Since I am now in my mid-fifties and must take care with my blood pressure, allow me to leave that bitter scene behind for a moment before returning for the final games of the rematch and a closing press conference that made the one after game three look like a child’s tea party in comparison.
There is a long and ugly history of recriminations and accusations of foul play and worse during world championship matches. The most popular anecdotes are trotted out in every layman’s book about chess because they are amusing from a distance. Fischer’s protests about the cameras in the playing hall in his 1972 match against Spassky led to him forfeiting the second game and to the third being played in a little room instead of the main stage. Karpov and Korchnoi often feuded viciously during their matches, especially their 1978 world championship in the Philippines. Karpov had a psychologist, some say a parapsychologist, named Dr. Zukhar on his team who stared at Korchnoi during the games. A series of dueling protests moved the man around the room in nearly every game. Korchnoi retaliated by inviting some American members of an Indian sect, who meditated and stared at the players and Karpov’s man. There were protests and investigations about the chairs—including having Korchnoi’s dismantled and x-rayed—Korchnoi’s mirrored glasses, and Karpov’s yogurt.
The 2006 world championship match between Vladimir Kramnik and Veselin Topalov sank to new lows—all the way to the plumbing fixtures. After accusations by Topalov’s camp that Kramnik was spending a suspicious amount of time in his personal bathroom during games, the organization closed it, leading to Kramnik forfeiting the fifth game in protest in a scandal quickly dubbed “Toiletgate” by the chess media. (Kramnik went on to win the match anyway.)
My epic rivalry with Karpov was not immune to such adventures, naturally. In the 1986 rematch, Karpov repeatedly displayed an almost magical intuition regarding my opening preparation. He met several of my novelties almost instantly, with the strongest responses, and seemed completely prepared even for lines I could not have been expected to play. I felt that the only way this could keep happening was if someone on my own team was sharing my preparation with Karpov. Two of my team members ended up leaving, though not before I lost three games in a row. A later article by one of Karpov’s own team includes remarks about how Karpov had spent a sleepless night analyzing a variation “he was sure would occur” in our next game, despite it being a completely different opening from my previous two games with white, which I won. Needless to say, his premonition was correct.
In sum, you can either believe that there is a great deal of treachery at the top level of chess, that some Grandmasters are as paranoid as the stories say, or that gamesmanship and off-the-board maneuvers are a standard element of an all-out psychological war. Or you may select “all of the above” and join the consensus.
My next clarification is about the dangerous words Ashley floated into the air at the press conference: “human intervention.” I have spent twenty years dealing with the many loaded meanings of this phrase, although I did not coin it and my insinuations were more complex. A certain amount of human intervention was allowed on Deep Blue’s behalf during the match. They were allowed to fix bugs, reboot after crashes, and to change its book and evaluation function between games, for example, and they did so. Later human-machine matches would limit this sort of activity, judging it as an unfair advantage for the computer.
There were at least two crashes of the machine during play, requiring a manual restart. According to the Deep Blue team, this happened in game three and game four. While neither incident seemed relevant to them because it didn’t affect Deep Blue’s next move, having to ask Hsu what was going on during a tense endgame in game four was far from ideal. And, as was subsequently pointed out to me by several chess programmers, a system restart changes everything from the perspective of reproducibility. The memory tables the machine uses to retain positions are lost and there will never be a way to confirm that the machine would repeat the same moves.
Putting allowed user actions aside, most people take the idea of human intervention to mean that Karpov or some other strong Grandmaster would be hiding in a box somewhere making moves in the style of Wolfgang von Kempelen’s chess-playing hoax automaton, the Turk. But in the modern day, with all the backups and remote access points, it wouldn’t require a chess master dwarf hiding inside the big black box. An amusing thought, but not really the point. Simply rebooting the machine, or triggering an event that forced Deep Blue to take extra time in a tricky position could be enough to make a big difference. Remember that the Deep Blue prototype in the 1995 Hong Kong tournament had to be restarted during its key game against Fritz, and it came back online with an inferior move. Bad luck, but it could also have come back with a superior move, especially if, for example, it had been programmed to take extra time after a crash.
On September 15, 2016, I was in Oxford to speak at the Social Robotics and AI conference, and I jumped at the chance to meet Noel Sharkey. From the University of Sheffield, Sharkey is one of the world’s great experts on AI and machine learning, and he’s currently involved in various projects on ethical guidelines and the societal impact of robots. But he’s best known in the United Kingdom as an entertaining expert and head judge on the popular television show Robot Wars. We only had a short time to speak during a lunch break before his conference keynote. I wanted to talk about machine learning and his United Nations debate on robot ethics. But he wanted to talk about Deep Blue!
“I’ve been annoyed about it for years,” he told me. “I was very excited about the prospect of an AI system beating you but I wanted it to be a fair contest and it wasn’t. The crashes? All the connected systems they put in? How do you monitor that? They could change software or hardware between moves. I can’t say IBM cheated but I can’t say that they didn’t. They certainly had the opportunity. Forget it! If I had been adjudicating. I’d have ripped out all their wires, put a Faraday cage around Deep Blue, and said, ‘Okay, now play, you’re on your own.’ Otherwise I’d have forfeited the damn thing in a second!” The mental image of Noel Sharkey ripping network cables out of Deep Blue made me sure I’d want him on my team no matter who I was playing against.
Lastly, the argument that IBM would never do or allow anything inappropriate in order to help Deep Blue’s winning chances was popular at the time but sounds almost quaint today. Nineteen ninety-seven was still four years before the Enron scandal rocked the corporate world, exposing the American energy giant as an empire of malice and fraud. It was a Watergate moment for the corporate world, and a preview of the revelations that came out of the 2007–08 financial crisis. I’m not putting a chess match on the scale with a catastrophic financial meltdown, of course. I make the point because after Enron, people stopped telling me that “a big American corporation like IBM would never do anything unethical.” Especially after they found out how much IBM’s stock price went up after the match.
Thanks to the honesty of Miguel Illescas, we do know that IBM was willing to push the boundaries of ethical behavior to improve Deep Blue’s chances in any way. In his 2009 New In Chess interview, he shared this remarkable revelation: “Every morning we had meetings with all the team, the engineers, communication people, everybody. A professional approach such as I never saw in my life. All details were taken into account. I will tell you something which was very secret. Well, it’s more of an anecdote, because it’s not that important. One day I said, Kasparov speaks to Dokhoian after the games. I would like to know what they say. Can we change the security guard, and replace him by someone that speaks Russian? The next day they changed the guy, so I knew what they spoke about after the game.”
As he says, perhaps not that important in practice, but it’s a bombshell in exposing the lengths IBM went to in order to gain any competitive advantage. I can hardly imagine the scandal that would have erupted had it been revealed during the match that IBM had hired Russian-speaking security personnel stationed in my personal rest area specifically to spy on me and my second during the match, but it would have been ugly.
After saying all of that, we come to my own confession. On what mattered most, on what really destroyed my composure, I was wrong and owe the Deep Blue team an apology. The moves in game two that left me with a lost position and crushed morale were unique only for the time. Within five years, commercial engines running on standard Intel servers could reproduce all of Deep Blue’s best moves, even improving on some of the “humanlike” moves that so impressed me and everyone else at the time. The engine on my laptop today slightly favors the “shockingly humanlike” move 37.Be4 from game two in less than ten seconds, although it rates it nearly equal to the queen sortie I had expected because 37.Be4 turns out not to have been as superior as we all believed at the time. Had I played better defense instead of collapsing and resigning, game two would have been considered a very impressive game for a machine but nothing more, no matter the eventual result.
This also highlights why it was so critical that I never saw a single game of Deep Blue’s before the match. Had I seen it make a single move demonstrating the uncomputerlike positional approach of game two’s Be4, for example, or the surprising h5 pawn push from game five, my play and my reactions would have been completely different. Keeping Deep Blue completely hidden was the strongest move of the match, but it was made by IBM, not by either of the players.
In turn, by understanding now that Deep Blue was very strong but still making plenty of inaccuracies, the fact that it missed the perpetual check draw at the end of game two becomes more comprehensible as well. Still strange, considering its powers of calculation, but no longer inconceivable. If I had had any way of knowing this during the match, perhaps the story would have turned out differently, but I’m not sure. My premature resignation in game two and the intense shame and frustration it produced in me were what made it nearly impossible to play.
I have my regrets, but I was not wrong to be shocked and confused at the time. In 1997, Deep Blue’s play was completely inexplicable to me, and IBM went to great lengths to keep it that way. Maybe they really didn’t have anything to hide, but they realized it couldn’t hurt if they acted as if they did, stoking my suspicions. They continued to stonewall us on releasing the printouts from game two, which, if they had showed nothing amiss in the eyes of Ken Thompson, could have relieved some of my stress about what was going on behind the scenes.
My agent Owen Williams told the organizers before game four that if Thompson did not receive the game two printouts, he wouldn’t be able to appear as a member of the Appeals Board. IBM took this as a warning that I might not appear either if Thompson didn’t, and they warned the media that there might not be a game that day. Thirty minutes before the game started, we got a message from Newborn saying that the printouts had been given to the Appeals Board, but when we arrived at the thirty-fifth floor, Thompson said that they were only for one move, 37.Qb6. Without the other moves to show context, this was useless.
This secretive and antagonistic behavior manifested in other ways as well, as reported in the New York Times after game five: “One reporter, Jeff Kisseloff, who had been hired by IBM to report on the Kasparov team for the match Web site, lost his reporting privileges after he included damning comments about Deep Blue from the champion’s supporters in his report. IBM also engaged Grandmasters John Fedorovich (sic) [a.k.a. Fedorowicz] and Nick DeFirmian to work on openings with Deep Blue, though no one on the Deep Blue side has said so publicly, even when asked directly in a news conference about additional help. It was Mr. DeFirmian who confirmed his involvement and that of Mr. Fedorovich, but declined to discuss it, he said, because IBM had insisted he sign a secrecy agreement.”
All of this prompted my mother to say, “It reminds me of the 1984 world championship match against Karpov. You had to fight Karpov and also the Soviet bureaucracy. Here we are, thirteen years later, and you have to fight a supercomputer and also a capitalist system using psychological warfare.” (If her use of “capitalist” sounds like a Marxist anachronism, remember what the first match did for IBM’s stock price!)
THE GAMES had to continue at the board as well, and I had black in game four. Obviously I would not repeat the disaster of game two, in which the typical roles of human-machine chess had been reversed. The computer had built up a dominating position with strong strategic play while I had been reduced to defensive shuffling. But when the machine finally broke through to convert its advantage, it made a tactical slip that could have been immediately exploited to lead to a shocking forced draw. (As everyone still thought at the time.) It was the same pattern as countless games since the first human-machine contests, only the computer and human had switched positions. In games four and five, the players would resume their normal roles.
I returned to a flexible defensive system in the fourth game and achieved a solid position after several diffident moves by Deep Blue. It still occasionally showed the downside of not being able to logically connect moves the way a human does. It advanced its pawns on the kingside and then seemed to forget about them as it found other alternatives, giving a strange impression. Again, there are advantages to this extreme objectivity, but there is a reason we say that a bad plan is better than no plan, at least in human chess. If you have a plan and it fails, you learn something. If you act aimlessly, from move to move, from decision to decision, whether in politics or business or chess, you don’t learn and will never become anything more than a skillful improviser.
The machine was pushing hard, too hard, and by so doing it created weaknesses in its own camp. On move twenty, I played a strong pawn sacrifice to free my pieces and to turn the tables. The machine again made a couple of strange moves that the commentators, at least the human ones, were quick to dub “ugly” and “pointless.” GM Robert Byrne wondered, “How can it be very strong one day and loony the next?” And perhaps they were loony to the commentators, but I had already come to appreciate that Deep Blue had a knack for making its moves work, however ugly they might be to a Grandmaster. This makes sense, because even if a machine doesn’t employ the goal-oriented strategy that humans use, if it evaluates a move as the best it’s because something in its evaluation likes it and the positions resulting from it. It’s an alien version of how Grandmasters have different styles. A move made by former world champion Tigran Petrosian, famed for his defensive skills, might look completely pointless to an attacker like me. Indeed, that move would in effect be weak if I made it, but it was strong for Petrosian because he understood it and what would come of it. Deep Blue could still make genuinely weak moves and pointless moves, of course, but it was strong enough that its computerlike inconsistency often worked out fine for it.
In another terrible disappointment, game four turned into another example of this. I missed one good attacking chance, but still had the clear upper hand all the way through to the endgame, only to find that the machine had a series of incredible drawing maneuvers that I could never have foreseen. Even today, looking at the game after move thirty-six, I cannot believe I failed to win that position, and even more incredibly, that the position might not even be objectively winning at all. With two rooks and a knight each, and a scattering of pawns, every aspect of the position was in my favor. My pieces were more active, its pawns were isolated and vulnerable. Even my king was better placed for the endgame. I estimate that I would win that position against a very strong Grandmaster four out of five times.
It was almost as if Deep Blue was taunting me by getting as close to losing as possible before coming back to draw. Material on the board slowly dwindled and I was finding it hard to calculate clearly as I began to tire. The forced win I was so sure was just around the corner never stopped being just around the corner. Commentators and later analysts were as surprised as I was, and kept looking for mistakes in my play that had let Deep Blue off the hook in the ending. But while perhaps I did not play flawlessly, it appears that there was simply no win to be had. Any strong player could explain why black’s position was clearly superior, but even Grandmasters with strong engines have failed to demonstrate how to win it. It was another demoralizing and exhausting day at the board.
After the game, I asked Frederic if he thought Deep Blue had used its secret weapon to help it achieve the miraculous draw. There were rumors that the machine could access endgame tablebases during its analysis, and, if so, I would have Ken Thompson to blame for it. In 1977, Thompson showed up at the World Computer Chess Championship with a new creation, a database that played the king and queen versus king and rook endgame perfectly. (KQKR is the abbreviation.) It wasn’t an engine; there was no thinking required. Thompson had generated a database that essentially solved chess backwards, what we call retrograde analysis. It started from checkmate and worked its way back until it contained every single possible position with that material balance. Then it worked out the optimal move from every one of those positions. For example, in KQKR, for the side with the queen it always played the moves that led to checkmate quickest. For the side with the rook it always played the move that delayed checkmate the longest. It didn’t play like a god, it was God. Or more accurately, the goddess of chess, Caissa!
It was a revolutionary contribution to computer chess, where the subtleties of endgame play had long been a machine weakness. A human can look at a pawn endgame and see instantly that if one side has two pawns versus one on the same side, he can create a passed pawn that will become a queen. That might take fifteen or twenty moves to actually happen on the board, but you don’t have to calculate them all to know what will eventually occur. A computer, on the other hand, does have to calculate all the way to the pawn queening to see the truth in the position, and that was often far too deep even for strong engines to reach.
With tablebases, all that started to change. Instead of calculating all the way, a machine only had to reach a tablebase position in its calculations to know if it was winning, losing, or a draw. It was like gaining second sight. Not every chess game reaches an endgame, so their utility was limited, but as tablebases grew bigger and bigger, incorporating more and more pieces and pawns, they became a powerful new weapon in the computer arsenal.
Thompson’s endgame databases were also the first computer chess innovation to have an impact on human chess. When he started with KQKR, he challenged Grandmasters to play against it, to see if they could win with the queen against his database. Keep in mind that it was generally considered not that difficult for a strong player to win queen versus rook; the general algorithm was taught in every endgame book. Incredibly, the machine showed how hard it really was, and it did it by playing moves that were inexplicable even to Grandmasters.
Six-time US champion Walter Browne lost a bet with Thompson when he failed to beat the database in under fifty moves—the amount of moves the rules of chess allow you to try to win such positions before the defender can claim a draw. Shocked, Browne, ever the gambling man, studied for a few weeks and returned for another try, mating it in exactly fifty moves and getting his money back. The position was actually a win in just thirty-one moves with perfect play, according to the database. For the first time, humans were being exposed by computers as far from perfect chess players.
The massive data storage required for each new piece added at first made tablebases impractical for most engines. One common set requires 30 megabytes for all positions with four pieces, 7.1 gigabytes for all positions with five pieces, and 1.2 terabytes for all positions with six pieces. Their use became commonplace as new data generation and compression techniques came along, and as hard drives kept getting bigger and bigger.
Just as the search tree from the beginning of a chess game grows too quickly to ever solve chess from the start, tablebases are far too huge and too difficult to generate to ever solve chess from the end. Theoretically, a thirty-two-piece tablebase could be generated, but we cannot even conceive of how much storage space it would require. Seven-piece bases only started to appear in 2005, due to the massive computing resources they require to generate and store. There are now full sets of seven-piece tablebases that take months to generate and occupy 140 terabytes. Now accessible online, they were originally generated by Russian researchers Zakharov and Makhnichev using a Lomonosov supercomputer at Moscow State University.
These have revealed some fascinating things about the complexity of chess while also refuting centuries of chess analysis and studies. For example, the longest mating position for seven pieces is KQNKRNB (king, queen, and knight versus king, rook, knight, and bishop). If configured just so, it takes exactly 545 perfect moves on both sides to force checkmate. More practical and well-known positions have also had to be reevaluated. It was assumed for a century that in some positions it was impossible to win with two bishops against an ideally placed lone knight, but the tablebases showed this was false.
There is a long history of chess studies and problems, in which the composer artfully arranges the pieces and presents the reader with a stipulation, usually “white to play and win” or “white to move and mate in three.” These are often found in the chess columns of local newspapers—if newspapers still have chess columns (and if we still have newspapers). Many of them look simply impossible and their solutions often reveal great wit and beauty. Databases care not for such things and many compositions have been refuted by the machines.
In a few cases, how the databases play some common positions can be useful for the human player to study, but this is rare. We need useful patterns and heuristics like “put your rook behind the passed pawn” or “keep your rook near your king when defending against the queen” in order to play. Tablebases generally provide no help in how to make these endgames easier for humans to understand. Even to me, 99 percent of tablebase moves in some positions are completely incomprehensible. I have flipped through several six- and seven-piece endings that require over two hundred moves to solve, and often the first one hundred and fifty moves looked like nothing was happening at all, revealing no pattern I could grasp. Only as mate came within forty or fifty moves could I start to see method in the machine madness.
It was one thing to face a giant opening database that had been prepared by a team of Grandmasters. It was quite another thing to play against an endgame database that played literally perfectly. Later human versus machine matches took steps to balance this part of the playing field as tablebases became bigger and more common. For example, in my 2003 match with Deep Junior, this line was added to the rules: “Should a position be reached which is in the machine’s endgame databases and if the result from that position with correct play is a draw, then the game ends immediately.” Otherwise, a game could become more of a strange form of solitaire than a competition.
Tablebases are the clearest case of human chess versus alien chess, and of the huge difference in how humans and machines achieve results. A decade of trying to teach computers how to play endgames was rendered obsolete in an instant thanks to a new tool. This is a pattern we see over and over again in everything related to intelligent machines. It’s wonderful if we can teach machines to think like we do, but why settle for thinking like a human if you can be a god?
This question was on my mind when I looked over Deep Blue’s incredible defense in game four. The rook endgame had been drawn with eight pieces on the board, too many for tablebases then or now to render a perfect verdict. But what if Deep Blue was accessing tablebases during its search? Could it be looking ahead and checking to see which positions were winning and losing in order to improve its evaluations? This “probing” of tablebases in the search later became standard for engines, but we weren’t sure if Deep Blue was doing it. If so, it was cause for concern. Would I have to add some endgames to the realm of positions I had to avoid against Deep Blue?
According to papers published later by the Deep Blue team, the machine did have access to tablebases during the match, and indeed used them briefly in its search in game four, the only game that reached a simplified endgame position. Six-piece tablebases were quite rare at the time, so I was surprised to read that Deep Blue’s contained “selected positions with six pieces” they had specially requisitioned from an expert.
Game four also included another crash, after I made my forty-third move. Every computer user knows what a crash is: your machine freezes, or the screen turns blue, and it’s time to curse and reboot. I’ve had many laptops and projectors crash on me during my lectures, which gives me the chance to quip that it’s because computers still hate me! But in discussing these events with experts, including one of the creators of the multiple computer world champion program Deep Junior, Shay Bushinsky, I realized how oversimplistic my understanding was. He pointed out that just about anything can take place during the recovery process, especially if it was a “controlled crash” instead of a catastrophic halt. Programmers often insert code that will restart all or part of their program’s processes under certain conditions. In fact, this is what happened to Deep Blue, according to Hsu’s book Behind Deep Blue. He calls them “self-terminations,” not crashes, and describes it as a “piece of code that monitored the efficiency of the parallel search and terminated the program itself if the efficiency dropped too low.”
This is a remarkable admission because it says that these distracting crashes—sorry, these distracting “self-terminations”—were a feature, not a bug. Not exactly intentional, as in occurring on demand, but a working part of the system used to “clear the pipes” if Deep Blue’s parallel processing system became clogged. This isn’t to say that they directly improved Deep Blue’s play, or that it would necessarily be unfair if they did, depending on the rules in place. But aside from annoying me during play as they fiddled with the machine, it made the games impossible to reproduce.
This was the biggest problem, according to Shay. “Once it crashes, the entire thing is kind of a sham because you can never confirm what happened is authentic,” he told me over dinner near his home on a sweltering evening in Tel Aviv in May 2016. I was in Israel to give two lectures, one on education and another on the human-machine relationship, and took advantage to gain the input of an old friend and colleague who also happened to be a world-class expert on machine chess. “The move timing changes, the hash tables change, who knows what else? There is no way to say afterward, ‘This is exactly why the machine made this move’ with any conviction. That’s not so bad in testing or in a friendly game, but in a high-profile competition, with millions of dollars at stake, it’s unacceptable.”
The game four crash took place on Deep Blue’s move where, in a stroke of luck, there was only one legal move in the position. I had just checked its king with my rook and its reply was forced, so there was little concern this time that it would be aided or harmed by the reboot.
IBM CEO Lou Gerstner made a visit to the match during the game, though I doubt he was informed that his computer star had crashed again. All the great PR Deep Blue was providing IBM would have taken quite a blow had the media started asking about crashes or self-terminations. Gerstner gave his team a pep talk and told the press that the event was “a chess match between the world’s greatest chess player and Garry Kasparov.” Considering that the match was tied and Deep Blue’s only win was in a drawn position that I had resigned, this seemed more insulting than accurate.
I felt completely drained, but we had two days off to prepare for the final two games of the match. I very much wanted to use my turn with white pieces in game five to make Gerstner eat his words.
We had already scheduled a special dinner for my team and friends for that night, although I really just wanted to go to sleep for ten hours. On the first rest day, we prepared a little for my black in game six. Then, on Friday, we started on game five and decided to stick with the anti-computer strategy that had done reasonably well in games one and three. The Réti Opening it would be. Meanwhile, we had asked that the printouts from games five and six be sealed immediately after the game and given to the Appeals Committee for safekeeping.
The opening of game five again showed the ups and downs of my anti-computer, anti-Kasparov strategy. I got the maneuvering position I wanted despite losing some time in the opening. I hadn’t gained any real advantage with the white pieces, but there was still a long game ahead. Deep Blue’s eleventh move was a surprising one, pushing its h-pawn forward two squares. The commentators thought this might be another case of Deep Blue making silly computerlike moves, but I wasn’t so sure. It created a threat on the kingside and appeared to me to be less the move of a machine than one in the style of a very aggressive human player. It was early in the game, so there were many logical moves for black. Its choice of this surprising thrust at the edge of the board again had me shaking my head at what Deep Blue was capable of. I think I even glanced up at Campbell for a second after he played ..h5, as if to confirm it wasn’t a slip by the operator.
It turns out that ..h5 wasn’t very good and that I could have gotten a large advantage by moving my knight to the e4 square, but I responded weakly. Once again, it was a case of a strange but weak move by Deep Blue turning out to be more effective than a good move because of how it affected me psychologically. I never got a sense of what to expect, never felt sure of how I should play, and I let it ruin my concentration. And when these odd moves were combined with all the off-board conflicts, I also let my imagination get the better of me.
The position opened up while I searched for a way to secure an advantage. Analyzing today, I am again struck by how many opportunities I missed. I was at my peak as a player and as of this writing I have been retired from professional chess for over a decade. And yet some of my moves seem obviously bad to me, and analysis backs this up. As poorly as I played, I was lucky the match did not turn out even worse for me.
After some exchanges, it looked like the position was about equal. I didn’t see how either side could play for a win. Then, to my delight, Deep Blue played a terrible queen move, allowing me to exchange the queens. Without the powerful queens on the board to generate threats, black’s structural weaknesses became more prominent. Now I had targets I could go after the way I did in game two of the Philadelphia match.
It worked for a while; I was making progress as more pieces were exchanged. Just as in game four, I would look at this endgame and be absolutely confident I could win it against any human player. But once again Deep Blue defended aggressively, finding remarkable tactical resources to hold on. It brought up its own pawn and king to create threats against my king and I was forced to accept a pretty repetition draw with my pawn one square away from becoming a queen. I had seen the forced draw coming much earlier than the commentators, who were still under the impression I was winning until almost the last minute. For the second game in a row I felt shattered, certain I had squandered a winning opportunity and disgusted with the low quality of my play.
Before I left the board, I demanded that the printouts be turned over to the arbiter or Appeals Committee immediately. The room filled with people, much to the confusion of the spectators watching on the video screens. After more promises were made by C. J. Tan, who had earlier told the Appeals Board there wouldn’t be any printouts until after the match, we went downstairs to discuss the game with the audience. Afterward, we went up for the printouts and no one was there. I went back to the hotel while Michael and my mother waited and tried to reach someone. The printouts would, finally, be delivered by arbiter Carol Jarecki. (Deep Blue’s full analysis logs wouldn’t appear in public until several years after the match was over, when they were quietly uploaded to the website IBM created for the event.)
In the auditorium, I was again met with cheers. I was simply unable to feel buoyed by the crowd’s support by that point, as nice as it was to hear. I felt like I couldn’t see anything anymore. Even after it gave me several chances, I couldn’t find the win I was sure was there, leading to another incredible escape by the machine. It was incredibly frustrating. That was an accurate assessment, analysis with modern engines shows. I had missed two good winning attempts and Deep Blue had again blundered badly, but yet again I had failed to exploit its mistakes. It turned out much later that I did miss a win in the game five endgame, not that this made me feel any better.
At the press conference, I was again frank about how impressed and surprised I was by some of Deep Blue’s moves, especially the one that had elicited laughter from the commentators. I said, “I was very much amazed by ..h5. Many discoveries in this match, and one of them is that sometimes the computer plays very human moves. ..h5 is a good move and I have to praise the machine for understanding very, very deep positional factors. I think it’s an outstanding scientific achievement.”
I want to enter that statement in my defense for when I’m told I did not give enough credit to Deep Blue and its creators, especially since it turns out that ..h5 wasn’t even a very good move! When the match ended the next day, and because of how it ended, I was in no mood to be flattering.
When asked about remarks by Illescas that I was afraid of Deep Blue, I was again candid. “I’m not afraid to admit I am afraid! And I’m not afraid to say why I’m afraid. It definitely goes beyond any known program in the world.” At the end, Ashley asked me if I was going to try to win the final game with the black pieces and I replied, “I’ll try to make the best moves.”
IN A MATCH of many firsts and many records, game six of the rematch would add several others, none of them good for me. It was the shortest loss of my career. It was the first classical match loss of my career. It was the first time a machine had defeated the world champion in a serious match. As with exhibition games, such things acquire an asterisk in the record books when the game or match is against a computer, but I wasn’t concerned about asterisks or my place in history. I had lost, and I hated losing.
The stories around the sixth and final game have grown in a way that I suppose is fitting for such a historic moment. It has acquired its own mythology, with different factions arguing for their interpretations. Rumors about what really happened in game six are passed around among the faithful like the shreds of a prophet’s shroud.
The chess always came first with me, and so I very much wish the game itself were worthy of the moment and of so much attention. Losing a battle, even losing a masterpiece, would have been far more to my liking. Instead, it is little more than an ugly joke of a chess game, promoted into a historical artifact by circumstance.
The match was tied, 2.5–2.5. Should I play it safe and aim for a draw or should I risk everything and play for a win with black? With no rest day, I knew I would have no energy for another long fight of the sort that resulted from my anti-computer lines. My play was already shaky. I knew my nervous system very well from two decades of competition, and it would not withstand the strain of another four or five hours of tension against the machine. But I had to try something, didn’t I?
It was the second time in the match that I played a “real” opening. The first time was the failed Ruy Lopez experiment in game two. This time I played the Caro-Kann, a solid positional choice that was a favorite of my nemesis Karpov, who used it against me several times in our games. I played it extensively in my youth, but decided fairly early on that the sharp Sicilian was far more in keeping with my attacking style. Deep Blue continued with a main line that I knew very well from having played it with white on numerous occasions. Perhaps Deep Blue’s opening book coaches had a sense of irony, or maybe they just thought that if it was good for me, it would be good for their machine.
On the seventh move, still following the main line, I reached out and played my h-pawn one square, instead of the bishop move that usually precedes the pawn move. There were shouts of disbelief in the commentary room as Deep Blue responded instantly, crashing its knight into my position with a devastating sacrifice. My king was exposed, my pieces were undeveloped, and white had overwhelming threats. You can see on my face that I knew the game was already over. I went through the motions of trying to defend a position that would be very difficult against any Grandmaster and, I knew, was absolutely hopeless against Deep Blue.
I played another dozen moves on autopilot, barely registering what was happening. I ignored it when operator Hoane picked up the wrong bishop on move ten. On move eighteen I had to give up my queen and on the next move, facing further losses, I resigned. The whole game had taken less than an hour. The match was over.
If you can, for a moment, imagine what that moment felt like for me, take one extra step in my shoes and imagine then having to face hundreds of reporters and a large audience asking you about it. The press conference felt like a strange continuation of the game. I was in shock, exhausted, and bitter about everything that had happened on and off the board. When it was my turn to speak, I told the audience that I certainly did not merit their applause after what had happened in the final game, and I admitted that I had felt like the match was already over after I failed to win the endgame in game five. I said I was ashamed. I admitted that it had been a big mistake not to prepare for the match properly and to play my normal preparation, that my anti-computer strategy had not worked.
I reiterated both my praise and my concerns over Deep Blue’s inexplicable moves, and threw down a challenge to IBM to let Deep Blue participate in regular tournaments, when, I promised, “I will rip it to shreds.” I said I would play Deep Blue under any conditions, with the only caveat being that IBM could only participate as a player, not as sponsor or organizer. I announced I would play it again with my world championship title on the line.
When I read over the transcripts of the press conference to refresh my memory, I didn’t think I sounded quite like the villain I was portrayed as afterward. I went on too long from sheer adrenaline, and repeated myself more than once. And I was far from gracious toward the victorious Deep Blue team in their moment of glory, and for that I must apologize.
But when I listened to the audio of the press conference, I could understand why they later said that I had “taken the joy out of it.” Drained and disappointed, the anger and confusion are palpable in my voice. I cannot say that I regret speaking my mind, because it is my nature to say what is in my heart. But I could have waited until the next day, after some rest and contemplation. It is fair to say that I had failed to rise to the occasion in game six and then I failed again at the press conference.
So, what did happen in game six? When asked several times at the press conference I deflected the question: “It doesn’t even count as a game.” “I was not in the mood of playing at all, I have to tell you.” “When you allow this piece sacrifice you can resign and there are many games played in competitive chess in which this line has happened, but I can hardly explain what I did today because I was not in a fighting mood.”
This was all true, but it does not explain why I played the horrible 7..h6 instead of the normal 7..Bd6. Several competing theories have formed the mythology of game six. One, that I was so discombobulated and tired that I had transposed these routine moves by accident, playing them out of order. My defenders and friends have advanced this theory, which made its way into various news reports and books. Two, that I was trying to lure Deep Blue into a trap, based on some recent analysis in a computer chess journal that showed black could defend after the knight sacrifice. Three, that I played the Caro-Kann as a last-minute inspiration and didn’t prepare, leaving me ignorant of this devastating blow.
Honestly, I find the suggestion that I blundered in my preparation to be more insulting than the idea that I suffered a complete nervous breakdown. Of course I was aware of Nxe6. I was also aware that it would be a killing move if Deep Blue played it against me in game six. I simply knew that it wouldn’t.
Machines are not speculative attackers. They need to see the return on their investment in their search before they invest material. I knew that Deep Blue would decide to retreat its knight instead of playing the sacrifice, after which my position would be fine. I knew I didn’t have the energy for a complex fight and that I would achieve stable equality this way. We tested it out on a few engines and they all retreated the knight. They thought the sacrifice was playable for white as well, but even when coached ahead a few moves, they did not like giving up a whole piece without concrete gains and evaluated the retreat higher.
While looking at the horrible positions for black that resulted, I realized that only a computer would be able to defend them, and that was the point. Computers love material and are incredible defenders. I was sure that Deep Blue would apply its fantastic defensive prowess to my position, evaluate it as fine for black, and therefore would decline to sacrifice the knight. I lost this bet, obviously, and spectacularly so, but the reason I lost it would not be clear for over a decade.
It may surprise you to find out that I was completely right in my evaluation of Deep Blue. It would never sacrifice the knight. And yet, it did. Why? Because of one of the most remarkable coincidences in the history of chess, or perhaps in history, period.
Here once more is Deep Blue coach Miguel Illescas in his 2009 interview, speaking about the fateful sixth game: “We were looking at all kinds of rubbish, such as 1.e4 a6 or 1.e4 b6, giving as many forced moves to the computer as we could. On this same morning we also introduced the move Knight takes e6 in the Caro-Kann, on the same day that Kasparov played it. That very morning we told Deep Blue, if Garry plays h6, take on e6 and don’t check the database. Just play, don’t think.… This was his bet, that the machine would never like this piece sacrifice for a pawn. And indeed, if we had given freedom to Deep Blue to choose, it would never have played it.”
I will not repeat here the stream of profanities in Russian, English, and languages not yet invented that escaped my lips when I first read that paragraph. What in the hell was this? Two paragraphs after Illescas says IBM had hired Russian speakers to spy on me, he says the team entered this critical line into Deep Blue’s book that morning? An obscure variation that I had only discussed with my team in the privacy of our suite at the Plaza Hotel that week in New York?
I’m no Nate Silver, but the odds of winning the lottery are quite attractive in comparison to those of the Deep Blue team entering a specific variation I had never played before in my life into the computer’s book on the very same day it appeared on the board in the final game. And not only preparing the machine for the 4..Nd7 Caro-Kann—even during my brief dalliance with the Caro-Kann as a fifteen-year-old I played the 4..Bf5 line exclusively—but also forcing it to play 8.Nxe6 and doing this despite generally giving Deep Blue “a lot of freedom to play,” in Illescas’s own words.
Am I alone in failing to make the leap of faith required to believe that the timing of this could possibly be innocent? I am trying, but I am failing. The IBM team went to great lengths to ridicule me about my “hand of God” remarks, and maybe I deserved it. Deep Blue’s moves were inexplicable, partly because IBM refused to explain them, but they were not human. But perhaps that was all part of the psychological warfare while other games were afoot. As Pynchon’s Proverbs for Paranoids, number 3 says in Gravity’s Rainbow, “If they can get you asking the wrong questions, they don’t have to worry about the answers.”
HAD I NOT melted down during game two and resigned prematurely, none of this would have mattered. Not only was the early resignation my fault, but allowing it to ruin my composure was the real fatal mistake. I played so far below my usual level after that that it has been a little embarrassing to go over the games for this book. As I said the day after the match on the Larry King show, rested and calmer, “I do not blame IBM, I blame myself.” I then again challenged Deep Blue to the rematch I believed I deserved after winning the first and losing the second. I wanted to play under neutral conditions, and I wanted to see if I could beat it playing normal chess. Not anti-computer chess; Kasparov chess.
Of course this never happened. Deep Blue never played another game. I can sympathize a little with those who say that IBM had gotten what it wanted already, a giant PR boost and an increase in its stock value of $11.4 billion in just over one week. If the entire project cost IBM the estimated $20 million they said, it’s an enviable return on investment even if only a fraction of those billions was due to the match. A loss to me in a rematch would be embarrassing, and, even if they won again, nobody remembers the second man to scale Mount Everest.
Later that night, I shared the elevator at the Plaza with the actor Charles Bronson. After a flicker of mutual recognition, he said, “Tough luck, man!” I said, “Yes, I’ll try to do better next time.” He shook his head and replied, “They’ll never give you the chance.” He was right.
A few days after the match, a Wall Street friend arranged a phone call between me and IBM CEO Lou Gerstner. I told him that since I had given his machine a rematch, he owed me and the world a rubber match. He was friendly and talked about the great potential, but I could tell it was never going to happen. It was polite, but it was a polite brush-off. He wasn’t interested and, if Gerstner wasn’t interested, IBM wasn’t interested.
The argument that IBM abandoned Deep Blue and chess because I was mean to them at a press conference is a little odd. If that was their excuse for not participating in a rubber match, all right, but why take Deep Blue apart? “It’s already directing traffic in Pittsburgh,” wrote one columnist. Why not let it play in tournaments, or analyze games? Why not put it on the Internet to let millions of chess fans challenge it? Deep Blue was the biggest thing to come out of IBM in ages, so why shut it down overnight instead of capitalizing on a machine that enjoyed better name recognition than sports stars like Pete Sampras? If IBM was offended by my “wild allegations” about the integrity of Deep Blue’s capabilities, then shutting it down immediately and limiting the team’s ability to speak about it was a strange way to respond to them. Even a single game against anyone else would mean lowering Deep Blue from its pedestal, exposing it to scrutiny and criticism. It beat the champion and retired, Fischer-like, becoming as much myth as machine.
Chess fans and especially the computer chess community were outraged. They called it a crime against science, against the spirit of the quest for the holy grail started by Alan Turing and Claude Shannon. As Frederic Friedel put it to the New York Times, perhaps poking fun at Monty Newborn’s comparison of Deep Blue’s win to the moon landing, “Deep Blue’s victory over Kasparov was a milestone in artificial intelligence, but it’s a crime that IBM didn’t let it play again. It’s like going to the moon and returning home without looking around.”
As this book was headed to press in December 2016, my coauthor Mig Greengard exchanged emails with two members of the Deep Blue team, Murray Campbell and Joel Benjamin, and they kindly shared several items of interest. Campbell is still working on AI at IBM Research, and is still a chess enthusiast. As such, he says he would like to have seen a third match, and that they had already been working on how to further improve Deep Blue. He corrected contemporary press reports with the surprising news that Deep Blue was kept online in their lab until, he writes, “It was finally powered down in 2001. Half was donated to the Smithsonian (in 2002) and the other half to the Computer History Museum (in 2005).… It was still a respectable supercomputer. We didn’t routinely run the chess hardware on the full system.” More is the pity then, that it was hidden away from an inquiring public. Campbell also told Mig that his favorite part of his decades in computer chess (starting as a student in the late 1970s) was not the 1997 rematch itself, but the preparation for it, because the stress level was so high during the match. If only that had affected Deep Blue’s play as it did mine!
GM Benjamin wrote to contradict his colleague Miguel Illescas’s published recollections about game six, saying that it was he (Benjamin) who entered the fateful 8.Nxe6 move into Deep Blue’s opening book, “a month or so before the match.” That is, not “that very morning” of game six, as emphasized by Illescas with such vigor that this revelation was the headline of his interview. Benjamin said that he didn’t dissent when the interview, and my incredulous response, were published in 2009 due to not wanting to publicly contradict his old teammate. This dispute between twelve-year-old and twenty-year-old human memories is another reason that all of Deep Blue’s files and logs should have been released at the time, especially if it was never to play chess again. By dismantling Deep Blue, IBM killed the only objective witness.
As for me, I moved on. The world still needed a human world chess champion after all, it turned out. I was very disappointed that I was never going to have a chance to get my revenge on Deep Blue. And it was always in the back of my mind that we never got to reproduce all of Deep Blue’s moves for posterity. It was an inverted Agatha Christie whodunit. There was ample circumstantial evidence and no shortage of motives, but it wasn’t clear that there had ever been a crime.
I have been asked, “Did Deep Blue cheat?” more times than I could possibly count, and my honest answer has always been “I don’t know.” After twenty years of soul-searching, revelations, and analysis, my answer is now “no.” As for IBM, the lengths they went to to win were a betrayal of fair competition, but the real victim of this betrayal was science.