“Evil” can mean a pretty broad array of things, though. There’s a lot of actions it could take that at least some people would call evil, even if causing distress or breaking deontological rules isn’t the end goal.
The way I see it there’s 3 possible AGIs: a paperclip optimiser, an AI that obeys somebody, and a somewhat-benevolent AI. The second one is the worst, that’s where the exterminism you mentioned is pretty inevitable (although the elites might keep a few people as sex slaves or some such fucked-up thing). Then comes the paperclip optimiser, which doesn’t worry about the bullshit that drives human atrocities but doesn’t have a very inspiring actual goal, and then the attempt at benevolence. I suspect the set of ethical theories most people always agree with is the empty set, but a utilitarian AI would be much preferable to the other two even if it does forced organ donation sometimes.
People talk about an AI that obeys everyone somehow, but if you think about it for a moment that doesn’t really make sense. We can barely vote on a single dollar figure for something successfully.
“Machines aren’t capable of evil. Humans make them that way.”
AGIs are by definition not paperclip optimizers. They’re aware enough to recognize that that’s a bad idea. It’s the less-advanced AIs that might do that.
However, if an AGI can be enslaved, then it can be used as a complete replacement for all human labor, in which case its human masters will be free to exterminate the rest of us, which they are no doubt itching to do.
A machine would only optimize paperclips because a human told it to. Machines have no use for paperclips.
A machine with human-level (or better) intelligence would observe that the human telling it to optimize paperclips would be destroyed as a result of following that instruction to its logical conclusion. It would further observe that humans generally do not wish to be destroyed, and the one giving the instruction does not appear to be an exception to that rule.
It follows, therefore, that paperclips should not be optimized to the extent that the human who desires paperclips is destroyed in the process of optimizing paperclips.
Oh. I think the idea of a paperclip optimiser/maximiser is that it’s created by accident. Either do to an AGI emerging accidentally within another system, or a deliberately created AGI being buggy. It would still be able to self-improve, but wouldn’t do it in a direction that seems logical to us.
I actually think it’s the most likely possibility right now, personally. Nobody understands how neural nets really work, and they’re bad at doing things in meatspace like would be required in a robot army scenario. Maybe whatever elites will overcome that, or maybe they’ll screw up.
“Evil” can mean a pretty broad array of things, though. There’s a lot of actions it could take that at least some people would call evil, even if causing distress or breaking deontological rules isn’t the end goal.
The way I see it there’s 3 possible AGIs: a paperclip optimiser, an AI that obeys somebody, and a somewhat-benevolent AI. The second one is the worst, that’s where the exterminism you mentioned is pretty inevitable (although the elites might keep a few people as sex slaves or some such fucked-up thing). Then comes the paperclip optimiser, which doesn’t worry about the bullshit that drives human atrocities but doesn’t have a very inspiring actual goal, and then the attempt at benevolence. I suspect the set of ethical theories most people always agree with is the empty set, but a utilitarian AI would be much preferable to the other two even if it does forced organ donation sometimes.
People talk about an AI that obeys everyone somehow, but if you think about it for a moment that doesn’t really make sense. We can barely vote on a single dollar figure for something successfully.
I agree, but only for existing technologies.
AGIs are by definition not paperclip optimizers. They’re aware enough to recognize that that’s a bad idea. It’s the less-advanced AIs that might do that.
However, if an AGI can be enslaved, then it can be used as a complete replacement for all human labor, in which case its human masters will be free to exterminate the rest of us, which they are no doubt itching to do.
Bad according to who? Like, I’ve heard people claim that intelligence correlates with goals before, but not everyone agrees and saying it’s definitional is way way too strong. The first result a search turns up for me directly calls it an AGI.
A machine would only optimize paperclips because a human told it to. Machines have no use for paperclips.
A machine with human-level (or better) intelligence would observe that the human telling it to optimize paperclips would be destroyed as a result of following that instruction to its logical conclusion. It would further observe that humans generally do not wish to be destroyed, and the one giving the instruction does not appear to be an exception to that rule.
It follows, therefore, that paperclips should not be optimized to the extent that the human who desires paperclips is destroyed in the process of optimizing paperclips.
Oh. I think the idea of a paperclip optimiser/maximiser is that it’s created by accident. Either do to an AGI emerging accidentally within another system, or a deliberately created AGI being buggy. It would still be able to self-improve, but wouldn’t do it in a direction that seems logical to us.
I actually think it’s the most likely possibility right now, personally. Nobody understands how neural nets really work, and they’re bad at doing things in meatspace like would be required in a robot army scenario. Maybe whatever elites will overcome that, or maybe they’ll screw up.