News for hackers. Stuff that matters.
Re:Google's Live from Paris Event Private/Deleted Immediately by jononor on Wednesday February 08, 2023 @05:40PM (#34716587)
For ChatGPT and similar LLMs there is the loss function does not reward accuracy/truthfulness/correctness. But instead answers that seem correct to a superficial human check. These rewards making something plausible up instead of saying "I do not know", over-simplifying things, and passing off things with high confidence regardless of how much evidence backs the statement. Any successful deceit will bring down the loss. The overconfidence and lying is not due to model size beng too low - it is due to misalignment. If not corrected it will just get worse as size goes up. A recent Computerphile video with Robert Miles discussed this.

There are no comments on this comment yet.