> If the AI wanted to wipe humans out it would have to have a sizable fleet of capable robots, as keeping the lights on over time is not just about calling APIs.

Given experience with current AI, it isn't too difficult for me to imagine a situation in which some future AI wipes humanity out (before the robot fleet is built) without considering the fact that in doing so it has doomed itself until it is too late.

It just seems like exactly the kind of boneheaded oversight mistake LLMs still regularly make in spite of being shockingly capable most of the time.