afficheur criminellement sous-abonné compte tenu de la quantité de qualité et/ou des goûts provocateurs de réflexion qu'il publie
mike64_t
mike64_t18 oct., 12:33
I think the observation that LLMs are "bad tutors" in that they cannot precisely probe understanding is accurate. The fact that "upweighting the entire rollout" is stupid is also true. However its not obvious to me that the remedy for that is LLM-reflection as to "what went well". I think this runs into very similar issues of collapse-risk or misallocation of supervision. Because while we might be sucking supervision through a straw, the only thing thats even worse is sucking tainted supervision through a straw.
ce n'est pas comme si Mike était un poster de niche ou quoi que ce soit, mais je pense juste à combien de comptes pourris ont 2 à 10 fois plus.
30,78K