Google's ERA writes expert-level scientific code

Google has a Nature paper out on ERA, short for Empirical Research Assistance, a system that uses Gemini to write the kind of code scientists spend weeks tuning by hand. The Google research team describes it as searching the literature, writing code, trying combinations of techniques, then running a tree search over thousands of candidate programs and keeping whatever scores best against the stated goal.

What makes the paper worth reading is that the results are external, not internal benchmarks. ERA's epidemiological forecasts rank at or near the top of the CDC's public leaderboards for flu, COVID-19, and RSV hospital admissions up to four weeks out. For California water planning, it produced spring runoff predictions more accurate than the state's official Bulletin 120 outlook. It also mapped CO2 concentration at a finer resolution than usual, picking up hourly cycles and clear urban hotspots, and designed a solar collector layout with no backward shading. Eight peer-reviewed manuscripts now use it on specific problems, and Google is rolling the tool into Gemini for Science under the name Computational Discovery.

The claim is not that the model invents new science. It is that the slow, iterative part, writing and refining empirical software, can be largely automated and still match expert work. You can read Google's writeup here.

Why it matters

If your research depends on hand-built modeling code, a tool that matches expert results on public forecasting leaderboards is worth a look, though the honest test is whether it holds up on your own problem rather than the eight cases Google chose to publish.