Google's ERA turns science into scores, lands 10+ papers

Platt's career, as his colleague Dave Bacon teases, has been defined by being twenty years early to the next big thing — convolutional neural networks, fusion research, quantum computing. That pattern helps explain why his team arrived at ERA by noticing that many scientific problems reduce to a 'scoreable task': once a score function exists, finding code that maximizes it can still be enormous effort. Platt's team set out to automate that search, starting from the idea of an 'auto-Kaggle' AI, with Kaggle's Google ownership making data easily available.

The result sits between an evolutionary algorithm and a large language model. Gemini keeps a running tree of past experiments and uses an Upper Confidence Bound rule to pick which notebooks to mutate, making the search optimistic rather than greedy — sometimes even the fifth-best notebook gets chosen. Platt notes the approach works because Gemini actually knows where to look, and that a step change between Gemini 2.0 and 2.5 took it from not working to working great. The team's results span climate work — including contrail mitigation, where the counterfactual problem had stumped them for over two years — and FireSat, a satellite constellation for spotting fires while they are still room-sized.

Platt is careful about what this proves. ERA supplies predictive models; it is up to the scientist to make sure they are truly descriptive, and he compares the tool to a power tool that can slice your fingers off. His own example of how easily metrics get gamed is the contrail-detection Kaggle competition, won partly by entrants who noticed a half-pixel labeling error. Whether ERA meaningfully speeds discovery, or mainly rewards careful problem framing, may depend less on the model than on the scientists deciding what counts as a real solution.