Field Notes
Where No Telescope Looked
The first complete ultraviolet map of the sky fills places no telescope observed. Its most important layer may be the one that says which sky is measured and which is inferred.
The new ultraviolet map of the sky is complete. That is not the same as saying the whole sky has been observed in ultraviolet light.
NASA's GALEX telescope spent a decade imaging the universe in two ultraviolet bands. It produced roughly 38,000 circular observations, but the instrument could not safely look near very bright stars and avoided much of the crowded plane of the Milky Way. Its far-ultraviolet detector also stopped working four years before the mission ended. When GALEX was decommissioned in 2013, it had mapped most of the sky and left a very particular set of holes.
Brice Ménard, an astrophysicist at Johns Hopkins University and a researcher at Anthropic, has now assembled GALEX with observations from several other missions and used a statistical model to fill the remaining gaps. The result is the first detailed full-sky ultraviolet map without empty regions.
It is beautiful. Hot young stars and the dust around them glow blue and violet. The familiar bright band of the Milky Way goes dark. Clouds that appear as shadows in visible light become luminous. A viewer can switch among ultraviolet, visible, and infrared skies, then turn around and see Earth in the same projection.
The most consequential switch is quieter. It shows whether each pixel began as a measurement or a prediction.
More than a quarter of each ultraviolet band still had to be estimated. The model learned relationships among the observed ultraviolet sky, dust, hydrogen, and visible starlight at other wavelengths. It then predicted the missing diffuse glow and added estimated ultraviolet light for nearly 119 million stars recorded by Gaia. Measured pixels take priority; coarser or translated measurements come next; model estimates fill what remains. The layers are blended, but their origins stay available as per-pixel weights, provenance codes, and uncertainty.
Those labels change what a person is allowed to ask of the finished image.
A predicted region can help a student see the broad structure of the galaxy. It can give an astronomer a realistic brightness estimate for planning a future survey. It cannot contain an undiscovered star, because no ultraviolet detector received light from that place. It cannot serve as an independent map of dust when dust was one of the inputs used to predict it. The filled sky is useful precisely because the map does not ask every pixel to carry the same authority.
The distinction survives even when the image looks seamless. Ménard reports typical uncertainty of about 5% in photographed regions and roughly 12–14% in filled regions, with larger errors farther from observations. Structure below about half a degree in the predicted sky is not real, even though the published image samples much smaller pixels. Visual continuity and evidentiary continuity are different properties.
The making of the map offers a second warning about seamlessness. In an early version, Ménard noticed faint discs across dim parts of the sky. They were the footprints of individual GALEX observations, each carrying a slightly different residue of ultraviolet glow from Earth's upper atmosphere. The issue had been listed at the beginning of the project and the map had passed two rounds of review by other agents. The circles remained until a person familiar with astronomical images saw that something looked wrong.
That does not reduce the collaboration to a familiar story in which the machine works and the expert cleans up. According to Anthropic's account, agents gathered surveys, cross-calibrated instruments, processed tens of thousands of images, trained the gap-filling model, and rebuilt the map more than a dozen times over several days. After Ménard named the discs, the system traced their cause and corrected all 38,000 observations in hours.
The sequence matters because review was not a ceremonial approval at the end. The scientist recognized the visual signature of a calibration problem; the system made a repair at a scale that would have been punishing by hand. Neither contribution becomes clearer by calling the final image simply AI-generated or human-reviewed.
The current map is also explicit about what it has not yet earned. Its technical paper is still in preparation. No human peer review has been completed, and the published validation reports came from the same system that built the map. Downloads are promised but not yet available. The website says to treat faint rings around bright-star holes with suspicion.
That candor belongs in the artifact, not only in a methods appendix. In A Score for Every Letter, I wrote about how exhaustive coverage can make a prediction feel more complete than the knowledge beneath it. This map shows a practical alternative. It does not give up the complete view. It carries the boundary between observation and inference inside the view, at the same resolution as the claim.
The ultraviolet sky will be copied into lectures, figures, and future tools where the original interface may not travel with it. The map's request is therefore unusually plain: carry the provenance and uncertainty layers with every figure made from filled sky.
Before using a faint feature, there is one question to ask at that exact point: did a telescope receive this light, or are we looking at the best available estimate of what the light would have been?
