Alpha Zero
What does it mean? Is it a big step or a small one? How could we know?Is Artificial Intelligence becoming a massive anticlimax? Is it inevitable that AI will be better than us?
For more information on Aleph Insights visit our website https://alephinsights.com or to get in touch about our podcast email podcast@alephinsights.com
Transcript
Hello and welcome to the Cognitive Engineering podcast produced by Tell Me Studios for Aleph Insights. In this series of podcasts, we take a look at interesting topics and discuss what we think they tell us about analysis and decision making. I'm Fraser McGruer, and I'm here with Peter Coghill and Nick Hare of Aleph Insights. And this week, we're discussing AlphaZero.
Speaker B: th of December: Speaker A:Sorry to interrupt, is it designed by the same people?
Speaker B:Yeah, it's DeepMind. So Google DeepMind. And this is a sort of more general version that allegedly can play not only Go, but also chess and shogi, which is a Japanese version of chess. And so what Google claim is that they set up the neural network and it didn't give it any domain knowledge. So they didn't feed in any data about the results of chess. They simply sort of put it on, you know, gave it the rules of chess essentially, and got it to more or less play itself and learn what the best moves were that way, without relying on any kind of domain knowledge, no opening books, no chess theory, none of that. Just, you know, here are the rules of chess. Now work out how to play it. And in apparently nine hours, it had learned to play so well that it was able to beat the best other chess artificial intelligence, which is Stockfish. In 100 matches, it got 28 wins and 72 draws. And no, Stockfish didn't win a single match. Now Stockfish is the one that the professionals use, you know, to train on. It's better than any human. So, you know, this is this amazing result really, that in less than a day, it went from nothing to being able to beat the leading chess artificial intelligence out there. So anyway, I'd say it's kind of, there's a bit of controversy. They haven't released all the details of what they did, of exactly how it worked, of whether or not it was a fair fight, because, you know, was Stockfish running on a sort of effectively a laptop, whereas AlphaZero is running on, you know, absolutely high-end processors and stuff. There's lots of things people don't know. But I mean, look, the point is that something like this is going to happen eventually. And I think I was kind of interested in a few things. First of all, you know, does it mean that we're a step closer to an artificial general intelligence? But also, this sort of interesting phenomenon, really, which I think is more of a psychological thing about our reaction to artificial intelligence, which is this almost sense of kind of anticlimax and despondency you get when you suddenly realise that, you know, thousands of years of chess, of chess expertise has been kind of more or less chucked out the window in an afternoon. Yeah. And, and I think there isn't that that's something which I think we're going to see more and more in a range of different domains.
Speaker A:So let's turn to I'm sure we'll come back to you. But let's turn to Peter in answering those two questions. What was the first question?
Speaker C:Are we a step closer to generalised artificial intelligence?
Speaker A:And the second question was, oh, isn't it all a bit underwhelming or disappointing or something like that? So have a cracker. Perhaps the first of those, Peter?
Speaker C:Yeah, well, yeah, well, this is this is promising. This is interesting, because it's the same system. And it's not specifically designed for any particular game that once given the rules of the game and enough time, will learn it and exceed human and the best performing artificial intelligence in it. So in that way, yeah, it does sound like it's generalisable. But yeah, without the without the full release of the code and the experiments that they've conducted, it's difficult to know. But yeah, so it sounds like a small step towards general generalised artificial intelligence. But as we have discussed before, general AI is a long way away. It's a really big thing. That is a really big thing.
Speaker A:Sorry, for a layman, what is talk to me about general AI? What does that mean? I don't know if all our listeners will understand what that means.
Speaker C:So we are general intelligences. Humans are generally we were able to apply our learning skills, assimilate data of all different types, all sorts of different problems. So if you took a human player and trained them intensively for 24 hours, they would get better. But they would likewise get better at pottery, or playing the piano or any number of any other tasks that we've got very generally applied learning skills. Whereas artificial intelligence to date is very specific tasks, tasks specific is optimised and is only capable of learning about one specific.
Speaker A:So for example, that's why it's a step up from AlphaGo because AlphaGo is specifically for Go. But what didn't you say that this was a general kind of intelligence that was that was they pointed it at chess?
Speaker C:Yeah, so it seems more general because it was able to learn three different games. So it learned Go, chess and shogi. But they are still quite, quite narrow problems in that you can define the rules of these games as a set of quite simple instructions. So you know how the knight moves, how the queen moves, what good looks like.
Speaker B:Yeah, it's perfect information as well, which is also, you know, very different from a lot of real world situations where you know the state of the board is at any given time. It's sequential as well. So people, you know, you take a move, then the other player takes a move. The real world, of course, isn't like that.
Speaker A:And so we've heard from Peter talking about, you know, his response to that initial question of well, what does this mean? And I think your answer is, yeah, it's a step.
Speaker C:It's a step, but a small step on a very long road to a general artificial intelligence.
Speaker A:Nick, anything to add to that?
Speaker B:Yeah, just actually, I'm not sure we know how long the road is in that, you know, we still haven't really worked out what to do to create a general artificial intelligence. There isn't consensus on what we need to do. So we don't know, for example, whether it's a matter of, well, Stuart Armstrong from the Future of Humanity Institute, they look at this kind of issue pretty much all the time. And he says there's a distinction really between tasks which are grind based, you know, where you can say, well, we've got to work on this thing for exactly 20 years and then we'll have done everything versus things that are insight based. And the problem with what an artificial general intelligence would look like is that we don't really know which of those things it is. Now, it might be just an accumulation of lots of narrow artificial intelligences. It might be that we can just sort of solve one problem at a time. You know, once we've done chess, you know, and we've done driving, then we can go from driving to making a cup of tea. And once we can do that, we can get it to go shopping for us. And eventually we've got something which pretty much behaves like a human or better than a human. But maybe not. It might be that the real world poses so many problems that it's not simply a matter of scaling things up. But we need a whole new different approach. And the thing is, we don't know. And we won't really know until we're in a position technologically to try those things out. And at the moment, we're a long way off being able to do that. Well, I mean, we're a long way off being able to have something which is capable, really, of acting in a general way in the real world. In other words, assimilating the kinds of data that we do and processing it in the way that we do. But that may be something which we'll sort of tackle quite soon. So the thing is that we don't know. This may be a big step, or it may be a small step. Or it may prove that what they've stumbled upon, or perhaps to be kind of what they've invented, is an approach which actually could just be scaled up. There might be that you take 100 of these things, you point them at a problem like doing the washing up or some other really hard problem. They will solve that as well. And maybe even come up with an exciting, innovative way of doing it.
Speaker A:Okay.
Speaker B:Sorry, probably worth saying, Peter touched on it there. I think the one really big difference, the biggest challenge probably, apart from the technological bits, the sort of trying to work out how we're going to process data and collect it, is actually how we specify what the objective is. Now, the really easy thing about chess is you know what winning looks like. It's precisely specified which situations are winning and which aren't. The problem with the real world, that's where a lot of the fears and risks about artificial intelligence come from, is that it's really hard to say what we want it to do. So if I said, look, I want you to do the washing up, we as humans all understand really well what that is. But a machine, you know, from scratch actually doesn't know which states of the world does having done the washing up correspond to? And which states of the world must it avoid during that? So, you know, is it okay to do the washing up in such a way that causes massive environmental degradation? Probably not. Is it okay to do the washing up if you're killing people on the way or causing a nuclear war or turning the entire world into a massive computer? You know, we instinctively think, well, that's obvious that you don't do that. But to a machine, it's not obvious unless you've said that. You know, if all you say is I want the washing up to be done, there's no guarantee that it'll take what we would consider to be an easy route. Yeah, I mean. So that's that's a big challenge is how do we specify the objectives for an artificial intelligence in the real world?
Speaker A:I mean, also, I don't start to get a little bit prosaic, but often seems I don't really want to go down this avenue. But it seems to me the barriers are often just physical ones as well, rather than the programming side of it.
Speaker B:I mean, like walking and picking things up. Yeah, yeah.
Speaker A:Yeah. Like, you know, yeah. Picking up dishes. Yeah.
Speaker B:I mean, what we used to. Well, that's another interesting thing. We're learning is kind of gradually invading. I mean, in the old days, we used to sort of think, well, what we need to do is precisely specify how to walk. We're going to get a pair of legs, a pair of robotic legs and tell it how to walk. And these days, the approach is much more. Well, we will accept that we don't really know. We find it hard to specify that. So we'll just get we'll simulate a pair of legs and we'll get it to we'll get it to try out thousands and millions of different ways of walking until it finds one that works. So the learning really is. That's why learning is so such a powerful approach compared to having artificial intelligences that you program. You don't really have to. You can just say, well, I want you to get from A to B as quickly as possible using this pair of robotic legs and it will learn that. So, you know, the suggestion is, well, we found it hard up till now. But then we found creating artificial intelligences for chess hard up to now. Well, now suddenly it seems like it's easier. Well, if it turns out that we can apply a similar approach to learning how to walk or pick things up, then then, you know, maybe maybe that's not a problem anymore.
Speaker A:OK, I want to turn to this second question of. Oh, great. It's managed to do that and just feeling slightly underwhelmed or. Yeah, I mean, I'm not sure what my question is, but do you want to wax lyrical on that, Peter?
Speaker C:Yeah, well, the the the way that the press the press found it quite exciting, as often they do. And there's been there's been a bit of a backlash. People saying, oh, well, you know, the Alpha Zero was in a better position, et cetera, et cetera. But to me, it sounds like they've take they've apart from this sort of slight generalization. It's still it still sounds like a bit of an engineering challenge. It feels like a grind problem to me because they've they've no doubt they've got some new insights and new new approaches. But they've thrown a lot of processing at a problem that was quite well understood. And it's sort of underwhelming because, you know, computers have been beating humans for a long time at chess. So why is this saying it? Why is this new? But what I think when you when you scratch the surface a bit, there is a this generalization thing is a unique insight that's new.
Speaker B: which beat Garry Kasparov in: Speaker C:Maybe because we know for, you know, being what we are, how hard it is to learn another language, how hard it is to get good at Czechs. And so it's sort of if something comes along, you know, the next generation of children who are brighter and more fit and active than you are. You feel there's a remorse about that sort of generational kind of remorse.
Speaker B:Yeah, like, you know, do we need to tell people to learn? I mean, I'm rubbish at languages. I just really hate the idea that I have to sit down and learn a whole nother language. Really bothered me. I mean, in a way, I kind of welcome it that, you know, we won't need, you know, we won't in a way need to learn foreign languages. We can just, we'll just have, you know, be able to translate in real time through our phone or whatever. Or through a fish that we can put in our ears. Yeah, indeed. I mean, at the very least, the benefits of learning a foreign language are getting lower and lower. And, you know, if you look ahead and you ask, well, what does it mean when we've got machines writing poems and writing articles and, you know, playing musical instruments and doing all of these things? Will we want, you know, to do that anymore? Will it be? I mean, what's going to be left for us? There's a sense of all of these great achievements being swallowed up by technology, which is a bit depressing.
Speaker A:It is actually now. Yeah. I mean, you're quite right. I mean, let's say imagine well known right now. Let's imagine someone who can't really speak any, who doesn't know any languages, who doesn't know how to map read, who can't play any instruments and can't play chess because they don't need to have done any of those things or not wanted to. There's no point because there's something out there that can just do it immediately. What kind of person is that? It's quite an anthropomorphic way to look at it. Anthropomorphic, is that the right way to put it? Anthropocentric. Anthropocentric way to look at it. Because maybe, yeah, we're asking the wrong question because us being humans.
Speaker C:Maybe it opens up opportunities to go and do other inventive things that machines can't do. What are they? That's what I'm worried about. Yeah, exactly. What they can't do.
Speaker B:I mean, you know, and it's not, this is a purely psychological point. I accept in a way and in some sort of sense, well, in a fairly obvious sense, the world will be better when you don't need to have done any of those things. When you get all the benefits from learning French and music and, you know, chess, you get all the benefits of those, you know, in the sense that you can play chess like a grandmaster thanks to this app on your phone. You don't need to speak French. You can still converse with French people. You get all the benefits without any of the costs. And in one very obvious sense, that's great. But I wonder if, you know, in a sort of psychological sense, whether we will suffer, there's a kind of stress of being outperformed by things, of feeling irrelevant, feeling that your skills are pointless. And, you know, I think it's something we may have to face and worry about.
Speaker A:Well, I mean, sorry, but we kind of have to wrap up. But before we do, I mean, just responding to that, imagine if you're a grandmaster. I'm sure there's been all sorts of responses from chess grandmasters to this AlphaZero development. But also, first of all, if you're a grandmaster for probably quite a long time now, you might be thinking, well, what's the point? Or people have been asking, well, what's the point? Because there's machines that can always beat you. But also, secondly, given one of the things you said about the AlphaZero doing surprising moves or whether it be AlphaGo doing surprising moves, presumably chess grandmasters and Go grandmasters are looking and studying and analyzing what the AI was doing and going, ah, OK, and learning from that. Peter, this has got to be our last point. Peter?
Speaker C:I think there's an analog. So we've had recorded music for nearly 100 years now and definitely commonplace in the household for 30, 40 years in most people's homes. And yet still people still choose to play instruments because it's fun to challenge yourself and to learn. So I don't think that I think there's still a place for human. But it becomes rather than a necessity. It becomes a personal challenge thing. Right. And deriving pleasure from that.
Speaker B:Yeah. Nick? Yeah. No, it's just I don't think this is quite the same thing. Recorded music, live music in the home is very different. I mean, it's very different in terms of performing it and the sound and everything else than recorded music. But I think, you know, just to take that point, even if it even if it was the same, you know, first of all, there are far fewer people who play musical instruments. Certainly, you know, every home had a piano in it 100 years ago. That's no longer true now. And secondly, it's kind of it's a bit of a sad vision for doing things simply because, you know, humans can do them. It's a bit like being in a zoo. See these see these human chess players. Isn't it amazing? They can play chess to nearly 1% as well as, you know, this this app on my phone. It's sort of sad. It reduces something that was a pinnacle of achievement to something which is now a sort of rather intriguing foothill.
Speaker A:We need to stop there. But just just we do just reminds me of as ever, you know, in Star Trek, the next generation where Data is, of course, is an excellent virtuoso violin player. But his shipmates just actually don't like listening to what he does just because it's a little bit too perfect. And it's sort of. Yeah. And it's it's it just sounds like recorded music. Anyway, that's that side of things. But actually, one of the things you said, I want to leave the last words to Paul McCartney, because just to round it back, you know, where is this a big jump? Is this a big leap? What is the end final destination? All I'll say or Paul would say is it's a long and winding road that leads us to who knows where. Thank you. Right. Yeah. OK.
Speaker B:Possibly a really short road that will lead us to somewhere predictable. But, you know, a road of indeterminate length leading us to an indeterminate destination.
Speaker A:Beautifully put, poet. Oh, Nick. OK, we'll stop there. Thank you, as always, to listening to the Cognitive Engineering podcast. I'm Fraser McGruer. We've been here with Nick Hare and Peter Coghill of Aleph Insights. And until next time, goodbye.
