Google’s ERA Uses LLM-Guided Search to Automate Science, John Platt Says

In a Latent Space podcast episode, Google researcher John Platt discussed Empirical Research Assistance (ERA), a system his team built to automate science problems that can be written down as a score function. ERA keeps a running tree of past experiment “notebooks” and uses a language model to propose around ten mutations at a time; an Upper Confidence Bound rule decides which branches are most promising, occasionally picking an optimistic option rather than the current best. Platt likened the result to a hyper-eager graduate student who never sleeps, but noted that the approach only began working well with the step change from Gemini 2.0 to 2.5.

According to Platt, ERA has produced at least ten papers, including work on climate change. One notable result involved condensation trails from airplanes, which account for about 1% of human-induced global warming. The practical fix is to have planes drop a flight level or two in ice-supersaturated air, but the hard part was estimating how much warming was actually prevented. Platt said this counterfactual problem stumped his team for over two years until ERA found a simple model with confounders they had not considered.

Platt also cautioned that powerful optimization can fool scientists. He invoked Goodhart’s law and cited Google’s contrail-detection Kaggle competition, where winners exploited a half-pixel label error in the training data—good for a prize, but not for solving contrails. His advice, repeated emphatically, was to fit a linear regression first, or an SVM, before reaching for more elaborate machinery.