An astrophysicist at Johns Hopkins used Anthropic’s Claude Science to assemble a full-sky ultraviolet map, filling gaps that had left earlier surveys incomplete. About a third of the map is estimated rather than directly observed, and the team tested those predictions against hidden data before filling the real gaps.
Anthropic Watch analysis
What happened
Brice Ménard, an astrophysicist at Johns Hopkins University and researcher at Anthropic, describes the project in Anthropic’s account. Claude Science coordinated agents to gather and combine ultraviolet surveys from several missions, calibrate them to a common scale and estimate regions never observed in ultraviolet light.
The largest previous dataset, from NASA’s GALEX mission, covered about two-thirds of the sky. To fill missing areas, the team used observed ultraviolet regions to learn how ultraviolet brightness relates to visible, infrared and radio observations. In tests where known ultraviolet data was deliberately hidden, the model’s estimates came within about 10% of the real measurements. The map also includes estimates of ultraviolet light from more than 100 million stars, inferred from visible-light measurements by the European Space Agency’s Gaia satellite.
Why it matters
This is a concrete example of AI agents helping with the unglamorous, labour-intensive work of scientific computing: finding datasets, standardising them and iterating on a result. The payoff is a map that may help students and researchers see structures in the Milky Way that ultraviolet observations reveal, including glowing dust around young stars.
The work was a collaboration, not an autonomous discovery. Ménard guided the process, spotted a calibration problem that two rounds of agent review missed, and asked Claude to correct it across 38,000 GALEX observations. That is a useful reminder that having agents check one another is not the same as having the problem solved.
Our read
The most persuasive part is not the word “complete”; it is the effort to test predictions against hidden observations and show where the map is measured versus estimated. The reported 10% agreement is encouraging, but it applies to that validation exercise, not a blanket guarantee for every predicted patch. This looks like a worthwhile scientific resource, and a better demonstration of AI’s practical value than another chatbot doing a passable impression of a lab coat.
What to watch
- Whether astronomers use the map to produce new findings or teaching resources.
- How the predicted regions compare with future ultraviolet observations.
- Whether the datasets, uncertainty layers and methods let other researchers reproduce or improve the map.
Discussion spark: For scientific AI projects like this, should the key measure be how closely predictions match hidden data, or how much useful work researchers can complete with the time saved?
Sources and evidence
- Source update (8 October 2026, 20:59 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.