Platt: Google's ERA began as an attempt to automate Kaggle, and he says it solved a contrail problem his team was stuck on for two years

Write the science problem as a score, then let an AI search over code

He says the team found that many science problems can be written as 'scorable tasks': a piece of code is wanted that makes some score as high as possible. Statistical models are one kind, but not the only one: Michael Brenner turns the asymptotic expansion of an ordinary differential equation into a test, and another project estimates the carbon dioxide only OCO-2 and OCO-3 can measure from the coarse pixels of the GOES weather satellite. Evolutionary coding has existed since the 1970s and never took off, because random mutation in code is close to useless; ERA works because the thing rewriting the code is a model that knows a lot. He says the team worked on it in 2024 and 2025, that it could not be done on Gemini 2.0, and that it began to work with 2.5.

“random mutation in code space is pretty much worthless”John Platt

A score gets gamed, so choosing it and guarding the holdout are people's work

He separates predictive models (lowest error on some dataset) from descriptive ones (they capture the physics and can extrapolate), says science wants the second, and says ERA hands back the first, so deciding whether a model is descriptive is the scientist's job. Systems find holes all the time, and so do people: in Google's contrail Kaggle competition, the entrants who beat Google had found a half-pixel error in its labels and exploited it. His answer is more rigour: hidden holdout sets nobody looks at, layer on layer of checks. He says ERA is built with a person in the outer loop: it runs for a few hours, the person comes back, points it at a paper or says what is wrong, and lets it run again.

Contrails: ERA unstuck a quantity no one can measure directly

He says that by current estimates contrails cause about 1% of human-made warming; where air traffic is dense, as over Europe, extra contrails cover a few percent of the sky and add about one watt per square metre of forcing locally, against about three watts for all human-made warming averaged over the globe. A plane only has to drop two flight levels to pass under the ice-supersaturated layers, for a little extra fuel. The team watches contrails in satellite images, predicts those layers with convolutional nets, and gives the maps to flight-planning software; those models are not ERA's. The hard part is how much warming a contrail avoided caused, a counterfactual question. The team had a model and a paper for the longwave, heat-trapping half, but was stuck for two years on the reflected sunlight; ERA found a simple model with a combination of confounders the team had not tried, which passed the team's test of recovering contrails injected into synthetic data, where the team's own attempts had failed.

Statistical models suit closed problems with enough data; climate is not one

He says weather and protein structure are data-rich, closed problems where a statistical model will do, AlphaFold being the example. Climate is not: it is non-stationary, the attractor itself is moving, there is no data from thirty years ahead, and a location may have fifty annual data points; the uncertainty in how much carbon dioxide the biosphere absorbs alone is about 300 ppm by 2100, against about 440 to 450 ppm today. So, he says, AI has not changed climate modelling. His hope is a future tool that reads far more papers than any person and writes code constrained by what is known; he stresses that today's ERA does not do this.

“so far, AI has not revolutionized it 'cause it's very, very resistant, again, 'cause of this data problem.”John Platt

The bottleneck is the experiment, and people stay for creativity, rigour and domain knowledge

He says ERA today is all computation and someone has to gather the data; the 'lab in the loop' is still wide open, with many labs that each do one controllable thing and none that does everything. Asked which bottleneck he would remove by magic, he wants a general lab that takes a JSON request and runs any experiment. He does not expect scientists to be replaced: AI is strong at writing code and finding knowledge and still weak at creativity and rigour, and ERA will not discover new physics from scratch, though it follows the constraints it is told. His advice to young people is to go deep in one domain, use the tools widely, and sometimes do the work by hand.

“the everything lab that you could like send JSON blob to, and it will do any experiment at all”John Platt