Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Everyday tasks

ProtoStar: public science, made AI-ready

Written November 12, 2026

Publicly funded science is public, but rarely usable by software. ProtoStar, announced October 29, 2026, pushes universities to release datasets, models and code in formats AI systems can work with directly — starting with a climate forecasting challenge scored against observed weather.

A catch-up read, not a launch

Project ProtoStar was announced on October 29, 2026, about two weeks before this page. It is an initiative to make publicly funded scientific research easier to discover, reuse and build on, by releasing datasets, models and code in AI-ready, machine-actionable formats, and to show the value of doing so through a series of open challenges. If you work with public science data, the useful question is not whether the announcement is new. It is what AI-ready concretely means, and what you can do with the first release.

Why public is not the same as usable

The United States federal government invests approximately $190 billion a year in research and development, and agencies such as the National Institutes of Health and the National Science Foundation increasingly expect the results to be shared publicly, with a plan for sharing the underlying data. Sharing on paper is not the same as sharing in practice. ProtoStar's announcement names the gap plainly: datasets, models and code are often scattered across incompatible systems, stored in static or legacy formats, or released without the documentation, metadata and provenance needed to understand and reuse them. An AI system trying to combine resources across disciplines hits that gap first.

The raw material is already large. Hugging Face hosts more than 17,000 public datasets across core scientific domains — genomics, protein design, chemistry, weather and climate, Earth observation and astronomy — and that figure approaches 29,000 when medicine and mathematics are included. Researchers shared more than three times as many new science datasets on Hugging Face in the first nine months of 2026 as in the same period of 2025.

Who is behind it

The founding university partners are Carnegie Mellon University, Purdue University, Georgia Institute of Technology, Arizona State University and The George Washington University, working through the Open Forum for AI. The Astera Institute is a partner, Hugging Face is the infrastructure partner, and the Open Research Community Accelerator is the community partner. Organizations formalize their involvement through a memorandum of understanding supporting the release of suitable datasets, models and code in AI-ready formats on open infrastructure. Carnegie Mellon's Open Source Programs Office is already encouraging its researchers to share suitable outputs on Hugging Face as an early example of that commitment.

The first challenge: climate, with a baseline

The first challenge, Climate Risk Intelligence, is hosted by Carnegie Mellon University and runs from November 2 to November 13, 2026. Teams use historical weather data from 210 United States cities: records through 2018 are the input, and the task is to predict actual conditions from 2019 to 2026, so forecasts are scored against observed weather rather than hypothetical scenarios.

There are four forecast targets, all relevant to urban planning: average annual temperature, total annual precipitation, the hottest day of the year, and cooling degree days, a measure of demand for air conditioning. Predictions are compared against simple statistical baselines, including historical averages and linear trends, to test whether AI approaches improve accuracy at all. Teams also propose a short climate adaptation policy idea based on what they learned from the data. Entries are submitted as a Hugging Face dataset, forecasting performance goes on a public leaderboard, and final results are due in early 2027, once complete 2026 weather records are available. Later challenges, hosted by other participating institutions in sequence, will follow the same pattern around different resources.

The evidence that documentation is the bottleneck

Two figures in the announcement carry the argument. The 18 public repositories maintained by the National Center for Biotechnology Information on Hugging Face have received more than 11 million downloads, and NASA IMPACT's 27 repositories approximately 770,000 downloads, as of October 26, 2026. And 90 percent of dataset downloads on Hugging Face come from fully documented datasets. The demand follows the documentation. That is the bet ProtoStar is making: fix the packaging, and reuse follows.

Limits

Related playbooks

Sources

Back to: Everyday playbooks hub · Also: Scientist World: 1,000 executable research tasks · A human-agent team ran X-ray experiments you can audit · AlphaGenome 2 reads variants per cell type