Garry Kasparov: IBM Deep Blue, AlphaZero, and the Limits of AI in Open Systems | AI Podcast Clips

You're lost to IBM D Blue in 1997, and I apologize for being romanticizing the notion, but in the history of our civilization, this is with everything you.

Said, I, I agree with probably one exception is that considering chess, you know, as the sort of, as a pinnacle of intellectual activities was our mistake because, you know, we just thought, oh, it's a, it's a game of the highest intellect. And it's just, you know, you have to be so, you know, anyone who.

Cares about science, innovation, an entire generation of AI researchers. And yet to you, that loss, at least if reading your face was, seemed like a tragedy, extremely painful, like you said, physically painful. Why, when you look back at your psychology of that loss, why was it so painful? Were you not able to see the seminal nature of that moment? Or was that exactly why it was that painful?

As I already said, losing was painful, physically painful. And the match I lost in 1997 was not the first match I lost to a machine. It was the first match I lost. Period. Yes. That's. Oh, wow.

Oh, wow. Yeah, it's right.

Yeah, that makes all the difference to me.

Yes.

First time I lost, it's just now I lost. And the reason I was so angry that I just, you know, I had suspicions that my loss was not just a result of my bad play.

Yes.

So though I played quite poorly, you know, just when you started looking at the games today, I made tons of mistakes. But, you know, I, I had all reasons to believe that, you know, there were other, other factors that had nothing to do with the game of chess and that's why I was angry. But look, it was 22 years ago. It's what, under the bridge we can analyze this match. And this is with everything you said, I, I agree with probably one exception is that considering chess, you know, as the sort of, as a pinnacle of intellectual activities was our mistake because, you know, we just thought, oh, it's a game of the highest intellect. And it's just, you know, you have to be so, you know, intelligent and you could see things that, you know, the ordinary mortals could not see. It's a game. And all machines had to do in this game is just to make fewer mistakes, not to solve the game because the game cannot be solved. I mean, according to Kuala Shannon, the number of legal moves is 10 to the 46 power. Too many zeros just for any computer to finish the job, you know, in next few billion years. But it doesn't have to. It's all about making fewer mistakes. And I think that's this Match actually and what's happened afterwards with other games, with Go, with Shogi, with video games, it's a demonstration that it's. The machines will always beat humans in what I call closed systems. The moment you build a closed system, no matter how the system is called Chess Go, Shoggy, Dota machines will prevail simply because they will bring down number of mistakes. Machines don't have to solve it, they just have to. The way they outplay us, it's not by just being more intelligent, it's just by doing something else. But eventually it's capitalizing on our mistakes. When you look at the chess machines ratings today and compare this to Magnus Carlsen, it's the same as comparing Ferrari to Usain Bolt. The gap is, I mean by chess standards is insane. 34 3,500 to 2,828 50 on Magnus. It's like difference between Magnus and an ordinary player from an open international tournament. It's not because machine understands better than Magnus Carlsen, but simply because it's steady, machine has steady hand. And I think that is what we have to learn from 1997 experience and from further encounters with computers and sort of the current state of affairs with AlphaZero beating other machines. The idea that we can compete with computers in so called intellectual fields.

It.

Was wrong from the very beginning. It's just, it's. By the way, the 1997 match was not the first victory of machines over.

Or grandmasters over Grandmasters.

Yeah, no, actually it's. I played against first decent chess computers from late, from late 80s. So I played with the prototype of Deep Blue called deep thought in 1989. Two rapid chess games in New York. I won handily. Both games we played against new chess engines like Fritz and other programs. And then it's the was Israeli program Junior that appeared in 1985. Yeah. So there, there were, there were several programs. I, you know, I lost few games in blitz. I lost one match against the computer chess engine 1994 rapid chess. So I lost one game to D Blue in 1996 match the man the match I won. Some people, you know, tend to forget about it that I won the first match. Yes. But it's, it's. We, we made a very important psychological mistake thinking that the reason we lost blitz matches, five, five minutes games. The reason we lost, lost some of the rapid chess matches 25 minutes just because we didn't have enough time. If you play a longer match, we will not make the same mistakes. Nonsense. So this. Yeah, we had more time but we still make mistakes. And machine also has more time. And machines, machine will always, you know, will always be steady and consistent compared to humans instabilities and inconsistencies. And today we are at a point where, yes, nobody talks about humans playing as machines. Now machines can offer handicap to top players and still will be favoring. I think we're just learning that it's no longer human versus machines. It's about human working with machines. That's what I recognized in 1998, just after licking my wounds and spending one year and just ruminating. So the. So what's happened in this match? And I knew that though we still could play against the machines, I had two more matches in 2003, playing both Deep Fritz and Deep Junior. Both matches ended as a tie. Though these machines were not weaker, at least actually probably stronger than Deep Blue. And by the way, today, chess app on your mobile phone is probably stronger than Deep Blue and Deep Blue. I'm not speaking about chess engines that are so much superior. And by the way, when you analyze games we played against DB 1997 on your chess engine, they'll be laughing. Yeah. So this is. And it also shows that's how chess changed. Because chess commentators, they look at some of our games like Game four, Game five. Brilliant idea. Now you ask Stockfish, you ask Houdini, you ask Commodore, all the leading chess engines. Yeah. Within 30 seconds, they will show you how many mistakes both Gary and D. Blue made in the game that was trumpeted as the. As a great Chess match in 1997.

Well, okay, so you've made an interesting. If you can untangle that comment. So now in retrospect, it was a mistake to see chess as the peak of human intellect. Nevertheless, that was done for centuries.

So by the way, in Europe, because, you know, you move to the Far east, they will go.

Games, games again.

Some of the games like, you know, board games. Yes, yes. Yeah, I agree.

So if I push back a little bit. So now you say that. Okay, but it was a mistake to see chess as the epitome. And now, and then now there's other things maybe like language, that conversation, like some of the things that in your view is still way out of reach of computers, but inside humans, do you think, can you talk about what those things might be and do you think just like chess, that might fall soon with the same set of approaches, if you look at Alphazero, the same kind of learning approaches as the machines grow in size?

No, no, it's not about growing in size. It's about Again it's about understanding the difference between closed system and open ended system.

So you think that key difference so the board games are closed in terms of the rules set, the actions, the state space. Everything is just constrained. You think once you open it the machines are lost.

Not lost but again the effectiveness is very different because machine does not understand the moment it's reaching territory of diminishing returns. It's the to put it in a different way, machine doesn't know it's just constrained.

You think once you open it the machines are lost.

Not lost but again the effectiveness is very different because machine does not understand the moment it's reaching territory of diminishing returns. It's the to put it in a different way machine it doesn't know how to ask right questions. It can ask questions, but it will never tell you which questions are relevant. So this the it's like about the. It's the. It's a direction. So these it's. I think it's in human machine relations we have to consider how to ask right questions. It can ask questions but it will never tell you which questions are relevant. So there's the. It's like about the it's the it's a direction. So these it's I think it's in human machine relations we have to consider so our role and many people feel uncomfortable that the territories are so our role and many people feel uncomfortable that the territory that belongs to us is shrinking. I'm saying so what eventually will belong to the territory that belongs to us is shrinking. Last few decimal points. But it's like having a very powerful gun that's and all you can do there is slightly outward direction of the bullet. Maybe you know, 0.1 degree of this angle but that means a mile away, 10 meters of target. So that's. We have to recognize that is a certain unique human qualities that machines in a foreseeable future will not be able to reproduce. And the effectiveness of this collaboration depends on our understanding what exactly we can bring into the game. So the greatest danger is when we try to interfere with machine superior knowledge. So that's why I always say that sometimes you'd rather have by reading these pictures in radiology, you may probably prefer an experienced nurse rather than having top professor because she will not try to interfere with machines understanding. So it's very important to know that if machines knows how to do better things in 95%, 96% of territory we should not touch it because it happened. It's like in chess recognize they do it Better see where we can make the difference. You mentioned Alphazero. I mean alphazero, it's actually a first step into what you may call AI because everything that's being called AI today is just, it's, it's, it's one or another variation of what Claude Shannon characterized as a brute force is a type A machine. Whether it's Deep Blue, whether it's Watson, it's. And all these, these modern technologies that are being trumpeted as, as AI it's still brute force. It's the, all they do, it's. They do optimization, it's this, they are, you know, they, they keep improving the way to process human generated data. Now AlphaZero is the first step towards machine produced knowledge which is by the way is quite ironic way to process human generated data. Now Alpha, the first company that championed that was IBM 0 is the first step towards machine produced knowledge.

Yes.

Which is by the way, by the way it's quite ironic that the first company that championed that was IBM. Oh, it's in backgammon.

Interesting.

Yes, you just, you should, you should, you should look at IBM. It's a neurogammon. It's the, it's the scientist, he's still working at IBM they had in the early 90s. It's just, it's the, it's the program that played, you know, the Alphazero type. So just trying to come up with own strategies. But because of success of Deep Blue this project. Oh, it's in backgammon.

Interesting.

In backgammon had been not abandoned but just you know, it was put on cold and now just you know, it's, you know it's. Everybody talks about the machines. Yes, you should look at IBM. It's a neurogammon. It's the scientist called Cesaro, he's still working at IBM they had in the early 90s it's generated knowledge. So as revolutionary and it is. But there's still many open ended questions. Yes, Alphazero generates its own data. Many ideas that Alphazero generated in chess were quite intriguing. So I looked at these games with. Not just with interest but with you know, it's quite exciting to learn how machine could actually, you know, juggle all the pieces and just play positions with a broken material balance. Sacrificing material always being ahead of other programs. You know, one or two moves ahead by foreseeing the consequences not over calculating because machines is a program that played, you know, the alphazero type. So just trying to come up with own strategies but because of Success of Deep Blue this project other machines were at least as powerful in calculating but it's having this unique knowledge based on discovered patterns after playing 60 million games had been not abandoned but just you know, it was put on cold and now just, you know it's, you know, it's. Everybody talks about the machines almost something.

That feels like intuition.

Exactly. But there's one problem. Now the simple question if, if Alphazero faces superior point, let's say another powerful computer accompanied by a human who could help just to discover certain problems. Because I already I looked at many AlphaZero games, I visited their lab, spoke to Demis Kasabis at his team and I know there's certain weaknesses there. Now if these weaknesses are exposed, the question is how many games will it take for AlphaZero to correct it? The answer is hundreds of generated knowledge. So as revolutionary and it is but there's still many open ended questions, thousands even if it keeps losing because the whole system is based so now imagine Solar says you can have a human by just making a few tweaks. So humans are still more flexible. Yes, Alphazero generates its own data many and as long as we recognize what is our role where we can play sort of the most valuable part in this collaboration so it will help us to understand what are the next steps in human machine collaboration. Ideas at Alphazero generated in chess were quite intriguing. So I looked at these games with, not just with interest but with, you know, it was quite exciting to learn how machine could actually, you know, juggle all the pieces and just play positions with a broken material balance sacrificing material always being ahead of other programs, you know, one or two moves ahead by. By foreseeing the consequences not over calculating because machines, other machines were at least as powerful in calculating but it's having this unique knowledge based on discovered patterns after playing 60 million games.

Almost something that feels like intuition.

Exactly. But there's one problem. Now the simple question if alphazero faces superior point let's say another powerful computer accompanied by a human who could help just to discover certain problems because I already I look at many Alphazero games I visited their lab, spoke to Demis Kasabis and his team and I know there's certain weaknesses there. Now if these weaknesses are exposed, the question is how, how many games will it take for Alphazero to correct it? The answer is hundreds of thousands. Even if it keeps losing. It's this because the whole system is based. So it's Now Imagine Solar says you can have a human by just making a few tweaks. So humans are still more flexible. And as long as we recognize what is our role, where we can play sort of the most valuable part in this collaboration. So it will help us to understand what are the next steps in human machine collaboration.