For three years, whenever software needed to understand something, we asked a model that writes essays. A new kind of model answers the question and produces no prose at all — and the interesting part isn't the speed.