AlphaZero's Immortal Zugzwang vs Stockfish — 2017

In December 2017, a neural network that had taught itself chess in a matter of hours sacrificed pawns and pieces against the strongest engine on earth, asking for nothing in return but squares. The result was a positional bind so complete that grandmasters reached back to 1923 for a name worthy of it: the Immortal Zugzwang of the machine era.

A Machine That Learned Alone

On 5 December 2017, DeepMind published a paper announcing that its program AlphaZero had surpassed the world's strongest chess engine, Stockfish, after roughly four hours of training. AlphaZero was given only the rules of the game. No opening book, no endgame tables, no centuries of accumulated human theory. It played millions of games against itself, learning by reinforcement, and emerged with an evaluation of chess that no human and no conventional engine had ever produced.

The contrast in method was stark. Stockfish, the culmination of decades of engine development, examined around 70 million positions per second using handcrafted evaluation rules. AlphaZero searched roughly 80,000 positions per second, a thousand times fewer, guided instead by the judgement of its deep neural network. In the 100-game match reported by DeepMind, played at one minute per move, AlphaZero won 28 games, drew 72 and lost none. Of the ten sample games initially released, one stood above the rest, and it earned a place among the famous chess games of any era, human or machine.

The Pawn That Never Came Back

The celebrated game arose from a Queen's Indian Defence, one of the most solid openings in chess history. Early in the game AlphaZero, playing White, offered a pawn with the thrust d4-d5, a long-term sacrifice that no traditional engine would have chosen, because its compensation could not be measured in material or in any forcing line. The compensation was lighter than air: open diagonals, a knight heading for a dominant outpost, and a lasting cramp in Black's position.

Stockfish took the pawn, as its evaluation said it should. For dozens of moves its assessment insisted Black was fine, even better. Yet AlphaZero never tried to regain the material. It manoeuvred with its knight to the f5 square, fixed Black's pawns, seized the light squares around the black king, and slowly took away every useful move Stockfish had. This was chess strategy in its purest form: not a combination to be calculated, but a verdict to be executed.

The Bind Becomes Zugzwang

Zugzwang, the compulsion to move when every move worsens your position, is usually the property of chess endgames, where kings and pawns run out of safe squares. What made this game extraordinary is that AlphaZero imposed it on a full middlegame board, against an opponent calculating tens of millions of positions per second. Stockfish's queen was driven to the edge of the board and entombed, reduced to shuffling between a pair of squares while AlphaZero improved its pieces at leisure.

By the final phase, almost any move by Black lost material or collapsed the position outright. Stockfish, which is never short of a defensive resource when one exists, simply had none. Commentators immediately recalled Saemisch against Nimzowitsch at Copenhagen in 1923, the original Immortal Zugzwang Game, in which White resigned in the middlegame because every legal move lost. AlphaZero had recreated that famous fate, but inflicted it upon a machine that does not blunder.

How the Chess World Reacted

The reaction among famous chess players was close to awe. Peter Heine Nielsen, the long-time second of world champion Magnus Carlsen, said he had always wondered what it would be like if a superior species landed on earth and showed us how they play chess, and that now he knew. Garry Kasparov, who had lost to IBM's Deep Blue twenty years earlier, wrote that chess had been shaken to its roots, and praised AlphaZero for playing with a dynamic, sacrificial style that mirrored his own.

There was criticism too, and it deserves honest mention. The 2017 match conditions were contested: Stockfish 8 played without its opening book, at fixed time per move, and with hardware settings that some felt understated its strength. DeepMind answered the objections with a peer-reviewed study in the journal Science in December 2018, in which AlphaZero played 1,000 games against an updated Stockfish under more standard conditions and won 155, lost 6 and drew 839. The verdict stood.

The Games That Changed Human Chess

AlphaZero's influence spread far beyond a single match. Grandmaster Matthew Sadler and Natasha Regan analysed the published games for their 2019 book Game Changer, which won the English Chess Federation Book of the Year award, and catalogued the machine's signature ideas: rook's pawn advances against the castled king, exchange sacrifices for long-term pressure, and an almost contemptuous attitude towards material when piece activity was on offer. Within a few years those motifs were appearing in elite tournaments and World Chess Championships preparation, played by humans who had absorbed the lessons.

The open-source project Leela Chess Zero reproduced the approach publicly, and neural network evaluation was eventually built into Stockfish itself. Modern engine analysis, and through it modern opening theory and chess tactics training, descends directly from December 2017. The zugzwang game is the moment that future was first made visible on a chessboard.

A Bind Worthy of the Board

It is fitting that the most advanced chess ever played was still expressed through the oldest of standards. Every diagram of the game, every demonstration board on which it has been replayed, uses the pieces codified in the Staunton chess set, the 1849 design named for Howard Staunton and created by Jaques of London. AlphaZero needed no physical pieces at all, of course. Yet when humans sit down to study what it discovered, they reach for the same broad-based king, the same knight carved with a horse's head, that serious players have used for more than 175 years. The medium endures even as the understanding it carries is transformed.

Jaques of London — Making Chess Sets Since 1849

FIDE regulations mandate the Staunton pattern, originally created by Jaques of London in 1849, for all serious competition. So when humans replay AlphaZero's immortal zugzwang on a real board, they do so with pieces whose design traces directly to the Jaques workshop in London. The oldest standard in chess still carries its newest ideas.

Shop Staunton Chess Sets →

Frequently Asked Questions

Q: What is zugzwang in chess?

Zugzwang is a German term meaning compulsion to move. It describes a position in which a player would be perfectly safe if allowed to pass, but the obligation to make a move forces them to damage their own position. It is most common in chess endgames, where every pawn move and king step matters, but it can also appear in middlegames when one side's pieces are completely tied to defensive duties. AlphaZero's 2017 games against Stockfish showed zugzwang on a grand scale: Stockfish's queen, rooks and minor pieces were so constricted that almost any move lost material or allowed a decisive breakthrough, while AlphaZero calmly improved its position with no need to hurry.

Q: How did AlphaZero learn to play chess?

AlphaZero, developed by DeepMind, was given only the rules of chess and no human games, opening books or handcrafted evaluation rules. It then trained through self-play reinforcement learning, playing millions of games against itself and updating a deep neural network based on the results. According to DeepMind's December 2017 paper, the system reached a level beyond Stockfish after roughly four hours of self-play training on Google's specialised hardware. During play it used a Monte Carlo tree search guided by its network, examining about 80,000 positions per second, compared with the tens of millions per second searched by Stockfish. Its strength came from judgement rather than raw calculation, which is why its play looked so human, and at times so superhuman.

Q: What was the result of the AlphaZero vs Stockfish match in 2017?

In the December 2017 evaluation reported by DeepMind, AlphaZero and Stockfish 8 played a 100-game match at one minute per move. AlphaZero won 28 games, drew 72 and lost none, scoring 25 of its wins with the white pieces. DeepMind initially released only ten sample games, which included the famous zugzwang masterpiece. The match conditions drew criticism, since Stockfish ran without its opening book and under fixed move times. DeepMind responded with a larger study published in the journal Science in December 2018, in which AlphaZero played 1,000 games against an updated Stockfish under more standard conditions and still won decisively, with 155 wins, 6 losses and 839 draws.

Q: Why is the AlphaZero zugzwang game considered immortal?

The label echoes the Immortal Zugzwang Game, the name given to Friedrich Saemisch's loss to Aron Nimzowitsch at Copenhagen in 1923, where White resigned in a middlegame because every legal move lost. Commentators reached for the same phrase in 2017 because AlphaZero achieved something comparable against the strongest calculating machine ever built. Beginning with a long-term pawn sacrifice in a Queen's Indian Defence, AlphaZero gave up material for nothing but squares and piece activity, then tightened its grip until Stockfish's pieces, including its queen, were reduced to shuffling on a handful of squares. Grandmaster Peter Heine Nielsen said the games felt like watching a superior species play chess, and the name has stuck in chess history.