you cherry-pick the one possibility where somehow all this turns out to work except that the AI does not stop at a certain point but goes on to consume the universe
It doesn’t have to consume the universe. It doesn’t even have to recursively self-improve, or even self-improve at all. Simple copying could be enough to say, wipe out every PC on the internet or accidentally crash the world economy.
(You know, things that human level intelligences can already do.)
IOW, to be dangerous, all it has to be able to affect humans, and be unpredictable—either due to it being smart, or humans making dumb mistakes. That’s all.
It doesn’t have to consume the universe. It doesn’t even have to recursively self-improve, or even self-improve at all. Simple copying could be enough to say, wipe out every PC on the internet or accidentally crash the world economy.
(You know, things that human level intelligences can already do.)
IOW, to be dangerous, all it has to be able to affect humans, and be unpredictable—either due to it being smart, or humans making dumb mistakes. That’s all.