A Third of the Sky Was Never Seen in Ultraviolet. Claude Mapped It Anyway, From 38,000 GALEX Exposures to 119 Million Stars

For years, astrophysicist Brice Menard had to apologize to his students. When his astrophysics course reached the ultraviolet sky — the wavelength that reveals dust lit up by starlight, from clouds around young stars to rings left by stellar explosions — the only all-sky UV map he could show was full of holes. This week that changed: Anthropic published an account of how Menard, a Johns Hopkins University professor who also works as a researcher at Anthropic, produced the first complete map of the entire sky in ultraviolet light, with roughly one-third of it — including much of the Milky Way's plane — filled in by prediction rather than direct observation.

The hole had a simple physical cause. Earth's ozone layer absorbs ultraviolet light, so the only way to observe it is from space. NASA's GALEX mission, which ran from 2003 to 2013, provided the largest dataset: about 38,000 individual observations covering roughly two-thirds of the sky. But GALEX deliberately skipped regions crowded with bright stars, including the galactic plane, to protect its detectors from damage. Other telescopes such as NASA's Swift and South Korea's FIMS/SPEAR added partial coverage, yet significant gaps remained. No successor mission has flown: NASA's next UV surveyor, UVEX, carries an estimated price tag of about $300 million and is not scheduled to launch until 2030.

Menard's instruction to Claude Science, Anthropic's research workbench launched at the end of June, was easy to state and hard to execute: gather every available ultraviolet dataset, put them all on a common scale, merge them into one map, and fill in every patch of sky no UV telescope has ever observed. Claude orchestrated a team of AI agents to do it. The agents searched the web for publicly available UV surveys, downloaded tens of thousands of GALEX images taken under varying conditions across a decade, and made each survey internally consistent. Bright stars posed the trickiest problem: just as a streetlight adds glare to a photo, stars add glare to telescope images, and that glare had to be removed before the faint UV light around them could be measured. Many agents worked in parallel on different regions of the sky to even out these differences.

For the unobserved third of the sky, the system used a statistical form of inpainting: it learned the relationship between UV brightness and observations at other wavelengths — visible, infrared and radio — from the mapped regions, then applied that relationship to predict what the missing patches should look like. To check the predictions, Menard had the system mask known regions, treat them as blank, and reconstruct them; the filled-in values landed within about 10 percent of the real measurements. Every pixel of the final map is labeled as either "measured" or "predicted," with uncertainty estimates attached. On top of the background, the system used the European Space Agency's Gaia catalog to estimate the UV brightness of 119 million individual stars, layering them onto the map one by one.

The collaboration ran over several days and more than a dozen major versions, with Menard steering between computation runs while Claude worked for hours at a stretch. It was not flawless. One evening Menard noticed faint circular marks in the darkest regions — the footprints of individual GALEX exposures, caused by uneven ultraviolet airglow from Earth's atmosphere that had not been fully subtracted. The issue had been on the project's known-problems list from the start, and two rounds of AI review had missed it. Menard told Claude he could see the discs; the agents traced them to airglow residue and reprocessed all 38,000 observations over several hours until the circles vanished.

Menard framed the project as a category of work larger than astronomy. Every field, he argued, has a backlog of projects that would explain a key concept or help other researchers but never rises high enough on anyone's list — weeks of meticulous, pixel-level calibration that scientists postpone indefinitely. "This time, I did not sacrifice research time I would otherwise spend on my research," he wrote, noting that his own contribution was limited to guiding the agents. Anthropic's post is careful to note that complete does not mean fully measured, and encourages anyone using the predicted regions to carry the labels along with the map.

Not a single new photon was collected: everything in the map comes from decades of publicly funded survey data, reprocessed and stitched together in days. When UVEX finally launches in 2030, astronomers will get a fresh, direct view of the ultraviolet sky. Until then, the map stands as a demonstration of a division of labor that is quietly spreading through science — agents handle the patient, well-specified labor, while the human's remaining job is knowing what the result should look like and spotting the artifact that two rounds of machine review waved through.