Winning starts with what you know
The new version 18 offers completely new possibilities for chess training and analysis: playing style analysis, search for strategic themes, access to 6 billion Lichess games, player preparation by matching Lichess games, download Chess.com games with built-in API, built-in cloud engine and much more.
André Hollstein is an IU professor for information technology and digitisation. In addition he runs a podcast called DigitalXRoad ("digital cross road"), also availble on Spotify, for his students and for the general public.
Recently Prof. Hollstein recorded Episode 6 of his podcast, with Fredric Friedel, discussing the digitisation of chess. He writes:
Chess is one of the most played sports in the world and it has always fascinated people to develop programs or machines that are equal to or even better than humans in chess. Today there are chess programs that are clearly superior to the world chess champion (probability of winning 999:1). This is a sign of the innovative ability of this industry, which has major points of contact with topics such as "Artificial Intelligence (AI)", "BigData" or "Cloud Computing".
I discuss these aspects and more in the current edition of DigitalXRoad with Frederic Friedel, one of the computer chess pioneers in Germany, who revolutionized and democratized chess with his company Chessbase. He was part of Garry Kasparov's team against DeepBlue (which beat the world champion in 1997), and today works on AI-based chess programs.
Enjoy this episode. Reading Tips:
You can listen to the discussion on DigitalXRoad, or on Sportify. It was conducted in German, but Prof. Hollstein has kindly provided us with an English transcript:
AH: Hello and welcome to a new episode of DigitalXRoad. In today's episode we want to deal with a topic that certainly does not come to mind to everyone when talking about digitization or computers, namely the topic of computer chess. But computer chess is an innovation driver. There are many points of contact with topics such as AI, Big Data or Cloud Computing. To discuss this topic, I am really happy today that I was able to win Frederic Friedel over to my virtual studio. Frederic is undoubtedly one of the computer chess pioneers in Germany and revolutionized chess with his company Chessbase.
AH: Hello Frederic, nice that you found time to come to my virtual studio between your numerous book projects and many appointments.
FF: Hello.
AH: Frederic, you're certainly as well known in the chess scene as Garry Kasparow or Bobby Fischer, but probably not that many outside of that world know you, so a little bit about your career. You originally studied philosophy, philosophy of science and linguistics in Hamburg and Oxford, are a science journalist, produced documentaries for German television and later came to computer chess through chess documentaries. Can we put it that way?
FF: Yes, exactly.
AH: You then founded the company Chessbase a few years later, in the mid-1980s, which I think was quite remarkable, considering that the first PC came onto the market in 1981, and soon after you founded ChessBase, which developed a business model that revolutionized the sport of chess. Maybe you can tell us a bit about that time and also tell how it came about and how it developed.
FF: Yes, and I also want to emphasize that Garry Kasparov, who wasn't world champion yet, helped us immensely. At some point he came to Hamburg at the invitation of SPIEGEL and and visited me because I had sent him some computer games to Baku. We spent a week talking about computers. He comes to me, rings the bell, says "Okay, you're Frederik, I'm Garry, that's a nice house, this is your family. Now we're friends, now tell me everything you know about computers. I did that. And in return he told me everything he thought a computer should be able to do for chess. And so we the idea for a database. I wasn't a programmer, but luckily I found a young physics student who had developed a prototype database. We showed that to Garry. He's spoke to companies, said, "You have to support these people." And he told me we had to finish developing the program now. We did it. In 1986 he was the first user of Chessbase – #0001. And he helped us with advertising and so on. Never took a penny from us. I was no longer a friend but a "relative" of his and he helped us. And I find it unique that he had, like two or three other players in the world, a monopoly on chess knowledge – he actually had three grandmasters who helped him, and had access to libraries and so on, and he contributed to democratization of chess. Today everyone can get exactly what Magnus Carlsen has for his studies, and that means the whole world can prepare optimally.
AH: I read that you now have almost ten million games in your database and I think that's the much more exciting part, ten million high-quality games.
FF: We have about 100 million games played on servers and that's also quite interesting because we want to check or analyze how beginners play, not only grandmasters, international masters but also beginners.
AH: Exactly, but the exciting point is that most of these games are also commented on, Grandmasters sat down and evaluated these games and gave their comments on them. Which basically also helps to reflect and improve one's own game. But what basically also offers a kind of training opportunity in modern chess – and we'll come back to the subject again shortly – to recognize a computer move, just as well as that a human would do.
FF: There are about a hundred thousand analysed games, among the eight and a half million high-quality games, and they are very interesting because you can see what a human expert thinks about them. He played that, although apparently it's very dangerous and so on. Interesting because you can see what the computer itself thinks about it and you can evaluate that. Yes, he was wrong. Of course it wasn't dangerous because of this or that. You can correct what a person thinks or realizes spontaneously by looking at the position. The computer corrects or confirms it.
AH: It was or has always been fascinating, I think, to create machines for humans that are at least equal to humans, maybe even better than humans and that's why we already have them. In the past there were developments to develop chess computer programs, which today also grandmasters use for the analyses. Then there was a memorable time in the 1990s when Gary Kasparov, who was already world champion at the time, also played against DeepBlue. You were also part of his team at the time, on the side of the world chess champion, so to speak.
DeepBlue, the IBM developers, tried to create a computer program – with an extremely large amount of computing power – that was able to beat the world chess champion. What was it like back then from your perspective, was it more the computing power behind it or was the program developed by IBM so good that you could say the software, the technology behind it was better, simply superior to the world chess champion?
FF: The technology was ground-breaking, they programmed it in hardware and it ran much, many times faster than a normal program running on a giant mainframe. Gary played twice against DeepBlue, once in 1996 and he won, and then in 1997. And I think – and I saw it – they tricked him by psychological means. He was very nervous about being spied on, or that humans – grandmasters – might be helping the computer. And that wasn't the case, but they managed to get him paranoid. DeepBlue won, but it never played again. Of course, because that's what they wanted. Regarding playing strength I must say: Garry was stronger than DeepBlue. Today I have a mobile phone – an Android – and it is much more powerful than DeepBlue. Today he would not win against my mobile phone. And of course the normal computers have become infinitely more powerful.
AH: Just out of interest, was it clear to Garry at the time that this is going to happen eventually? After all, he worked with you guys from a for a long time. As you just mentioned, he developed ChessBase with you. Was he aware that this development will come at some point and that there will be computers that will be much better than humans. Was only a matter of time? Or was it surprising for him, even if they maybe tricked him a bit psychologically?
FF: It was surprising that it came so quickly. In 1991 we developed our Fritz program and there was a discussion and I had it with experts from all over the world. When will a computer beat the reigning world champion? And there were estimates like “next year” or “the year after next”. Some said 1995. Frederic Friedel said "in the year 2000". Some said in 2030 and others "never". And I have to say I think I was right because in 1997 when that happened it was too soon. He was still stronger and he demanded a rematch. But he didn’t get it from IBM. It was and is now that a computer is infinitely better than humans. I can hardly express it, sometimes I say 10,000 times. I don't know if you know the Elo numbers?
A grandmaster today has 2800, and that's sensational. It used to be 2700. A computer's Elo today is 3600, so I reckon the (current) World Champion Magnus Carlsen could draw one in ten games if he's lucky.
AH: I agree with that and that's a nice lead-in to the next topic, I think. This development is certainly not only due to the algorithms, but also to the developments on the side of artificial intelligence, i.e. AI (Artificial Intelligence). With AlphaZero, Google has managed to set up a chess program in a relatively short time, admittedly with a lot of computing power, which has the playing strength you just mentioned. And they did it essentially by giving the chess program the ability to play against itself without any kind of feedback from humans. That is, what is called reinforcement learning and with it a completely different quality of chess program than has been available up to now, isn't it?
FF: Actually our Androids etc. work with brute force programs – traditional brute force programs. For about 30 to 40 years that was the only option. You calculate everything – the most nonsensical moves. "If I play this, he plays this, I play this, he plays this, how about it? Oh, I lost two pieces, that's bad.” Brute force works with millions of positions, and with that they have become already much stronger than the world champion. The AI development is only a few years old, when Google, Demis Hassabis and his company (DeepMind), started. Demis had been a very strong chess player himself in his youth. They started building a neural network with this AI method, in which they let Google's hardware play eight hours or so, something like 40 million games, against itself. After a few hundred games it played abysmal badly. One side lost three pieces, but still won the game because the opponent lost four pieces senselessly. But after some 1000 games it was already recognizable as chess, after some million games it was strong. After 40 million games it became stronger than the strongest brute force program. So the world champion was left far behind. We reconstructed that, we developed a program called Fat Fritz. They provided their methods and the code – everything – and we reproduced it. The Google version never saw a human game, only the games against themselves. But we've also included our ChessBase Mega Database, as well as beginner's games so that it understands how to recognize or how to mate with "3 rooks" against the naked king. This never happens in master games. We reached a playing strength of about 3600, as I said. And now the brute force programs, which have also been further developed, are about the same – a little weaker than the neural networks. It's typical of AI: the thing is incredibly powerful, but can't tell you why. It doesn't even know if a queen is more valuable than a pawn, it just has some statistical evaluation of positions of this type and knows which move to play. Most chess players now use neural networks, which are slightly tactically inferior to brute force programs, but they find intuitive moves, strategic moves that brute force programs do not find. That's why all world championship contenders and the world champion himself use neural network programs. AI programs.
AH: Let's go back to what we discussed at the beginning – with the analysed games, does the chess engine then learn something and become even better as a result?
FF: This is our experiment. Google did it their way because they wanted to show that without any external knowledge an AI program can develop chess code which is superior. What we've done is to reproduce exactly that, plus human games to see if that does anything. I think yes. But very little of course, most of it is incomprehensible stuff that's figured out from over 20 million games.
AH: And what do you think, how will the development go on? We had spoken beforehand. From your perspective, development is also limited somewhere. If we stay with the Elo numbers, you probably won't get chess programs that have 5000 Elo points or even more, but it's sort of an approach to a limit value which is difficult to define now, you told me. How do you assess that?
FF: Yes, that is actually my prediction. We noticed during our development of Fat Fritz that if you still have a million... The question is why did Google, for example, have the AI play games eight hours or something, on huge hardware? Why not 18 hours or eight days or eight weeks? And reach infinite playing strength? Because they noticed, like we did that the curve is flattening. This means that if you play ten million more games and have them evaluated, then the increase in playing strength is no longer as great. And now my controversial opinion: No matter what we do, the playing strength will never increase further than 4000 Elo points. The question is why? Because chess is basically a draw. No matter what you do, you can always find a drawn position, and that means that a chess program that is 3900 Elo points will always find a way, or almost always find a way, to reach a drawn position against a 4000 Elo program. That means it will be able to defend itself, and then it will be infinitely difficult to win a game. And that's why it's going to stop at about 4000.
AH: Sounds like a very plausible theory.
Is it the case that what you have done in the AI environment is also a prime example for other topics that we will also encounter in everyday life in the future? There will certainly be further development in chess, but I think what you learned it will certainly be transferrable to other subjects, or do you see it differently?
FF: No, I see that the same way and I always invite people to come to Hamburg and see how we developed it, or to go to Google at all and see how it [works]. We are pioneers. We have experienced AI drastically, concretely and directly, just as the others are now beginning to experience it. AI will show similar developments in all areas in the future. With chess we were just the first. We were the first to experience this with brute force and now with neural networks, but it will happen everywhere, in law, in medicine, in business, in writing. I am a writer. I find that programs can write fairly well and are competitive. Maybe just like neural networks in chess, they will outperform us and write better and summarize better. But certainly in medicine and jurisprudence, i.e. in law, where similar phenomena will appear. That people will say: "The thing has absolutely no idea of justice, of human feelings and so on, and has only followed and studied a few 100 billion cases. It has no heart, no soul. But I want to go to the AI judge because he's much more balanced, fair, just, even if we don't know what he's doing. He judges that better than people who maybe know a few 1000 or 10,000 cases?”
AH: I also deal a lot with learning processes and the question of how people learn, for example. What we just discussed about AI also tempts us to think more intensively about how such AI-driven solutions can also help to improve the learning process. In chess maybe – but also in other contexts – do you see it that way too?
FF: Absolutely. It's like a human has a part of their brain reserved for chess, and the vast majority of people don't have that part. I know this from my own experience. I was a passionate chess player and actually quite good at it. But I realized that I could never become more than a weak International Master, even if I dedicated my life to chess. There's a small number of people who, as little kids, just absorb chess. They look at games and say, "Oh, I get that." Today it is the case that at the age of 11 they can become full grandmasters. This is a phenomenon that can hardly be described. It's like pointing to a child that is romping around in the garden, screaming, and saying: "That is Johann. He's just done his doctoral thesis in quantum physics and is working on the Very Large Hadron Collider CERN. They will say but you are pointing to a child! Yes, of course, that's him. That's the performance we see at 11 and 12 and 13 and 14. And I know dozens of boys like that, who already have full grandmaster titles. I think I know three that will go to the very top. Maybe four. And where does that come from? Not only because they watched how people play against each other, but they were able to see a hundreds of games on the computer. They absorb it like a sponge, the knowledge, the patterns. Incidentally, I believe that this part of chess resides in the right hemisphere, in very strong players, because they primarily work with pattern recognition. That's what the right brain does. Calculations of variants, that happens with the left brain, where mathematics and other things reside. Creativity comes from the right brain, where pattern recognition, speech recognition, image recognition takes place.
AH: Perhaps at the end, Frederic, another very beautiful story, I think, or even an amazing one. At the beginning you asked me what I think, when the first idea for a chess computer came about, and I was miserably wrong. You then explained to me that it was actually before the invention of the computer. And indeed Turing had already dealt with the idea, and maybe you can – you gave a nice lecture with Garry Kasparow at a Turing congress – maybe you can tell us something more about it. It's a very nice story.
FF: It's a question I often ask: "What do you think, when was the first chess program written, in relation to the invention of the computer?" And most people say ten years later, five years later (like you). No, the answer is two years before the first computer was built. At Bletchley Park (Alan) Turing cracked the German Enigma with a machine he built. He knew that freely programmable computers were coming. He was thinking how to do it, but he couldn't wait. He was a passionate chess player, but very bad. One of the smartest people in the world was a bad chess player. He built a paper machine. That means he formulated the algorithms, how a future computer should play, the evaluation criteria. And then he acted as the CPU. He played against an amateur, several games, but we have one. Hee calculated the moves himself – if I move the pawn like this, then the queen has so and so many possible moves, the king is so and so wecure, and so further. It took him 15 to 40 minutes to calculate a move. And when the first computers were created, other programs were then immediately implemented. Thus, chess is the very first application of AI in human history. Until then, for a year or two, computers, electronic programmable computers, had only been doing mathematical calculations. But the first thing that AI experimented with was chess. First Turing, and second in Los Alamos, where they were meant to calculate the implosion of atomic bombs. The scientists got the computers, put them together, got them running – and immediately wrote a chess program to experiment with.
AH: Maybe back to Turing for a moment. You told me that – you also tried to reprogram it, but found that there were still errors in the algorithm?
FF: It wasn't an error in the algorithm, it was an error in his calculations. We reconstructed the program, ran it on Fritz, and Garry Kasparow played a game in Manchester for Alan Turing's 100th birthday. Only during development we found – we had a single sample game – we found that our program deviated in 4, 6 or 8 places. We checked and looked for the mistake we had made. And then I called Donald Michie, a Turing employee who was still alive, and I asked him, "What are we doing wrong?" And he says, "You're doing something wrong? No, Turing did something wrong. He didn't have time to calculate things that precisely, he just wanted to understand the principle. And so of course he made a mistake every now and then.” In the very first move of the game we also play [pawn] e3. Turing played [pawn] e4 because that's so natural for a human. He was influenced by his intuition for chess. He said: "It has to be this", so that it doesn't take 40 minutes per move, but only 15.
AH: But still, I think it's a very nice story. And yes, there is an article about exactly that, which we will definitely link to for those who might want to read it again, because I found the story very, very nice.
Frederic, I think we're at the end now. I think we can go on talking for hours, because it's really a very exciting topic and I really enjoy listening to what you say, because I think you make it very lively and very understandable. So I thank you very much for your time and for your contribution and maybe we'll see each other again at one point or another, because I think there will certainly be a few developments on this topic in the future that can be discussed. So thank you again and of course all the best to our listeners.
FF: Thanks.
