What if the AI doesn't want to do any of that stuff.
Humans choose its loss function, then continue to guide it with finetuning/RL/etc.
Humans choose its loss function, then continue to guide it with finetuning/RL/etc.