up the mountain" and implement things yourself; hyper-optimization overfits.

Original post →

John Platt on Google's ERA: MCTS-style Gemini tree search that auto-solves scoreable science problems

More from AGI Musings