10 minutes in and he's articulated the actual stakes of x-risk vs. pause/slowdown better than anyone else, despite being completely wrong about ai for the past 10 years