I consider myself “pro-AI”, which isn’t quite “e/acc” (that’s an online cluster I’m not really part of & don’t speak for).
I think:
a.) generally technology is good for humanity in the long run and banning it is bad, and the prior ought to be that AI is no exception.
b.) I don’t know with much confidence what’s going to happen with AI. I can’t rule out the bad outcomes. I’m not “an optimist” in that sense. Though if you forced me to guess, my guess would be that humanity doesn’t get wiped out entirely, though there might be some major catastrophes. It’s more that I’m ~deontologically committed to not banning tech, come what may. In particular, I also am not “an optimist” about humanity’s future generally; I think civilizations collapse, catastrophes happen, and it’s quite possible that some kind of “the Shit Hits The Fan” will occur in our lifetime with or without AI, though again, probably not a true Homo sapiens extinction event. I would sort of rather we rolled the dice and tried to create the most powerful creations we can, so that even if it ends in catastrophe we’ve at least done something impressive with our time on this planet, than a long term future that’s basically dissipated energy, decline, tech collapse, massively catastrophic war, etc.
c.) I think AI is cool and interesting and powerful (fundamentally, because minds and computation always have been.) My sympathies are with people who want to work on it!
d.) Safety/alignment research is very important and good, so long as it doesn’t preclude the AI industry continuing to exist and trying to move forward. It’s not “anti-car” to think cars should have safety features. It’s not “anti-nuclear” to think nuclear power plants should be extremely careful about safety. I am more or less okay with regulation that requires AI to be built with certain safety practices, though ideally I’d prefer to see that done through voluntary means like insurance instead. I think “alignment” is sometimes ill posed, but lots of things that fall under that umbrella are straightforwardly good.
e.) I’m honestly a little bit down on planning ahead, partly because I’m not good at it or temperamentally suited to it. I accept, rather than rail against, the human tendency to not worry about problems until they show up. “Near mode” is for things that are, in fact, near; a bunch of intuitions kick in when something feels “like real life”, and also a bunch of information shows up that you didn’t have earlier. If you try to plan way ahead, it’s not fun, it’s restrictive, and it’s also kind of error prone. Yes, if you “fuck around and find out”, you may get bit quite badly; maybe this is foolish; but I’m pretty foolish. I’m not going to say everyone should always FAFO, again, I think it’s nice that airplanes don’t fall out of the sky and sanitary food gets delivered to the grocery store, and that takes planning, but usually that’s near-mode planning derived from experience, which is more robust than theory. I think a common rationalist failure mode is to assume coordinated planning ahead works, when in fact, ignorance, shifting motivations, and disagreement/misunderstanding/conflict between different people (and different groups of people) means you basically can’t do very much “steering” of anything too big or distant. I feel like, compared to the norm in some Silicon Valley spaces, I’m a bit of a controversial radical for believing that it’s worth it for a single, small, new organization to write down a plan for its own activities ahead of time. But even there, I think it would be silly to assume that the org would follow the plan without having to change it.
f.) I used to be much more firmly “actually-existing AI is fine to make progress on, because it’s not that powerful”, and now I’m uncertain. This stuff does seem pretty powerful! But we don’t know how powerful, exactly. I do think there are some fundamental differences between current AI and animal/organismal intelligence. It’s possible that, at least for a while, we could make a bunch of AI progress on within a paradigm that fundamentally can’t do the things necessary to cause x-risks. It would be a shame to miss out on economic productivity for nothing.
g.) Also I think it would be kind of lame to simply stop. “Humanity was about to create these impossibly powerful entities, and then it didn’t” is a sad story. What? We lost our nerve? We backed out? Even in the most sympathetic possible framing, it’s like stepping back from the cliff. That makes a terrible story! That’s how you want the story to end? The foolhardy tragic hero always walks off the cliff!
I consider myself “pro-AI”, which isn’t quite “e/acc” (that’s an online cluster I’m not really part of & don’t speak for).
I think:
a.) generally technology is good for humanity in the long run and banning it is bad, and the prior ought to be that AI is no exception.
b.) I don’t know with much confidence what’s going to happen with AI. I can’t rule out the bad outcomes. I’m not “an optimist” in that sense. Though if you forced me to guess, my guess would be that humanity doesn’t get wiped out entirely, though there might be some major catastrophes. It’s more that I’m ~deontologically committed to not banning tech, come what may. In particular, I also am not “an optimist” about humanity’s future generally; I think civilizations collapse, catastrophes happen, and it’s quite possible that some kind of “the Shit Hits The Fan” will occur in our lifetime with or without AI, though again, probably not a true Homo sapiens extinction event. I would sort of rather we rolled the dice and tried to create the most powerful creations we can, so that even if it ends in catastrophe we’ve at least done something impressive with our time on this planet, than a long term future that’s basically dissipated energy, decline, tech collapse, massively catastrophic war, etc.
c.) I think AI is cool and interesting and powerful (fundamentally, because minds and computation always have been.) My sympathies are with people who want to work on it!
d.) Safety/alignment research is very important and good, so long as it doesn’t preclude the AI industry continuing to exist and trying to move forward. It’s not “anti-car” to think cars should have safety features. It’s not “anti-nuclear” to think nuclear power plants should be extremely careful about safety. I am more or less okay with regulation that requires AI to be built with certain safety practices, though ideally I’d prefer to see that done through voluntary means like insurance instead. I think “alignment” is sometimes ill posed, but lots of things that fall under that umbrella are straightforwardly good.
e.) I’m honestly a little bit down on planning ahead, partly because I’m not good at it or temperamentally suited to it. I accept, rather than rail against, the human tendency to not worry about problems until they show up. “Near mode” is for things that are, in fact, near; a bunch of intuitions kick in when something feels “like real life”, and also a bunch of information shows up that you didn’t have earlier. If you try to plan way ahead, it’s not fun, it’s restrictive, and it’s also kind of error prone. Yes, if you “fuck around and find out”, you may get bit quite badly; maybe this is foolish; but I’m pretty foolish. I’m not going to say everyone should always FAFO, again, I think it’s nice that airplanes don’t fall out of the sky and sanitary food gets delivered to the grocery store, and that takes planning, but usually that’s near-mode planning derived from experience, which is more robust than theory. I think a common rationalist failure mode is to assume coordinated planning ahead works, when in fact, ignorance, shifting motivations, and disagreement/misunderstanding/conflict between different people (and different groups of people) means you basically can’t do very much “steering” of anything too big or distant. I feel like, compared to the norm in some Silicon Valley spaces, I’m a bit of a controversial radical for believing that it’s worth it for a single, small, new organization to write down a plan for its own activities ahead of time. But even there, I think it would be silly to assume that the org would follow the plan without having to change it.
f.) I used to be much more firmly “actually-existing AI is fine to make progress on, because it’s not that powerful”, and now I’m uncertain. This stuff does seem pretty powerful! But we don’t know how powerful, exactly. I do think there are some fundamental differences between current AI and animal/organismal intelligence. It’s possible that, at least for a while, we could make a bunch of AI progress on within a paradigm that fundamentally can’t do the things necessary to cause x-risks. It would be a shame to miss out on economic productivity for nothing.
g.) Also I think it would be kind of lame to simply stop. “Humanity was about to create these impossibly powerful entities, and then it didn’t” is a sad story. What? We lost our nerve? We backed out? Even in the most sympathetic possible framing, it’s like stepping back from the cliff. That makes a terrible story! That’s how you want the story to end? The foolhardy tragic hero always walks off the cliff!