AlphaZero: The AI That Broke Chess's Rules
In December 2017, DeepMind published a paper that sent shockwaves through the chess world. Their neural network, AlphaZero, had been given nothing but the rules of chess. No opening books. No endgame tablebases. No database of master games. After four hours of self-play, playing millions of games against itself, it defeated Stockfish, the strongest traditional engine in the world, in a 100-game match. More remarkable than the result was how it played.
What AlphaZero did differently
Traditional chess engines like Stockfish evaluate positions using hand-crafted rules: material count, king safety scores, pawn structure penalties, piece mobility bonuses. AlphaZero had none of these. Instead, it used a deep neural network trained entirely through self-play, it learned what positions were good and bad by playing millions of games against itself and observing the results.
The result was a playing style that was recognisably chess, but not recognisably engine chess. Where Stockfish calculated billions of positions with brute force, AlphaZero evaluated far fewer positions but with a deeper "understanding" of which ones mattered. It played like a human Grandmaster, but one who had never read an opening book.
The style: sacrificial, dynamic, beautiful
AlphaZero's games were unlike any engine games before. It routinely sacrificed material, pawns, exchanges, even pieces, for long-term positional compensation. It preferred active pieces over material equality. It played openings that Stockfish's evaluation considered dubious (the King's Indian, gambits, early piece activity over material) and won with them.
Most strikingly, AlphaZero played beautifully. Its games had a human aesthetic quality, bold sacrifices followed by precise exploitation, that traditional engines, with their focus on material-counting evaluation, never produced. Garry Kasparov, writing in Science, described AlphaZero's style as reminiscent of the great romantic players of the 19th century, Morphy, Anderssen, but with modern precision.
What it means for human chess
AlphaZero's most lasting impact has been on human opening theory. Its preference for certain openings, the King's Indian Attack, the English Opening, aggressive Sicilian lines, validated approaches that had been considered slightly inferior by engine-driven analysis. Several ideas from AlphaZero's games have been adopted by top human players, including Carlsen.
More broadly, AlphaZero challenged the chess world's over-reliance on engine evaluations. For years, players had been using Stockfish's numerical assessment ("+0.3") as gospel. AlphaZero showed that a position Stockfish evaluates as "slightly worse" can be practically superior if it offers long-term dynamic compensation that is hard to evaluate numerically. Activity, piece coordination, and the initiative, the "human" factors, turned out to matter more than the engine's material count.
The post-AlphaZero landscape
Stockfish has since incorporated neural network evaluation (via NNUE), becoming dramatically stronger. The latest versions of Stockfish are substantially better than the AlphaZero that beat the 2017 Stockfish. The two approaches, traditional search with neural evaluation, have converged. Modern engines play with a combination of brute-force calculation and "intuitive" neural assessment that produces the strongest chess ever played, by human or machine.
For human players, the practical lesson is clear: use engines for analysis (ChessBhumi game review is powered by Stockfish), but don't worship the evaluation bar. A position is not "equal" just because the engine says +0.0. Play the position that gives you the most practical chances, the lesson AlphaZero taught the world.
Try it yourself
Review your games with engine analysis, but pay attention to the positional insights, not just the numerical evaluations. When the engine suggests a move, ask why, what principle does it illustrate? Active pieces? King safety? Pawn structure? That understanding is what transfers to your own play. The engine is a tool, not a teacher, you have to extract the lessons yourself.