How DeepMind's AlphaZero taught itself chess in nine hours, crushed the world's strongest engine, and rewrote what grandmasters thought they knew about the game.
Two matches, two decisive victories. AlphaZero didn't just beat the strongest chess engine in history -- it did so while examining 1,000 times fewer positions per second, relying on understanding rather than brute computational force.
In twelve 100-game mini-matches starting from the 12 most popular human openings, AlphaZero won 290, drew 886, and lost only 24 games. It also maintained superiority even when given 10-to-1 time disadvantage against Stockfish.
| Metric | AlphaZero | Stockfish |
|---|---|---|
| Positions evaluated/sec | ~60,000 | ~60,000,000 |
| Search approach | Neural-guided MCTS (selective) | Alpha-beta (exhaustive) |
| Evaluation function | Deep neural network (learned) | Hand-crafted heuristics |
| Opening knowledge | None (discovered through self-play) | Extensive opening book available |
| Endgame tables | None | Syzygy tablebases |
| Training games played | 44 million (self-play) | N/A (hand-tuned over decades) |
Despite searching 1,000x fewer positions per second, AlphaZero's neural network allowed it to focus on the most promising variations -- evaluating fewer but better positions. As Kasparov observed: "It's the embodiment of the cliché, 'work smarter, not harder.'"
DeepMind's Demis Hassabis called AlphaZero's style "alien" because "it doesn't play like a human, and it doesn't play like a program. It plays in a third, almost alien, way." Here are the hallmarks that stunned the chess world.
AlphaZero willingly sacrificed pawns, exchanges (rook for minor piece), and even queens not for immediate tactical gain but for long-term positional compensation that wouldn't materialize for 20-30 moves. Traditional engines would never consider such moves because their evaluation functions penalize material deficit immediately.
Where Stockfish counted material, AlphaZero measured harmony. It would reroute knights through many intermediate squares to reach optimal outposts, coordinate bishops on long diagonals, and keep rooks on open files -- building "a beautiful position" before striking. Piece coordination was valued above material count.
AlphaZero's signature: pushing the h-pawn to h5, then h6, pinning the enemy king to the back rank. This restricted the opposing king's mobility and created long-term attacking chances on the kingside. The whole game then became about opening lines or getting a rook to the back rank. This concept was later adopted by Carlsen, Caruana, and other top GMs.
As Matthew Sadler and Natasha Regan identified, a core AlphaZero theme was "restricting the movement of the enemy king." Even when giving up material, AlphaZero would limit the opponent's king to a corner, then methodically build pressure -- what analysts called a "positional boa constrictor approach."
Advanced lone pawns deep in the opponent's position that created permanent weaknesses and cramped the enemy pieces. In one famous game, a thorn pawn resulted in Stockfish's queen being comically trapped on h8. Both AlphaZero and its open-source successor Leela Chess Zero independently developed this strategy.
GM Maxime Vachier-Lagrave noted: "The biggest edge is that the horizon effect for AlphaZero... AlphaZero was winning and Stockfish was like: everything's fine for like 20 moves." AlphaZero's neural network could foresee strategic consequences far beyond Stockfish's search depth, winning games that Stockfish thought were equal.
Chess historians immediately drew parallels to the Romantic Era of chess (1850s-1880s), when players like Paul Morphy and Adolf Anderssen played with daring sacrifices and aggressive attacks. AlphaZero seemed to rediscover this style at a superhuman level -- proving that the creative, sacrificial approach wasn't "refuted" by modern engine analysis but was, in fact, objectively correct when played with sufficient depth. It played like "watching Morphy play Kasparov using the mind of Lao Tzu."
From queen sacrifices to immortal zugzwangs, these are the games that made the chess world gasp. Click to expand each game's analysis.
Perhaps the most famous AlphaZero game. In this Queen's Indian, AlphaZero sacrificed a pawn to gain dark-square control, then methodically strangled Stockfish until it had no useful moves. Stockfish's queen ended up trapped and useless while AlphaZero's pieces dominated every diagonal and file.
AlphaZero chose the Evans Gambit -- a 19th-century romantic opening abandoned at the highest level decades ago -- and used it to crush the strongest engine in history. The message was unmistakable: the old attacking chess was alive, it just needed to be played at a superhuman level.
In this stunning game from the original 2017 match, AlphaZero placed its king on d2 at move 16 -- an unconventional placement that baffled analysts. Then came the devastating bishop sacrifice on g6 that created a "deadly bind," winning the game 20 moves later.
AlphaZero played the Fried Liver Attack -- 6. Nxf7 -- a knight sacrifice that has been known since the 16th century. Against the strongest engine ever built, using a children's tactical trick, and won in 55 moves with ruthless precision.
In this Queen's Indian, AlphaZero made a knight sacrifice that generated seemingly nebulous attacking chances. No traditional engine would consider the sacrifice sound -- the compensation was too abstract, too long-term. But AlphaZero's neural network saw the truth: Stockfish's position was slowly dying. The game lasted 56 moves.
The positional boa constrictor at its most extreme. AlphaZero used bishop coordination and strategic squeezing to gradually paralyze every one of Stockfish's pieces. By the end, Stockfish had no useful moves to make -- every piece was frozen in place while AlphaZero's pieces roamed freely.
Starting from zero knowledge, AlphaZero discovered, tested, and abandoned openings purely through self-play. Its final preferences reshaped how grandmasters think about opening theory.
In the Schliemann Variation of the Ruy Lopez, AlphaZero found "Bb7 at some stage, castling queenside and throwing the h-pawn forward -- something that has never been seen before," according to Matthew Sadler. AlphaZero independently rediscovered many standard openings but also found entirely novel approaches that no human or engine had considered.
The release of AlphaZero's games sent shockwaves through professional chess. From awe to skepticism, every top player had something to say about the machine that taught itself chess in nine hours.
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.
Programs usually reflect priorities and prejudices of programmers, but because AlphaZero programs itself, I would say that its style reflects the truth.
I can't disguise my satisfaction that it plays with a very dynamic style, much like my own!
To watch such a strong programme like Stockfish, against whom most top players would be happy to win even one game out of a hundred, being completely taken apart is certainly definitive. The chess world will get scrambled.
The more relevant thing is it figured everything from scratch. That is more scary and promising if you look at it.
In essence I have become a very different player in terms of style than I was a bit earlier, and it has been a great ride.
It's like discovering the secret notebooks of some great player from the past.
I was shocked. This is the new big thing. It totally changes chess. It might be rated, what, 3700? Close to 4000? That's really crazy.
I am very much surprised because we normally work with Stockfish... if we have a program which beats Stockfish so easily it might be a new generation for computers. I will pay very much to get access to this program. Maybe $100,000, today!
Of course the result is extremely impressive; I wouldn't even dream of winning one game against Stockfish. It's quite exciting in a way because you can see that chess is definitely not as drawish maybe as we thought.
I was amazed. I don't think any other engine has shown dominance like that. I think it was four hours of learning so who knows what it can do with even more.
AlphaZero is basically using the Google supercomputer and Stockfish doesn't run on that hardware; Stockfish was basically running on what would be my laptop. If you wanna have a match that's comparable you have to have Stockfish running on a supercomputer as well.
The match results by themselves are not particularly meaningful because of the rather strange choice of time controls and Stockfish parameter settings. Stockfish vs AlphaZero is very much a comparison of apples to orangutans.
GM Matthew Sadler and WIM Natasha Regan were granted unprecedented access to over 2,000 unpublished AlphaZero games and the DeepMind team itself. Their book, with a foreword by Garry Kasparov, won both the ECF Book of the Year 2019 and FIDE's Averbakh-Boleslavsky Award.
AlphaZero "likes 1.d4 and 1.Nf3" and "prefers" these over 1.e4. As Black, it chose 1...e5 rather than the Sicilian or French. It found novel ideas in nearly every opening system, including castling queenside in the Schliemann and throwing the h-pawn forward.
Regan shared how studying AlphaZero shifted her play: "I have learnt a bit more about pushing the h-pawn, restricting the king, opening the lines." The rook's pawn advance to h6 against a castled king became AlphaZero's signature weapon.
Sadler revealed that studying AlphaZero "changed the way I think about stuff," giving him confidence in positions he previously doubted, sometimes even defying engine evaluations. The strategic themes are "not so difficult to implement and quite intuitive."
Both professionals and club players can improve by studying AlphaZero's discoveries in "opening preparation, piece mobility, initiative, attacking techniques, long-term sacrifices and much more." The key distinction is between tactical sacrifices (hard to copy) and strategic themes (intuitive and usable).
AlphaZero's games didn't stay in a DeepMind lab. They reshaped how the world's best players think about and play the game.
14th World Champion Vladimir Kramnik collaborated with DeepMind to use AlphaZero as a laboratory for testing chess variants -- work that could shape the future of chess itself.
| Variant | Rule Change | Key Finding |
|---|---|---|
| No-Castling | Castling completely disallowed | 89% decisive games in human tournament |
| No-Castling (10) | Castling prohibited for first 10 moves | More complex opening play, reduced draws |
| Torpedo Chess | Pawns can move 1-2 squares from anywhere; en passant anywhere | Dynamic, highly tactical positions |
| Semi-Torpedo | Pawns can advance 2 squares from 2nd or 3rd rank | Increased pawn mobility, new tactical motifs |
| Pawn One Square | Pawns move only one square forward | Slower positional games |
| Stalemate = Win | Stalemate counts as victory | Changed endgame dynamics dramatically |
| Pawn-Back | Pawns can retreat one square (to 2nd/7th rank only) | More flexible pawn structures |
| Pawn-Sideways | Pawns can move laterally one square | Novel defensive resources |
| Self-Capture | Players can capture their own pieces | Surprisingly deep strategic implications |
The most promising variant. When a No-Castling tournament was held at the Kramnik-Gelfand Training Camp in Chennai (avg Elo: 2457), only 3 of 27 games ended in draws -- an 89% decisive game rate. King safety became a permanent concern, producing "simultaneous attacking and counter-attacking" and far more creative chess. The paper was co-authored by Kramnik, Demis Hassabis, Nenad Tomasev, and Ulrich Paquet.
AlphaZero was never released to the public. But the chess community built its own version from scratch -- and it inherited the alien aesthetic.
An open-source neural network chess engine inspired by AlphaZero's architecture. Instead of Google's TPU cluster, Lc0 uses distributed computing -- volunteers generate self-play games on their own hardware, creating a collectively trained neural network.
As of 2020, Lc0 played over 300 million games against itself. It uses the same Monte Carlo Tree Search approach as AlphaZero, guided by a learned neural network rather than handcrafted evaluation.
Lc0 proved that AlphaZero's approach wasn't a fluke:
Yes -- and fans noticed immediately. Early Lc0 network T10 was remembered for its strong stylistic similarities to AlphaZero's first games. Both engines independently developed the "thorn pawn" strategy (advanced pawns cramping the opponent's position) and shared a fondness for:
The convergence makes sense: both learned chess purely through self-play with no human bias. The alien style wasn't an artifact of Google's hardware -- it was what chess looks like when you strip away centuries of human assumption.
AlphaZero wasn't just a chess story. It was a landmark in artificial intelligence research, published in the most prestigious scientific journal in the world.
| Game | Training Time | Opponent Defeated | Hardware |
|---|---|---|---|
| Chess | ~9 hours | Stockfish 8 (then world #1) | 5,000 TPUs gen1 + 64 TPUs gen2 |
| Shogi | ~12 hours | Elmo (2017 CSA champion) | Same infrastructure |
| Go | ~13 days | AlphaGo Zero | Same infrastructure |
Training used 5,000 first-generation TPUs generating self-play games and 64 second-generation TPUs for neural network training, all running in parallel. Match play used only 4 TPUs + 44 CPU cores. AlphaZero surpassed Stockfish in playing strength after just 4 hours of training.
The chess world's relationship with computers was fundamentally transformed.
| Aspect | Before AlphaZero | After AlphaZero |
|---|---|---|
| Engine aesthetics | Dry, materialistic, drawish. "Boring but correct." | Dynamic, sacrificial, creative. "Beautiful and correct." |
| Sacrificial play | Considered unsound at the highest level. Engines "refuted" gambits. | Proven objectively playable. Romantic chess vindicated. |
| Material evaluation | Material advantage = winning. Piece values were near-absolute. | Activity, harmony, and king restriction can outweigh material. |
| Chess's future | "Chess is being solved. All games will be draws." | "Chess is definitely not as drawish as we thought" (MVL) |
| Engine development | Alpha-beta search with hand-tuned evaluation. Brute force. | Neural networks + MCTS. Even Stockfish adopted NNUE. |
| Opening theory | Decades of human analysis, refined by engines. Considered settled. | Everything reopened. Novel ideas in every opening system. |
| Human-computer relationship | "Engines are tools that tell us the right move." | "Engines are teachers that show us new ways to think." |
| The draw problem | Super-tournaments had declining decisive games. | No-castling variant: 89% decisive. New ideas for chess's future. |
AlphaZero shows us that machines can be the experts, not merely expert tools. The knowledge it generates is information we can all learn from. AlphaZero is surpassing us in a profound and useful way, a model that may be duplicated on any other task or field where virtual knowledge can be generated.