1974: "No-one would put an AGI in a bomb, that's just ridiculous"
1986: "No-one is going to put an 80386 in a vending machine, that's just ridiculous"
1986: "No-one is going to put an 80386 in a vending machine, that's just ridiculous"
We’re supposing that a conscious being makes significant advances in use cases where standard automation hardware+software is applied today.
I’m not saying ML isn’t a major shift, but we’re talking about AGI, I don’t think any use case exists unless it’s specifically meant to wow fleshy human beings. A trained ML model in the domain you’re working in, with out the pesky conscious, seems like it’s the boring efficient end state for automation.
The Japanese tried this with plain-ol' non-artificial intelligence, and it seemed to have been somewhat effective, if not the most sustainable.
Why would I want the bomb to think? I want the bomb to follow some pretty specific instructions.
It probably hinges on how rapidly the field of battle is evolving.
Do you want your bomb to at least attempt to adapt in realtime to novel targets? To novel means of camouflage, defense, and even interdiction (rather than wait for reports in the field to eventually prompt a software patch to upgrade the weapons system)? Well, the bomb is going to have to be considerably smarter to do so.
Of course, there are just as many ways that sort of on-board capability can go awry. I imagine that painting noncombatant symbols (or the equivalent adversarial input) on combat vehicle roofs may serve to fool overly smart munitions for a short while, for example.
I don’t know why you’d prefer a “human being” to control the bomb, when most of our technological history in this space is trying to remove humans from the equation considering how many errors they make.
In a true AGI, one advanced enough to make the original poster concerned about the morals of enslaving it, MOST of it’s training is on things unrelated to being a successful bomb.