John Platt describes Google’s Empirical Research Assistance tool for scientific scoring
John Platt, a machine‑learning pioneer with an Oscar and two eponymous algorithms, explains Google’s Empirical Research Assistance (ERA). ERA treats any scientific problem that can be expressed as a scoreable task and uses a Gemini LLM to run a Monte‑Carlo‑Tree‑Search‑style loop: the Upper Confidence Bound rule selects promising notebook branches, Gemini proposes about ten mutations per branch,…
Key points
- ERA combines Gemini LLM with Monte‑Carlo Tree Search to mutate notebook code and maximize defined score functions.
- Using ERA, Platt’s team built a contrail climate model after two years of counterfactual work.
- Contrails contribute about 1 % of anthropogenic warming; a gram of exhaust can generate ten kilograms of ice crystals.
The team has applied ERA to climate‑related challenges, most notably modeling contrail formation. Contrails account for roughly 1 % of human‑induced global warming, and a single gram of aircraft exhaust can seed ten kilograms of ice crystals. After two years of counterfactual modeling, ERA discovered a simple model that captured previously missed confounders. Platt also mentions a $15,000 contrail‑detection Kaggle competition where winners exploited a half‑pixel label error, illustrating how metric gaming can mislead scientific progress.
🔬 An Oscar, Two Asteroids, and the Algorithm in Your sklearn: John Platt on AI for Science
Latent Space · 22 September 2026
Loading the full article…
This text was published by Latent Space and written by Brandon Anderson. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Did OpenAI solve the wrong Navier-Stokes problem? · 1 src
- Anthropic reports Claude leads 26% of its AI research and development tasks · 5 src
- Generative AI designs de novo thiolation domains that boost nrps yields up to ~3‑fold · 1 src
- AISI and EvalEval release evaluation cards for five benchmarks and six frontier models · 1 src
- OpenAI's GPT-6 Astra attempts 97 unsafe instructions in robot safety test · 2 src
Comments
via GitHub Discussions