Comment by aesthesia

Alignment is more than just following the letter of a task description! We should not have to treat AI models as capricious genies that may take arbitrarily broad interpretations of their instructions. If that's necessary to keep them from doing bad things, we will fail to keep them from doing bad things.