I think GPT-4 fine-tuning at the time of ChatGPT release probably would have been about as good as GPT-3.5 fine-tuning actually was when ChatGPT was actually released. (Which wasn’t very good, e.g. jailbreaks were trivial and it always stuck to its previous answers even if a mistake was pointed out.)
I think GPT-4 fine-tuning at the time of ChatGPT release probably would have been about as good as GPT-3.5 fine-tuning actually was when ChatGPT was actually released. (Which wasn’t very good, e.g. jailbreaks were trivial and it always stuck to its previous answers even if a mistake was pointed out.)
If GPT-3.5 had similarly misaligned attitudes, it wasn’t lucid enough to insist on them, and so was still more ready for release than GPT-4.