🔬 An Oscar, Two Asteroids, and the Algorithm in Your sklearn: John Platt on AI for Science
Google's John Platt explains ERA, which uses Gemini to search and automate scoreable scientific experiments.
In a Latent Space interview, Google researcher John Platt describes Empirical Research Assistance (ERA), which treats scientific problems as scoreable tasks and searches a tree of experiment notebooks with Gemini. ERA uses an upper-confidence-bound rule similar to Monte Carlo tree search and proposes about ten mutations per step. Platt says performance jumped from Gemini 2.0 to 2.5 and that ERA contributed to at least ten papers, including airplane-contrail climate modeling that the interview links to about 1% of human-induced warming. He also cautions that optimizing a score can overfit, invoking Goodhart's law.
- ERA searches notebooks with Gemini and an upper-confidence-bound rule.
- About ten mutations are proposed at each iteration.
- Results improved sharply from Gemini 2.0 to Gemini 2.5.
- The team credits ERA with at least ten papers, including contrail work.
- Platt warns score targets can be gamed, citing Goodhart's law.
- Vendors
- Organizations
- AI models
- Gemini 2.0Gemini 2.5
From the original article
8 min read at latent.spaceWe talked to Google’s Oscar winning “Giganerd” about automating science, solving climate change, and how future generations can contribute to science in the age of superintelligent AI
By Brandon Anderson · latent.space
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.