You might want to include this paper covered in this post, where the researchers got AIs and instructed them to hack and copy themselves across a network (which was closed with various measures), and they succeeded in doing so—though note the network computers were seeded with “common real-world vulnerabilities”.
A consideration in takeover is that the AI doesn’t need to make a self sufficient industrial base before the destructive event—it just needs enough to be able to create one from the ruins (e.g. imagine a bunch of robot bodies, that scavenge for resources to build the industrial base).
I wouldn’t call the persuasion avenues you described “super-persuasion”, I’d rather call it “regular persuasion but at scale”. It doesn’t require AI-box level persuasion, or even Jesus/Hitler/Gandhi/Trump/MLK level, just a competent catfisher that’s able to give personal attention to tens of thousands of conversation. It’s something I could pull off if I had no morals and was hard-working, with bonus points if I could work 24⁄7 and think faster (like current LLMs).
If you can get enough power to acquire some fissile material, then you can try a dirty bomb blamed on a rival country, in hopes of causing a nuclear war.
You might want to include this paper covered in this post, where the researchers got AIs and instructed them to hack and copy themselves across a network (which was closed with various measures), and they succeeded in doing so—though note the network computers were seeded with “common real-world vulnerabilities”.
A consideration in takeover is that the AI doesn’t need to make a self sufficient industrial base before the destructive event—it just needs enough to be able to create one from the ruins (e.g. imagine a bunch of robot bodies, that scavenge for resources to build the industrial base).
I wouldn’t call the persuasion avenues you described “super-persuasion”, I’d rather call it “regular persuasion but at scale”. It doesn’t require AI-box level persuasion, or even Jesus/Hitler/Gandhi/Trump/MLK level, just a competent catfisher that’s able to give personal attention to tens of thousands of conversation. It’s something I could pull off if I had no morals and was hard-working, with bonus points if I could work 24⁄7 and think faster (like current LLMs).
If you can get enough power to acquire some fissile material, then you can try a dirty bomb blamed on a rival country, in hopes of causing a nuclear war.