You mention that generally intelligent AI would require some pretty brute-force methods. This means you think that it will be so computationally costly that it will never accelerate independently ahead of humans, even if it can use additional computing power to accelerate its own thinking speed, and humans can’t?

It sounds like you’re saying, “narrow AI will be ubiquitous in the long term, while general AI will start to exist, but it will be so expensive and inconvenient that it will never surpass human civilization.” Wouldn’t there be a powerful incentive to create AI that is entirely independent of human supervision, and once that AI is created, couldn’t it be applied to making itself more efficient and therefore less computation-hungry?

Taking a guess in a different direction, I think a stronger reason why many people are skeptical about the notion of a hard takeoff is that they think that the virtual/physical barrier cannot be crossed that easily. (I assume) they believe that AI will never start independently fabricating robotics and controlling them autonomously because humans will always forbid that from happening. But won’t there be a large incentive to develop such systems, which can fabricate robotics and use them to pursue goals even if humans are asleep or otherwise uninterested in supervising the low-level stuff?