by Valentin, June 10 2021, in books

TL;DR: Deep Thinking (Gary Kasparov, 2017)

  • Chess engines:
    • In average, there are 35 legal moves available when it's someone's turn to play. A skilled human player is able to immediatly spot 3 or 4 moves that are "plausible" and then study them to elect the one they think is the best. The intelligence that allows them to spot those 3 or 4 candidate moves is not well understood, so the machine has no other choice than studying all the available moves. Basically, a machine plays with a branching factor that is tenfold this of a human player. Chess engines do have various pruning methods, although they're not necessarily reflective of human reasoning, nor as effective.
    • Both humans and computers need to be able to evaluate a given snapshot of the board to determine what advantage/disadvantage this position holds; this is called the evaluation function. Humans' evaluation function is far more sophisticated than computers'. The reason computers' evaluation function is weaker is because 1. We don't know how to turn all this kind of human intelligence into programming and 2. Even if you add some intelligence, the evaluation function costs more time, so the engine can't go so deep. There is a trade-off to make between the sophistication of the evaluation function and the depth of the exploration. Modern engines go so deep that the evaluation function doesn't need to be that much sophisticated.
    • The only chess skill that a machine has is the tactical skill, which is the ability to calculate a sequence of moves in advance ("if I attack his Queen, he'll need to move her, the only available square is here, at which point I'll be able to put my knight here, but this is defended by his bishop, unless I advance my pawn here", etc). In skilled human play, the game is driven by a high level strategic skill, and tactics are used to enable the strategy at each step of the way. The intelligence to strategize is not well understood and haven't been programmed into machines. So computers play without strategy.
    • How do computers can win with having a branching factor handicap, an evaluation function handicap, and a strategy handicap? => By adding more processing power. No need to care about the branching factor if you have enough CPUs to handle it, and no need to care about strategy if you extend your tactics to 18 moves ahead.
    • The improvement of chess engines had nothing to do with resolving the mysteries of human cognition. It had everything to do with Moore's law (computing power doubles every 2 year) as well as various algorithmic improvements, most notably in pruning.
    • In 2017, AlphaZero was the first neural-network-based chess engine to beat the top brute-force-based search engine (Stockfish), after 9 hours of training based on self-play. Its gameplay was described as "alien" by top humans, and shows how far humans are from actually playing top chess.
    • Endgame tablebases contain all possible combinations of endgames where only a limited number of pieces remain on the board. The biggest tablebases, the Lomonosov tablebases, compiled by the university of Moscow, know the perfect outcome of any position involving 7 pieces of less.
  • Kasparov vs. Deep Blue:
    • Although the most publicized event (Kasparov's defeat) happened in 1997, it was actually the rematch of a first edition that happened in 1996, in which Kasparov beat Deep Blue.
    • In the 1996 event, there was great collaboration between IBM and Kasparov, and the whole match was more of a science project than anything else. When IBM saw the potential in publicity and organized a rematch for 1997, they switched to battle mode, cut communications with Kasparov, hired grandmasters to help them improve Deep Blue, and refused to publish any game from Deep Blue prior to the rematch.
    • Because chess engines have no notion of strategy and their search horizon at the opening can't see beyond the middlegame, the way to beat them is to play a very defensive and passive game to force them to attack, at which point weaknesses will appear in their structure, that will be exploitable in the endgame. The problem then is that chess engine are absolute beasts at defending endgames; if there is one single combination that allow them to survive, they will find it; if the opponent make one single mistake, the outcome could be entirely reversed.
    • This put Kasparov in inherent difficulty, because Kasparov's natural style was to attack vigorously. Restraining himself from playing is own style was psychologically draining.
    • Kasparov won the first game using this strategy, with the white piece. However, he lost the second game, which psychologically destroyed him, for multiple reasons:
      • Because this time Deep Blue actually attacked quite vigurously and solidly, and Kasparov became entangled in his own passive defense, which was absurd as he was the one supposed to have an aggressive style.
      • Because Deep Blue was playing quite differently than the day before, which indicated that custom tuning was taking place between games, and Kasparov started to become suspicious that one move in particular could be coming from a human and not an engine.
      • Because he resigned on move 45, angry at himself for having been directed to a losing position, without analysing the position well enough. There was in fact a perpetual check that could have allowed him to draw the game. His team informed him of that the following morning, which devastated him.
    • Games 3, 4 and 5 were drawn.
    • In game 6 (the final game), Kasparov volontarily played a dubious move which was risky, but only if Deep Blue was able to realize that it needed to sacrifice a knight to take advantage. Kasparov believed that the engine was not powerful enough to find it. Yet Deep Blue played the knight sacrifice, which allowed him to wreck Kasparov's defense. Kasparov resigned. It was later revealed that the engine was not powerful enough to compute this move, but engineers hard-coded it the very morning of this game as they were analysing this specific variation with grandmasters.
  • Opinions on technology:
    • Kasparov is a proponent of automation, which expands our intellectual horizon by making us able to focus on things on a higher intellectual plane. In the specific domain of chess, chess engines did not made the game uninteresting, but to the contrary more interesting. The generation of grandmasters who grew up with engines are the strongest chess players in history because of the insights they get from the engines.