字幕列表 影片播放 列印英文字幕 I'm going to ask a question, but you can only answer by saying either 'yes,' 'no,' or 'it's complicated.' Alright? So, let's start over here. Is some form of superintelligence possible, Jaan? 'Yes,' 'no,' or 'it's complicated.' Yes. Yes. Yes. Yes. Yes. Yes. Definitely. No. Well, this was disappointing, we didn't find any disagreement. Let's try harder. Just because it's possible doesn't mean that it's actually going to happen. So, before I asked if superintelligence was possible at all according to the laws of physics. Now, i'm asking, will it actually happen? A little bit complicated, but yes. Yes, and if it doesn't then something terrible has happened to prevent it. Yes. Probably. Yes. Yes. Yes. Yes. No. Shucks, still haven't found any interesting disagreements. We need to try harder still. OK. So, you think it is going to happen, but would you actually like it to happen at some point? Yes, no, or it's complicated? Complicated, leaning towards yes. It's complicated. Yes. Yes. It's really complicated. Yes. It's complicated. Very complicated. Well, heck, I don't know. It depends on which kind. Alright, so it's getting a little bit more interesting. When I think, we had a really fascinating... When is it going to happen? Well, we had a really fascinating discussion already in this morning's panel about when we might get to human level AI. So, that would sort of put a lower bound. In the interest of time, I think we don't need to rehash the question of when going beyond it might start. But, let's ask a very related question to the one that just came up here. Mainly, the question of well if something starts to happen, if you get some sort of recursive self improvement or some other process whereby intelligence and machines start to take off very very rapidly, there is always a timescale associated with this. And there I hope we can finally find some real serious disagreements to argue about here. Some people have been envisioning this scenario where it goes PHOOM and things happen in days or hours or less. Whereas, others envision that it will happen but it might take thousands of years or decades. So, if you think of some sort of doubling time, some sort of rough timescale on which things get dramatically better, what time scale would you guess at, Jaan? Start now or starting at human level? No, no, so once we get human level AI or whatever point beyond there or a little bit before there where things actually start taking off, what is the sort of time scale? Any explosion, as a nerdy physicist, has some sort of time scale, right, on which it happens. Are we talking about seconds, or years, or millennia? I'm thinking of years, but it is important to act as if this timeline was shorter. Yeah, I actually don't really trust my intuitions here. I have intuitions that we are thinking of years, but I also think human level AI is a mirage. It is suddenly going to be better than human, but whether that is going to be a full intelligence explosion quickly, I don't know. I think it partly depends on the architecture that ends up delivering human level AI. So, this kind of neuroscience inspired AI that we seem to be building at the moment that needs to be trained and have experience in order for it to gain knowledge that may be, you know, on the order of a few years so possible even a decade. Some numbers of years, but it could also be much less. But, I wouldn't be surprised if it was much more. Potentially days or shorter, especially if it's AI researchers designing AI researchers Every time there is an advance in AI, we dismiss it as 'oh, well that's not really AI:' chess, go, self-driving cars. An AI, as you know, is the field of things we haven't done yet. That will continue when we actually reach AGI. There will be lots of controversy. By the time the controversy settles down, we will realize that it's been around for a few years. Yeah, so I think we will go beyond human level capabilities in many different areas, but not in all at the same time. So, it will be an uneven process where some areas will be far advanced very soon, already to some extent and other might take much longer. What Bart said. But, I think if it reaches a threshold where it's as smart as the smartest most inventive human then, I mean, it really could be only a matter of days before it's smarter than the sum of humanity. So, here we saw quite an interesting range of answers. And this, I find is a very interesting question because for reasons that people here have published a lot of interesting papers about the time scale makes a huge difference. Right, if it's something that happening on the time scale of the industrial revolution, for example, that's a lot longer than the time scale on which society can adapt and take measures to steer development, borrowing your nice rocket metaphor, Jaan. Whereas, if things happen much quicker than society can respond, it's much harder to steer and you kind of have to hope that you've built in a good steering in advance. So, for example in nuclear reactors, we nerdy physicists like to stick graphite sticks in a moderators to slow things down maybe prevent it from going critical. I'm curious if anyone of you feels that it would be nice if this growth of intelligence which you are generally excited about, with some caveats, if any of you would like to have it happen a bit slower so that it becomes easier for society to keep shaping it the way we want it. And, if so, and here's a tough question, is there anything we can do now or later on when it gets closer that might make this intelligence explosion or rapid growth of intelligence simply proceed slower so we can have more influence over it. Does anyone want to take a swing at this? It's not for the whole panel, but anyone who... I'm reminded of the conversation we had with Rich Sutton in Puerto Rico. Like, we had a lot of disagreements, but definitely could agree about paths slower being better than faster. Any thoughts on how one could make it a little bit slower? I mean, the strategy I suggested in my talk was somewhat tongue and cheek. But, it was also serious. I think this conference is great and as technologists we should do everything we can to keep the technology safe and beneficial. Certainly, as we do each specific application, like self-driving cars, there's a whole host of ethical issues to address. But, I don't think we can solve the problem just technologically. Imagine that we've done our job perfectly and we've created the most safe, beneficial AI possible, but we've let the political system become totalitarian and evil, either a evil world government or even just a portion of the globe that is that way, it's not going to work out well. And so, part of the struggle is in the area of politics and social policy to have the world reflect the values want to achieve because we are talking about human AI. Human AI is by definition at human levels and therefore is human. And so, the issue of how we make humans ethical is the same issue as how we make AIs that are human level ethical. So, what i'm hearing you say is that before we reach the point of getting close to human level AI, a very good way to prepare for that is for us humans in our human societies to try and get our act together as much as possible and have the world run with more reason than it, perhaps, is today. Is that fair? That's exactly what I'm saying. Nick? Also, I just want to clarify again that when I asked about what you would do to slow things down i'm not talking at all about slowing down AI research now. We're simply talking about if we get to the point where we are getting very near human level AI and think we might get some very fast development, how could one slow that part down? So, one method would be to make faster progress now, so you get to that point sooner when hardware is less developed, you get less hardware overhang. However, the current speed of AI progress is a fairly hard variable to change very much because there are very big forces pushing on it, so perhaps the higher elasticity option is what I suggested in the talk to ensure that whoever gets there first has enough of a lead that they are able to slow down for a few months, let us say, to go slow during the transition. So, I think one thing you can do, I mean this is almost in the verification area, is to make systems that provably will not recruit additional hardware or resigned their hardware, so that their resources remain fixed. And i'm quite happy to sit there for several years thinking hard about what the next step would be to take. But, it's trivial to copy software. Software is self replicating and always has been and I don't see how you can possibly stop that. I mean, I think it would be great if it went slow, but it's hard to see how it does go slow given the huge first mover advantages and getting to superintelligence. The only scenario that I see where it might go slow is where there is only one potential first mover that can then stop and think. So, maybe that speaks to creating a society where, you know, AI is restrictive and unified, but without multiple movers. Yeah, Demis, so your colleague Sean Legg mentioned that the one thing that could help a lot here is if there's things like this industry partnership and a sense of trust and openness between the leaders, so that if there is a point where one wants to... Yeah, I do worry about, you know, that sort of scenario where, you know, I think, I've got quite high belief in human ingenuity to solve these problems given enough time. the control problem and other issues. They're very difficult, but I think we can solve them. The problem is will there, you know, the coordination problem of making sure there is enough time to slow down at the end and, you know, let Stuart think about this for 5 years. But, what about, he may do that, but what about all the other teams that are reading the papers and not going to do that while you're thinking. Yeah, this is what I worry about a lot. It seems like that coordination problem is quite difficult. But, I think as the first step, may be coordinating things like the Partnership on AI, you know, the most capable teams together to agree, at least agree on a set of protocols or safety procedures, or things, you know, agree that, maybe, you know, you should verify these systems and that is going to take a few years and you should think about that. I think that would be a good thing. I just want to caveat one thing about slowing versus fast progresses, you know, it could be that, imagine there was a moratorium on AI research for 50 years, but hardware continued to accelerate as it does now. We could, you know, this is sort of what Nick's point was is that there could be a massive hardware overhang or something where an AI actually many, many, many different approaches to AI including seed AI, self-improving AI, all these things could be possible. And, you know, maybe one person in their garage could do it. And I think that would be a lot more difficult to coordinate that kind of situation, whereas, so, I think there is some argument to be made where you want to make fast progress when we are at the very hard point of the 's' curve. Where actually, you know, you need quite a large team, you have to be quite visible, you know who the other people are, and, you know, in a sense society can keep tabs on who the major players are and what they're up to. Whereas, opposed to a scenario where in say 50 or a 100 years time when, you know, someone, a kid in their garage could create a seed AI or something like that. Yeah, Bart, one last comment on this topic. Yeah, I think that this process will be a very irregular process and sometime we will be far advanced and other times we will be going quite slow. Yeah, i'm sort of hoping that when society sees something like fake video creation where you create a video where you have somebody say made up things and that society will actually realize that there are these new capabilities for the machines and we should start taking the problem as a society more seriously before we have full and general AI. We'll use AI to detect that. So, you mentioned the word 'worry' there, and you Nick went a bit farther, you had the word 'doom' written on your slides three times. No wonder there was one star on Amazon on that rating and that it was even in red color. I think it's just as important to talk about existential hope and the fantastic upside as downside and I want to do a lot of that here. So, let's just get some of those worries out of the way now and then return to the positive things. I just want to go through quickly and give each one of you a chance to just pick one thing that you feel is a challenge that we should overcome and then say something about what you feel is the best thing we can do, right now, to try to mitigate it. Do you want to start Jaan? To mitigate what? Mention one thing that you're worried could go wrong and tell us about something constructive that we can do now that will reduce that risk. I do think that AI arms races, I see like a lot of, like, good. I'm really heartened to see kind of great contacts between OpenAI and DeepMind, but I think this is just like a sort of toy model of what the world at large might come up with in terms of arms races. And for myself I have been spending increasing amount of time in Asia recently just to kind of try to kind of pull in more people elsewhere, what has been so far, just been, kind of like, an Anglo- American discussion mostly. So, like this is, I think, this is one thing that should be done and i'm going to do it. Well, as someone who is outside this field, I think the challenge i'm really in touch with is how hard it is to take the safety concerns emotionally seriously. And how hard it is for people in the field to do that as well. I can't tell you have many people outside this room who purport to be experts think the control problem is a total non-issue. I mean, it's just flabbergasting to meet these people and just therefore not worth thinking about.