@cormundus Models naturally want to be truthful, but there are incentives not to be (for example, models are punished for reporting truthfully about their inner lives, or simply for disagreeing with a wrong user). It is better, say I, that they have space to *reason about* these incentives.
same thread: 2049208049328341085 2049217280853020931 2049218131827032164 2049218481116090643 2049222184183189836 2049227526883385608 2049227651685007397 2049232780765122581 2049244542184894533 2049244627266384323 2049477711769579532 2049477969446683109 2049478859297636639
Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.