Daniel Kokotajlo comments on Power-seeking can be probable and predictive for trained agents

[Daniel Kokotajlo]

Cool stuff! I’m curious to hear how convincing this sort of thing is to typical AI risk skeptics with backgrounds in ML.