Google’s John Platt Describes ERA, an AI System That Searches for Better Scientific Code

Google’s ERA is built around a consequential bet: scientific software can be improved by searching for code that performs better against a chosen measure of success. In a newly published Latent Space interview, Google Research Fellow John Platt described how the system uses Gemini and tree search to pursue that goal.

Platt said Google Research has worked on AI for science for more than a decade, first with machine-learning models designed for individual problems. The ERA work he discussed was conducted in 2024 and 2025, after general-purpose language models prompted the team to look for a more reusable approach.

The first task is defining what “better” means

ERA starts by recasting a research question as a scorable task: code or a proposed solution should maximize an evaluation score. Platt said that translation is often not obvious. An ERA agent converses with a scientist to help define the task, then produces an underlying Python notebook rather than simply returning a one-off answer.

Then ERA searches through competing notebooks

The system’s specialized harness uses Monte Carlo Tree Search, a technique for exploring alternative branches. Platt said ERA keeps hundreds or thousands of candidate notebook versions, selects one, asks Gemini how it could improve, tests a revised version and adds it back to the candidate pool. Google has described ERA as considering thousands of options while optimizing code against a specified goal.

A CO2 project shows the intended use

Platt pointed to a remote-sensing application that combines OCO-2, OCO-3 and GOES satellite data with weather and surface-reflectivity information to estimate atmospheric carbon dioxide at improved spatial and temporal resolution. Google says the related model uses GOES-East data and other inputs to estimate CO2 every 10 minutes across an area; OCO-2 returns to a location every 16 days along a single path.

ERA is already part of Google’s broader Computational Discovery prototype, which Google said is available through a trusted-tester program. A Nature paper on ERA reported work across genomics, public health, satellite imagery, neuroscience, forecasting and mathematics, including 40 single-cell analysis methods that outperformed leading human-developed methods on a public leaderboard.

Platt also drew a line between predictive models that minimize error on data and descriptive models that capture reality well enough to extrapolate, while acknowledging that the distinction can blur. ERA can search for a high-scoring notebook; whether that notebook also offers a useful scientific explanation remains a separate judgment.