Doomed to Draw

Ben Jackson

  • The Grandmaster: Magnus Carlsen and the Match that Made Chess Great Again by Brin-Jonathan Butler
    Simon and Schuster, 211 pp, £12.99, November 2018, ISBN 978 1 982107 28 4
  • Game Changer: AlphaZero’s Groundbreaking Chess Strategies and the Promise of AI by Matthew Sadler and Natasha Regan
    New in Chess, 416 pp, £19.95, January, ISBN 978 90 5691 818 7

If you know anything about Magnus Carlsen, you probably know that he is supposed to be making chess cool. Before he was twenty, he was the subject of two books and a film; in the years since – he’s now 28 and the world’s best chess player – he has been one of Cosmopolitan’s sexiest men and one of Time’s hundred most influential. He is imposingly good-looking; it’s impossible not to be impressed by his quiff, even if his face always looks slightly swollen, as if he’s coming off an especially bad night’s sleep or a mild allergic reaction. His cool sometimes seems of the high school jock variety. Before the 2018 World Chess Championship, he tweeted a video claiming that his preparation involved ‘the three Ps’: pizza, Premier League and poker. On a rest day, he went off to play football and reappeared the next day with a black eye. Inscrutable over the board, he has difficulty hiding his emotions the rest of the time: you will always know if Carlsen is bored, and he is easily bored. He has no time for traditional education, which is understandable enough, but he also seems to take a relatively casual approach to chess. He claims to lie in bed until just before lunch and fired his one-time coach Garry Kasparov, arguably the greatest player of all time, for being too intense. ‘He does what he likes,’ his father explained proudly to the New Yorker. ‘It’s curiosity as opposed to discipline.’

The second of four children, Carlsen was born in Tønsberg, the oldest town in Norway, in 1990. His soft-spoken father, Henrik, introduced him to chess when he was five, after noticing his powerful memory and capacity for concentration. Henrik calls himself an ‘ego dad’, but when Magnus at first failed to take to the game, Henrik didn’t force things. Accepting that his children were ‘definitely not geniuses’, he allowed Magnus to focus on football and skiing. At home, the sibling rivalries were intense. It was partly the desire to defeat his elder sister that enticed Magnus back to chess a few years later. This time, his interest became all-encompassing. He abandoned schoolwork so that he could study the game and prepare for tournaments. He developed a taste for breaking egos (‘I enjoy it when I see my opponent really suffering’) and, in a country not renowned for its chess players, he was quickly identified as having serious potential.

In his early years, Carlsen didn’t have the grandmasterly assistance enjoyed by many chess prodigies. His trajectory was extraordinary. In a single year, his Elo rating rose from 904 – only slightly better than moving the pieces at random – to 1907, enough to be among Norway’s top four hundred or so players. When he was 12, his father requested a year’s leave of absence to take him on a chess-playing tour of Europe, beginning a career as a kind of valet to his son (so far as I know, Magnus still lives with his parents – or the other way round). As a teenager, Carlsen developed an aggressive, sacrificial style, happy to give up pieces in exchange for retaining mobility and initiative. At 13, he drew a game with Kasparov. At 18, he became the youngest ever world number one. In 2013, he took the World Chess Championship from Viswanathan Anand. ‘It would be a bigger deal if I hadn’t won,’ he said.

By this time, Carlsen’s style had evolved into one favouring positional control over dramatic attacks. It is often compared to Anatoly Karpov’s ‘boa-constrictor’ technique; as Kasparov described it to the New Yorker as ‘strangling pressure, not direct hits’. The word most commonly associated with Carlsen’s play is ‘intuitive’, which I take to mean that he doesn’t rely on preparation, particularly computer preparation, as much as his rivals do. He will sometimes pick unusual or apparently poor moves just to complicate the position or knock an opponent out of lines that have been prepared in advance with the assistance of a computer (these days, you can download a chess app on your phone that can handily defeat a grandmaster). In the documentary Magnus, Jon Ludvig Hammer, who assisted Carlsen during the 2013 Anand match, suggests that his style depends on getting players away from prepared moves, ‘so that it’s a battle of the mind, rather than a battle of who can use the computer best’. This has continued to pay dividends. He defended his title in 2014, again defeating Anand; and then in 2016 against Sergey Karjakin, a Crimea-born Russian who performed above expectations but still fell to the inevitable; and last year he saw off Fabiano Caruana to retain the title.

Brin-Jonathan Butler’s The Grandmaster, which focuses on the Karjakin match, doesn’t provide any evidence that it ‘made chess great again’, despite the book’s subtitle. The match left the peak of the chess world undisturbed. The main surprise was that Carlsen failed to come out on top over the 12 games played under lengthy classical time controls, meaning that he required the rapid games to retain his title (according to the present World Championship format, the time controls become progressively shorter the longer the players remain tied). The match was all but won by the time it happened, but the most spectacular moment was Carlsen’s queen sacrifice in the tiebreaker. Such moments of grace, which appear to come as a kind of miracle, rather than as a result of logical planning, are what make chess exciting for me, even if a player like Carlsen finds the accumulation of small advantages more satisfying. As I read this book, I started to realise what a hopeless task Butler had been assigned. Carlsen refused to spend any time with Butler, so he has to rely heavily on existing journalism. He sets out to illuminate Carlsen’s genius, but the paucity of new material forces him to discuss matters that seem irrelevant or outdated. Chapters are devoted to Bobby Fischer, Judit Polgar, chess hustlers and Josh Waitzkin (the subject of the film Searching for Bobby Fischer). Returning to his main subject, he concludes that ‘the secret’ to Carlsen’s greatness is that ‘it remains a secret,’ whatever that means.

*

During the last few weeks, I’ve spent a lot of time watching Carlsen play blitz (five-minute) and bullet (one-minute) chess online. He says he does this to relax and to rejuvenate his ego by thrashing wired bunglers after he’s lost a serious tournament game. There is, obviously, a large element of performance art in these streams, but they provide a rare chance to see Carlsen at his ease around chess. Sometimes, he drinks beer, listens to hip-hop and makes bad jokes with his friends. While he waits for an opponent’s move, he clicks rapidly, over and again, on the piece he has just moved, full of antic energy. When I play, I descend into a dreadful mood, provoked by losing or by the site’s chatbox, in which you are expected to belittle weaker opponents. Carlsen loses a surprising number of games, but he shrugs them off, accepting that sometimes an experiment fails.

Some players specialise in particular time controls, but Carlsen is formidable in all of them: he is currently the top-ranked player in classical, rapid and blitz chess. If he were a runner, he would win the 100m and the 5k, but would be best-known for his marathons. His coach, Peter Nielsen, describes him as an ‘incredibly annoying opponent’ because he is happy to grind away for hours at positions most top players would accept as a tie. He often upbraids colleagues who agree to early draws, arguing that they should play out the game. Once, he was offered a quick draw by Hikaru Nakamura. Carlsen played on for five hours, only to wind up with a draw anyway. He criticised the competitors in the 2012 World Championship for agreeing to draws in positions he evaluated as playable. He even suggested that, as the then world champion, Anand had started to fear loss more than he craved victory. All of which is to say: Carlsen doesn’t like quitters.

Still, Carlsen’s first 11 games in the 2018 World Championship against Caruana finished as draws. In itself, this wasn’t a story: draws have long been the most common result between top grandmasters. The real shock came in the 12th and final classical game. Taking an admirably active approach, Caruana, who had the white pieces, castled on the queen’s side in an attempt to target Carlsen’s king by opening up the king’s side. This manoeuvre was a bit of a gamble. Though he had prepared a defence with an unusual rook move, he had exposed his king to Carlsen’s active queen. Over the course of the next few moves, Carlsen managed to bring his rook, knight and dark-square bishop into the attack. He advanced a pawn close to Caruana’s king, kept his own king out of danger, and had an extra half hour on the clock. It wasn’t a clear win for Carlsen, but he was certainly ahead.

Rumours then began to circulate that Carlsen had offered a draw. I was watching Chess24’s stream, where analysis was provided by several grandmasters, none of whom believed a word of it. ‘I think it’s extremely unlikely,’ Peter Svidler said. Anish Giri – one of a small number of players to become a grandmaster at the age of 14 – added: ‘I never heard these words coming out of Magnus’s mouth.’ We sat, waiting, while the players settled in for a good long think. Finally, the news was confirmed, and Giri lost it completely. ‘When his life is on the line, suddenly he’s not Magnus any more! I have never, never seen Magnus offer a draw in a position like this … I think it is some sort of a nervous breakdown, really.’ The thought was widespread. Kasparov tweeted that Carlsen had lost his nerve and was no longer the favourite for the championship. Vladimir Kramnik, a former world champion, believed that Carlsen had lost his ‘fighting spirit’. But Carlsen had made a simple calculation: he and Caruana were, plainly, well matched under classical time controls; yet while he was the top-ranked player in rapid chess, Caruana was only eighth. In other words, here was his opportunity to win.

In the end, his decision was vindicated. He put in a dominant performance in the tiebreaker, winning three rapid games in a row to retain his title. After the championship, his streak of draws in classical play continued until he reached 21 in a row, beating Giri’s record of 20 (to be fair, he won his next game, and the tournament in which he had set the record, and has since been playing some of the best chess of his career). One possible explanation for the streak is that Carlsen’s diagnosis of Anand’s play as world champion – that he was more afraid of loss than desirous of victory – has come to apply to himself. But a more likely explanation is simply that chess players are getting harder to beat. This is mainly because of the influence of the supreme defenders, computers, which have repeatedly shown that positions that appear to be lost are defensible (if only one move can save a position, a computer’s sheer calculating ability ensures that it will find it). As a result, professional players have become more resilient, and more sceptical of daring, sacrificial chess. ‘Draw death’ seems to loom over the game, much as it did in the 1920s, when José Raúl Capablanca, the great Cuban player, proposed making the game more technically challenging and draws less likely by expanding the board and including two new pieces: a chancellor, which would move as both a rook and a knight; and an archbishop, which would move as both a bishop and a knight. Gone are the days of Kasparov’s dynamic attacks, the story goes. Instead, thanks to computers, we have entered a period of solid, overprepared chess, doomed to draw after draw.

*

Chess has always been a means to an end for computer programmers. The first chess-playing program was written on slips of paper by Alan Turing and David Champernowne in the late 1940s. At the time, it was easy to think that a computer capable of playing a decent game would require a kind of general reasoning capacity, that it might need to make logical deductions, think strategically and learn abstract concepts. ‘If one could devise a successful chess machine,’ Herbert Simon and others suggested in 1958, ‘one would seem to have penetrated to the core of the human intellectual endeavour.’ Combine this with its clear measures of success and readily formalised rules, and chess seemed to offer the perfect testbed for artificial intelligence.

In principle, chess can be played perfectly. There are a finite number of possible games, each of which must eventually end in a win, draw or loss. If you could lay out all the possible moves, you could work backwards from the result and find the optimal move in any given position. No one has calculated the exact number of possible chess games – it’s so huge it isn’t worth the computational effort – but Claude Shannon, one of the pioneers in the field, in 1950 ventured a conservative guess at the number of ‘typical’ games. Given an average of forty moves for each player per game, with an average of thirty possible legal moves in any given position, there would be around 10120 possible ways for a game to play out. Here is a back-of-the-envelope calculation, adapted from a paper by Nathan Ensmenger. The world’s most powerful supercomputer, IBM’s Summit, can perform two hundred quadrillion calculations per second, or 2x1017. If every atom in our galaxy (roughly 1067) were a Summit, and if they had been working for every second since the Big Bang (roughly 1018 seconds), they still wouldn’t have finished assessing these possibilities. Weirdly, the practical unfeasibility of solving chess is a stroke of luck for AI researchers. Raw number-crunching isn’t what we mean when we speak about intelligence. The challenge is to work out a way through the search space, as human players do, without testing every possibility; only our failure to reach perfection opens the window for creative intelligence to be applied.

Shannon proposed two approaches to the problem. ‘Type A’ brute force programs would investigate every possible line as thoroughly as they could, given the limits of time and computing power, and evaluate each move in order to select the best at any point. ‘Type B’ programs would play more like humans, concentrating on the most plausible lines and dedicating their resources to those. Six years later, the MANIAC I – designed at Los Alamos and, with 2400 vacuum tubes, weighing half a ton – became the first computer to beat a human at something resembling chess. It operated a Type A strategy, but its memory was so limited it had to play on a reduced 6x6 board, with no bishops. In its first game, a strong player competed without his queen and still beat it handily. After that, a novice lost in 23 moves, a historic achievement for computer chess.

And then, in 1958, Allen Newell and Herbert Simon invented a mechanism known as alpha-beta pruning, a search algorithm which ceases to investigate a particular line after at least one possibility has been found that proves it to be worse than a previously evaluated line. This approach was relatively easy to program and its performance relied so much on speed that the results steadily improved as computers got faster. For artificial intelligence, this was disastrous. Researchers abandoned Type B approaches altogether, along with Newell and Simon’s goal of using chess in a ‘deliberate attempt to simulate human thought processes’; instead, they just built faster hardware. Plenty of finesse still went into crafting the evaluation function and the openings, but in essence computer chess became an engineering project, with no application beyond winning games. By the time IBM’s Deep Blue beat Kasparov in the famous 1997 match, Joe Hoane, the principal software engineer on the team, said: ‘It is not an artificial intelligence project in any way … We play chess through sheer speed of calculation and we just shift through the possibilities and we just pick one line.’ (They weren’t in it only for the chess: in the days following the match, the publicity helped to increase IBM’s value by more than $11 billion.)

Computers designed like this can defeat any human, but they don’t produce very interesting chess. As Carlsen puts it, they ‘are really good tactically and they can’t play chess’. Their games usually include plenty of tedious manoeuvring and not much grand strategy. I’m often reminded of a remark by Clive James, writing about Game of Thrones: ‘Where contending forces are invincible, there can be no plausible conflict, only choreography.’ There is no drama when computers play, but their tactical insights, especially their ability to weasel out of tricky positions, influenced a generation of players. Even weak players like me feel their influence, if only because an engine is instantly available to tell us where we blundered. These days, Kasparov wrote, ‘a move isn’t good or bad because it looks that way or because it hasn’t been done that way before. It’s simply good if it works and bad if it doesn’t … Humans today are starting to play more like computers.’

*

On the eve of the World Championship last year, a Norwegian website released a chess program named Caruana, which was designed to pick the worst move in any position. I beat it inside twenty seconds. Caruana first gave up one knight, then the other, before sacrificing its queen for no reason and in the same move opening itself up to checkmate. All good fun, but there was no mistaking this for human play. In fact, no computer program that I know of can play convincingly bad chess – that is, humanly bad chess. Computers programmed to play weakly tend to alternate between world-champion-level moves and ridiculous blunders, as if finding themselves utterly at a loss as to how to let you win. For years, human chess has influenced computers and computer chess has influenced humans, but neither can really play like the other. Humans rely on pattern recognition, visualisation and the ‘feel’ of the position. Computers rely on raw calculation combined with preprogrammed rules and heuristics.

AlphaZero, a generic game-playing algorithm designed by Google’s DeepMind, is something different.[*] AlphaZero doesn’t use alpha-beta pruning, and is given no human knowledge except the rules of the game. Instead, it works out the best strategies by playing millions of games against itself. When a game ends as a win for one side, it shifts the parameters of its neural network to show that the positions encountered are more promising for the winner, and the moves chosen by the winner are more likely to be the best ones. In future games, it focuses its evaluation energies on moves that have proven successful before, or that bring the game towards board positions that look good, a method of prioritising so effective that when AlphaZero finishes ‘training’ it only searches eighty thousand positions per second, compared to Deep Blue’s two hundred million. This also means that it can direct a game towards positions that ‘feel’ more generally promising based on past experience. Taken as a whole, it seems to get closer to human intuition and to Shannon’s Type B approaches.

Game Changer by Matthew Sadler and Natasha Regan is primarily a book about chess strategy, but it also contains interesting non-technical descriptions of AlphaZero’s design and uninteresting interviews with people like Demis Hassabis, the founder of DeepMind, which you should blow straight past if you want to avoid advice such as ‘life’s so short, ideally you want to do activities that have more than one purpose.’ As a chess player, AlphaZero is extraordinary. After ‘learning’ the rules of the game, it trained for just nine hours, playing 44 million games against itself, before crushing the world-champion program, Stockfish, in a hundred-game match (Tord Romstad, one of Stockfish’s original authors, has complained about the conditions, but that’s by the way). Sadler and Regan pull all kinds of interesting things out of its play. For one thing, in many positions rated as equal by traditional engines, AlphaZero gave itself the advantage. As a result it surprised the other programs, seeing chances they missed. Some of these surprises probably came because AlphaZero learns for itself, and is not wedded to the conventional wisdom programmed into other chess engines. It also plays with the dynamic, sacrificial style that many people thought had been refuted by engines like Stockfish. Kasparov, whose style is quite like AlphaZero’s and not at all like Stockfish’s, wrote: ‘Programs usually reflect priorities and prejudices of programmers, but because AlphaZero programs itself, I would say that its style reflects the truth.’ Whether this is tongue in cheek, I don’t know, but what we do know is that the same algorithm has been successfully applied to Go and Shogi (Japanese chess), and that DeepMind has used similar techniques to control the cooling systems in Google’s datacentres, reducing energy use by up to 40 per cent.

It seems uncharitable not to feel excited about all this. But I can’t help having my doubts. For one thing, the sheer resources required to create something like AlphaZero – and DeepMind’s refusal to share any of the code – mean that it is probably just another tech project that will concentrate the benefits of progress in the hands of a small elite. For another, it’s still not clear that many real-world situations can be productively reduced to a process of optimisation, with a unitary goal and a predefined set of rules.

Hassabis has said DeepMind’s goal is ‘solving intelligence, and then using that to solve everything else’. The hubris of this is not only that it overestimates the potential of AI to do stuff, setting the bar so high that we can’t help but be disappointed by the results, but that it disregards the fact that people disagree about what stuff to do. The obvious question, given our deep political differences, is ‘solving everything for whom?’ And how do we ensure that crucial decisions about the future are not made solely by programmers and others within the tech industry, at the expense of the rest of us? One common response, at least to the former question, is that we should let computers ‘solve’ these political and ethical issues for us too. As Nick Bostrom put it in his 2010 book Superintelligence, which addressed the subject of machines surpassing human intelligence, ‘since the superintelligence is better at cognitive work than we are, it may see past the errors and confusions that cloud our thinking.’

That’s hard to argue with, but I can’t help thinking about the influence of computers on chess players. They see past our errors and confusions on the board, and they certainly help human players to win. They also have the virtue of offering non-binding advice: players retain the final decision over which move to make. But a computer is a monomaniac: it chooses moves that it calculates will help it win, no matter how ugly they are. And they have led to an impoverishment and a homogenisation of style. If such features were replicated in political decision-making, the consequences could be much greater. More to the point, much of our politics today revolves around the perception that decisions are being taken elsewhere, whether in Westminster, Brussels or Washington. Passing off the work of decision-making to the ultimate aloof elite, a computer, is not a serious way of confronting this issue. Sometimes it’s important to decide things for ourselves, and to feel like we’re deciding, even if we often go astray.

I think this is the reason I’m a Carlsen fan. No doubt he does plenty of computer preparation – you can’t survive in professional chess without it – but he seems constantly to be looking for ways to reduce the influence of computers, to pull the game into positions where intuition and judgment come to the fore. ‘When I covered the Kasparov-Deep Blue match,’ Steven Levy has written, ‘I thought the drama came from a battle between computer and human. But it was really a story of people, with brutal capitalist impulse, teaming up with AI to destroy the confidence and dignity of the greatest champion the world had seen.’ Carlsen may not be able to defeat computers at chess, but he keeps finding ways to surprise the people who team up with them.

[*] Paul Taylor wrote about AlphaGo – an antecedent of AlphaZero – and DeepMind in the LRB of 11 August 2016.