Study: LLMs should not be used to generate confidence scores

A widely discussed Hacker News post argues that having LLMs produce probability or confidence scores for agents is untrustworthy. Confidence should be computed from information via a rule, the author says, so agent designers should not rely on such numbers.