I think the most relevant takeaway is that we did end up with an arsenal of weapons that have now put us, at all times, hours away from nuclear winter by a very reasonable metric of counterfactual possibility.
And while nuclear winter in practice probably wouldn’t be quite an extinction-level event from what I hear, it was still a very counterfactually close possibility that a nuke’s surprisingly runaway chain reaction could have been just a little more runaway.
Participants in this kind of dialogue should come in with a healthy respect for the likelihood that a big extinction risk will become salient when research figures out how to harness a new kind of power.
As much as it maybe ruins the fun for me to just point out the message: the major point of the story was that you weren’t supposed to condition on us knowing that nuclear weapons are real, and instead ask whether the Gradualist or Catastrophist’s arguments actually make sense given what they knew.
That’s the situation I think we’re in with Fast AI Takeoff. We’re trying to interpret what the existence of general intelligences like humans (the Sun) implies for future progress on ML algorithms (normal explosives), without either a clear underlying theory for what the Sun’s power really is, or any direct evidence that there’ll be a jump.
Richard Ngo: You don’t have a specific argument about utility functions and their relationship to AGIs in a precise, technical way. Instead, it’s more like utility functions are like a pointer towards the type of later theory that will give us a much more precise understanding of how to think about intelligence and agency and AGIs pursuing goals and so on. And to Eliezer, it seems like we’ve got a bunch of different handles on what the shape of this larger scale theory might look like, but he can’t really explain it in precise terms. It’s maybe in the same way that for any other scientific theory, before you latch onto it, you can only gesture towards a bunch of different intuitions that you have and be like, “Hey guys, there are these links between them that I can’t make precise or rigorous or formal at this point.”
In my opinion the relevant detail is that we were not able to prevent the Soviets from getting the bomb. It took them all of about 3 years. It’ll take China, Russia, open source hackers et al about 18 months max to replicate AGI once it arrives. So much for your decisive strategic advantage.
Nuclear technology is hours away from reducing the value of all human civilization by 10%, and for all we knew that figure could have been 100%. That’s the nuclear threat. I wouldn’t even classify that as a “geopolitical” threat. The fact that Soviet nuclear technology pretty quickly became comparable to US nuclear technology isn’t the most salient fact in the story. The story is that research got really close, and is still really close, to releasing hell, and the door to hell looks generally pretty easy to open.
18 months is more than enough to get a DSA if AGI turns out anything we fear (that is, something really powerful and difficult to control, probably arriving fast at such state through an intelligence explosion).
In fact, I’d even argue 18 days might be enough. AI is already beginning to solve protein folding (Alphafold). If it progresses from there and builds a nanosystem, that’s more than enough to get a DSA aka take over the world. We currently see AIs like MuZero learning in hours what would take a lifetime for a human to learn, so it wouldn’t surprise me an advanced AI solving advanced nanotech in a few days.
Whether the first AGI will be aligned or not is way more concerning. Not because who gets there first isn’t also extremely important. Only because getting there first is the “easy” part.
I don’t really think advanced AI can be compared to atomic bombs. The former is a way more explosive technology, pun intended.
I think the most relevant takeaway is that we did end up with an arsenal of weapons that have now put us, at all times, hours away from nuclear winter by a very reasonable metric of counterfactual possibility.
And while nuclear winter in practice probably wouldn’t be quite an extinction-level event from what I hear, it was still a very counterfactually close possibility that a nuke’s surprisingly runaway chain reaction could have been just a little more runaway.
Participants in this kind of dialogue should come in with a healthy respect for the likelihood that a big extinction risk will become salient when research figures out how to harness a new kind of power.
As much as it maybe ruins the fun for me to just point out the message: the major point of the story was that you weren’t supposed to condition on us knowing that nuclear weapons are real, and instead ask whether the Gradualist or Catastrophist’s arguments actually make sense given what they knew.
That’s the situation I think we’re in with Fast AI Takeoff. We’re trying to interpret what the existence of general intelligences like humans (the Sun) implies for future progress on ML algorithms (normal explosives), without either a clear underlying theory for what the Sun’s power really is, or any direct evidence that there’ll be a jump.
That remark about the ‘micro-foundational explanation for why the sun looks qualitatively new but really isn’t’ refers to Richard Ngo’s explanation of why humans are so much better than chimps: https://www.lesswrong.com/s/n945eovrA3oDueqtq/p/gf9hhmSvpZfyfS34B#13_1__Alignment_difficulty_debate__Richard_Ngo_s_case
In my opinion the relevant detail is that we were not able to prevent the Soviets from getting the bomb. It took them all of about 3 years. It’ll take China, Russia, open source hackers et al about 18 months max to replicate AGI once it arrives. So much for your decisive strategic advantage.
Nuclear technology is hours away from reducing the value of all human civilization by 10%, and for all we knew that figure could have been 100%. That’s the nuclear threat. I wouldn’t even classify that as a “geopolitical” threat. The fact that Soviet nuclear technology pretty quickly became comparable to US nuclear technology isn’t the most salient fact in the story. The story is that research got really close, and is still really close, to releasing hell, and the door to hell looks generally pretty easy to open.
18 months is more than enough to get a DSA if AGI turns out anything we fear (that is, something really powerful and difficult to control, probably arriving fast at such state through an intelligence explosion).
In fact, I’d even argue 18 days might be enough. AI is already beginning to solve protein folding (Alphafold). If it progresses from there and builds a nanosystem, that’s more than enough to get a DSA aka take over the world. We currently see AIs like MuZero learning in hours what would take a lifetime for a human to learn, so it wouldn’t surprise me an advanced AI solving advanced nanotech in a few days.
Whether the first AGI will be aligned or not is way more concerning. Not because who gets there first isn’t also extremely important. Only because getting there first is the “easy” part.
I don’t really think advanced AI can be compared to atomic bombs. The former is a way more explosive technology, pun intended.