I actually don’t think the distinction between slow and fast takeoff matters too much here, at least compared to what the lack of instrumental convergence offers us. The important part here is that AI misuse is a real problem, but this is importantly much more solvable, because misuse isn’t as convergent as the hypothesized instrumental convergence is. It matters, but this is a problem that relies on drastically different methods, and importantly still reduces the danger expected from AI.
I actually don’t think the distinction between slow and fast takeoff matters too much here, at least compared to what the lack of instrumental convergence offers us. The important part here is that AI misuse is a real problem, but this is importantly much more solvable, because misuse isn’t as convergent as the hypothesized instrumental convergence is. It matters, but this is a problem that relies on drastically different methods, and importantly still reduces the danger expected from AI.