Alpha Zero’s achievement in chess is staggering. It showcases a quantum leap for Artificial Intelligence.
If there is one thing that sets human beings apart from other species, it is this: we think too much of ourselves. Just because we lucked upon opposable thumbs and a powerful brain, both of which allowed us to dominate other species, we behave as if we are masters of the universe. It’s pathetic. We’re bawling babies in front of a bacterial onslaught, and we will soon find ourselves inadequate in front of machines that we ourselves will make. It is time for humility.
A few days ago, Alpha Zero beat Stockfish. We humans talk about Ali-Foreman and Federer-Nadal and Fischer-Spassky, but the most momentous match in human history might well have been the chess match between these two machines. But first, some context.
Here’s the Artificial Intelligence context. In 1950, when AI was in the realm of science fiction, Alan Turing came up with the Turing Test. Wikipedia defines this as “a test of a machine’s ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human.” So if you’re having a text conversation with a party you cannot see, a machine would pass the Turing Test if you do not realise that it is a machine. I would hold that AI has achieved this easily, although many humans would probably fail. (Check out Donald Trump’s Twitter feed.)
Here’s the chess context. Until the early 1990s, the thought of a computer beating a human in chess was laughable. But technology progressed quickly, and in 1997 a machine called Deep Blue beat the then-World Champion, Garry Kasparov. Computers soon left humans far behind. Today, a program on your smartphone can thrash the best player in the world.
Now, you’d imagine that this would mean the end of chess. Everyone would use computers in their analysis and pedagogy, and we’d all start playing like machines. But exactly the opposite happened, and chess was instead enriched.
There was once a study that aimed to see how many moves a grandmaster and a novice could think ahead in a game of chess. The answer was that they saw the same number of moves ahead, but the GM saw the right ones. Learning chess is less about calculation and more about pattern recognition and heuristics. The more you play, the more patterns you learn to instinctively recognise, with an understanding of how they interact with other patterns. A strong player can glance at a position on the board and understand its salient aspects.
And then, the heuristics. Heuristics are simple rules that allow people to make decisions. For example, a chess player will be taught that it is important to occupy the center early, to take her king to safety by castling, to develop her pieces as much as she can, and so on. Now, humans cannot possibly calculate everything on the chess board. (The number of possible positions in a 40-move game is greater than the number of electrons in the observable universe.) So they use shortcuts – or these heuristics.
All humans learn chess by learning heuristics. These have evolved over centuries, and are a common body of knowledge that every player has to learn to reach a certain level. The famous Soviet School of Chess was the embodiment of this. Given this common body of knowledge, chess players actually played in a similar way, with individual style appearing on the margins.
Computers did not need heuristics, because they had the computing power to actually calculate every move and every position. (This is called ‘brute force’.) This did not make chess more homogenous, but less, as computers looked beyond the set of heuristics that were instinctive for players. This meant that the new generation of players who used chess programs as an analytical tool were no longer bound to the dogmas of the past, useful as they were. All the principles earlier generations had learned had exceptions, and all the exceptions could be explored using these programs.
As a result, the current generation of players has more stylistic variation than ones before. Younger players think about the game in unique ways that older ones can’t fathom, and is outside their playbook. And while all top players use programs like Stockfish for analysis, none of them plays games against it because Stockfish would thrash them, and it would be too demoralising. It’s like trying to race a car.
So what did Alpha Zero do? Well, Alpha Zero was built by Deep Mind, an AI division of Google. It is a self-learning program, and the rules of chess were fed into it, but nothing else. No opening databases, no heuristics. It played against itself for four hours to learn the game. Then it played Stockfish in a 100-game match. Alpha Zero won 28 games, and the rest were drawn. After four hours of learning, it beat a chess program into which years of development had gone.
Astonishingly, Alpha Zero achieved this by playing like a human. While Stockfish examined 70 million positions per second, Alpha Zero looked at only 80,000. While teaching itself chess, it discovered, developed and then used heuristics that seem to go beyond the ones humans discovered. For example, human are taught not to move the same piece multiple times in the opening when others lie undeveloped. Alpha Zero did this again and again, favouring activity over development. It also made long-term positional sacrifices, with no immediate gain, which machines otherwise do not do.
The games released by Alpha Zero are spectacular. Alpha Zero plays like a human, but an enhanced human. The grandmaster Peter Heine Nielsen, Magnus Carlsen’s coach, told chess.com: “After reading the paper but especially seeing the games I thought, well, I always wondered how it would be if a superior species landed on earth and showed us how they play chess. I feel now I know.”
The implications of the Deep Learning that Alpha Zero demonstrates are fantastic and unfathomable, and not just for chess. AI is already embedded in our lives – your smartphone would have seemed like science fiction in 1990 – and will become more so. It has become fashionable to be worried about AI, but I am optimistic. Technology will make us all better versions of ourselves – and that journey begins by accepting that we aren’t all that awesome to begin with.