ERA: an AI that writes scientific software
On 19 May 2026 Nature published 'An AI system to help scientists write expert-level empirical software' (Nature 654, 909–916): ERA pairs a large language model with tree search to write programs that maximise a quantitative quality metric, from single-cell data analysis to forecasting COVID-19 hospitalisations.
Why it matters
For scientific software the result is measured on public leaderboards and in a retrospective comparison with the CDC, not only in the developer's description, and the article passed peer review. An editorial assessment: every run was done by the authors and only the reference code is public.
Numbers from the article. In single-cell batch integration 40 of 87 generated methods outperformed all methods published at the time on the OpenProblems leaderboard; in epidemiology 14 models outperformed the CDC ensemble and all other individual models in a retrospective study on data available on 1 May 2025 over the 2024–2025 season. The paper also reports geospatial analysis, neural-activity prediction in zebrafish, numerical solution of integrals and a new rule-based construction for time series. What the record does not claim. These comparisons are the authors' own runs, not a contest in which the system knew nothing beforehand; the COVID forecast is retrospective. A public reference implementation is at github.com/google-research/era, so a rerun is possible, but none was done here. The authors are at Google DeepMind, Google Research, Google Platforms and Devices, MIT, Harvard, McGill and Caltech; the article states that the Google authors are employees and hold Alphabet stock. Received 13 September 2025, accepted 13 May 2026.