What's this about?
ERA keeps a branching history of notebooks, chooses promising branches with an upper-confidence rule and asks Gemini to propose mutations. Google has published code and application papers, but ERA is not a product.
What can go wrong?
A system can optimise a measurable proxy rather than the scientific question. Scientists must validate that a predictive result is descriptive and not a dataset artefact.
What's the community saying?
- Score design remains the scarce scientific skill.
- Simple baselines such as linear regression expose false sophistication.
- Domain taste and manual practice remain necessary checks on automated search.
Podcast deep summary
TLDR: Platt presents ERA as an optimistic evolutionary search over experiment notebooks. A capability jump between Gemini 2.0 and 2.5 reportedly made the approach useful, leading to work including contrail climate modelling. He repeatedly warns that both humans and models reward-hack, so a high score is not automatically a scientific explanation. The episode connects this lesson to FireSat, fusion, quantum computing and the history of neural networks. Practical advice: define a trustworthy score, start simple, inspect confounders, build domain expertise and sometimes “hike up the mountain” manually. Publisher quotations include “It’s almost like having a hyper-eager grad student who doesn’t sleep,” “Always just fit linear regression,” and “It’s a power tool. It can slice your fingers off.” No timestamped transcript was available, so timestamps are intentionally omitted.