Models are often wrong but sometimes useful. Models that provide answers couched in a certain level of confidence are miscalibrated when all answers are given c

Models are often wrong but sometimes useful. Models that provide answers couched in a certain level of confidence are miscalibrated when all answers are given confidently. New training paradigms attempt to better calibrate model confidence in post-training, but clearly there are competing incentives to give answers confidently given the economics of the AI arms race.