Game Changer

The method behind AlphaZero is where the story turns strange, because it is so much simpler than what came before. There was no library of human games poured in, no encyclopedia of openings, no list of principles about controlling the center or developing pieces. The engineers gave the system the rules — how each piece moves, what counts as checkmate, what counts as a draw — and then set it to play against copies of itself, millions of times over.

At the start, it played randomly, flailing like a beginner who has just learned which way the knight jumps. But after each game it adjusted. A neural network — the same broad kind of pattern-learning system used to recognize faces or translate languages — slowly tuned itself toward the moves that tended to lead to wins. It was, in effect, running its own tournament against itself and learning from every result, with no human ever telling it that a bishop is worth roughly three pawns or that an open file is good. Those ideas, to the extent AlphaZero used them at all, it discovered on its own.

The reason Game Changer runs to hundreds of pages is that Sadler and Regan treat the games as a curriculum, not a curiosity. Freed from the conventional engine's obsession with material, AlphaZero kept doing things human theory had always half-suspected were strong but could never quite trust. It valued the mobility of its pieces above almost everything, happily giving up a pawn if it meant its bishops could breathe and its opponent's could not.

One motif recurs so often that Sadler names it: the far-advanced wing pawn, pushed deep into enemy territory not to promote but to jam the opposing position, taking away squares and tying pieces down. Conventional wisdom treated such pawns as weaknesses to be defended. AlphaZero treated them as long-term investments in space, and it was right often enough to force human players to reconsider. The same held for its handling of the initiative — its willingness to sacrifice material for a lasting attack that might not pay off for twenty moves, if it paid off at all.

Step back from the sacrifices and the pawn pushes, and the deeper claim in Game Changer is about the shape of the intelligence involved. For most of computing's history, the machine's advantage over the human was calculation — speed, tirelessness, the ability to check every possibility. Human beings held the opposite territory: intuition, pattern, the feel for a position that no one could fully write down as a rule. The two seemed like different currencies. AlphaZero blurred the line, because it won the way humans were supposed to be good — by judgment — using a method that produced judgment out of nothing but experience.

That matters because AlphaZero's designers were never really trying to solve chess. Chess was a controlled arena, a closed world with clear rules and clear outcomes, ideal for testing whether a system could teach itself competence without being told the answers. The same DeepMind lineage went on to attack problems with no tidy rulebook at all, most famously the prediction of how proteins fold — a challenge with enormous consequences for medicine and biology, and nothing like a game. The chess story was a proof of concept for a way of learning.

The ten games DeepMind published in December 2017 still hold up as the thing they were meant to be: evidence that a machine given only the rules could, in an afternoon of playing itself, arrive at chess of startling quality. What Sadler and Regan added, with their two thousand games and their access to the team, was the explanation underneath the spectacle — the self-play, the trained sense of where to look, the willingness to prize activity and initiative over the safe accumulation of material that conventional engines had always chased.