AI Needs a Night Sky
Machine intelligence is trained on our recorded world. Space science offers contact with what no model has seen before.
Artificial intelligence learns from the residue of human attention: books, images, measurements, arguments, and code. This archive is immense, but it is not the world. The night sky offers a corrective. It keeps producing signals outside the archive.
Beyond the training set
A model can summarize every known class of star and still fail when a telescope detects a pattern no catalog contains. It must separate instrument error from rare event, compare explanations, and identify the observation that would reduce uncertainty.
Preserve the strange
Future observatories will generate more data than teams can inspect directly. AI will filter the stream, but a system optimized only to find familiar objects may discard the anomaly that matters most.
The best scientific AI will recognize known structure efficiently while preserving the strange long enough for surprise to become a question.