What was built
Anthropic published an account of a full-sky ultraviolet map on 8 October, assembled by AI agents under the direction of Brice Ménard, an astrophysicist at Johns Hopkins University and a researcher at the company. Ménard describes his role as guiding the agents; the agents did the bulk of the work.
Ultraviolet light shows dust lit by starlight. The ozone layer blocks it, so it can only be measured from space, and no complete map existed. NASA’s GALEX ran from 2003 to 2013 and imaged about two-thirds of the sky in roughly 38,000 observations, deliberately skipping bright regions including the plane of the Milky Way. The map combines far-ultraviolet light at 154 nanometres and near-ultraviolet at 232.
How the agents did it
The agents searched for public ultraviolet surveys, then made each one internally consistent, removing glare from bright stars with many agents working in parallel across regions of sky. They cross-calibrated the surveys, matched their resolution and mapped them onto a common coordinate system before merging.
GALEX supplied the bulk. NASA’s Swift, Korea’s FIMS/SPEAR and Europe’s TD-1 filled in some of what was left. For the roughly one-third of the sky never observed in ultraviolet at all, the agents used inpainting, trained on the two-thirds that had been observed together with visible, infrared and radio data. ESA’s Planck supplied gap-filling templates alongside Hα templates, and Gaia provided visible-light measurements used to estimate the ultraviolet output of more than 100 million individual stars.
How they checked it
The validation was to hide parts of regions that had been measured and ask the model to predict them. Those predictions landed within about 10% of the real ultraviolet measurements.
That figure is the reason the map is usable rather than decorative, and it is also the reason the labelling matters. Every pixel in the final map is marked either measured or predicted, and carries an uncertainty estimate. About a third of the map is prediction.
The error the agents missed
The most instructive detail in Anthropic’s account is a failure. Circular footprints from individual GALEX observations showed up in the merged map, caused by residual atmospheric glow. Two rounds of agent review did not catch them.
Ménard did, and asked in plain language: “I can see discs with the imprint of individual observations; can you correct that?” Claude then corrected all 38,000 observations, which took a couple of hours of processing.
That is the shape of the result. The agents handled a volume of calibration work no single researcher would attempt, and a human noticed an artefact two automated review passes had signed off.
What it is for
The work ran over several days during the summer and went through more than a dozen successive versions. Anthropic presents the map as teaching material rather than a research instrument, and Ménard’s wider point is about the projects that did not happen: he suspects many scientists have postponed similar work that is now feasible.
What to watch
The claim worth testing is not the map but the method. A survey merge with published uncertainties and a measured-versus-predicted label on every pixel is checkable by anyone who works with the same data, and the ten per cent validation figure is Anthropic’s own.
The useful signal will be whether astronomers outside the company take the map apart, and whether the inpainted third holds up where new ultraviolet observations eventually land on it.
