Hallucination, calibration and grounding

A · 10-second intuition A language model produces likely text, not checked facts; when it does not know, a fluent guess is often more likely than “I don't know”.

Not started
Predict before you read

Why does a base model state wrong facts fluently?