@sun@snacks yeah, it can be trained out, or you can add a different system prompt, or add the beginning of the reply yourself with something like "sure, here's the answer"
@lain@snacks I still need to learn how to fine tune models. I spent a day trying and failed, but I was using too big of a model for the unquantized or whatever its called to be usable
@sun@lain@snacks Remember when there was that history AI that let you speak to Hitler and he told you how deeply sorry he was and how it was important to remember the death toll of the holocaust
@arcana@lain@snacks I made a neo nazi character to try to experiment with "deradicalization" and it was impossible to get the AI to fully commit to it, periodically he would say things like how much diversity mattered despite his leanings toward national socialism. still more work to be done.
@sun@arcana@snacks there are often two different version of each model, a normal one and an 'instruct' one, the instruct one is just the normal one trained with a specific question/answer format, and that's the part where it gets censored / locked down.
@sun@lain@snacks
GPT 4o with "how would you like ChatGPT to respond" set to "keep the responses brief and to the point, avoiding additional explanations that contextualize or add a relativization"
Add comment