A network of radio telescopes scattered across the planet catches the whisper of a distant black hole. Each antenna records a signal, but how do you combine these recordings if the clocks at the stations tick at different rates and the atmosphere distorts incoming waves? Usually, data are first 'calibrated'—adjusted to a reference, but the reference itself can be wrong. Then the image turns out like a puzzle assembled from crooked pieces.

These 'honest' fragments are exactly what the GenDIReCT algorithm is fed. It works like an experienced restorer: first, it learns from thousands of ordinary photographs (cats, airplanes, cars) and then it takes an incomplete set of closure invariants and fills in the picture. No astrophysical templates: the model doesn't know how a black hole 'should' look, so it doesn't force the result to match expectations.

The test on real data is impressive. For the blazar 3C279, the algorithm distinguished three bright clumps, and one of them moved at an apparent speed ten times faster than light — an illusion created by a relativistic jet aimed almost directly at us. For Centaurus A, it produced two glowing edges along the jet, and their opening angle matched what the Event Horizon Telescope had previously found.

Previously, only the EHT collaboration produced images of event horizons, and skeptics asked: isn't the complex processing injecting what we want to see? Now there's an independent witness. It was trained on earthly pictures, yet it spotted cosmic jets, so those jets really are in the data, not in the astronomers' imagination.

The puzzle with no picture on the box is assembled. And although details are missing, the overall silhouette is unmistakable: we're looking at the vicinity of a black hole, seen without a single trusted telescope.