ChessWorld.net, founded in 2000, is an online chess site.
If you would like to play relaxed, friendly online chess, then...
or

📚 Chess Courses – Openings, Tactics, Middlegame, Endgames

AlphaZero, Neural Networks & The Revolution of AI Chess

On December 5, 2017, the history of chess was split into "Before AlphaZero" and "After AlphaZero." Google's DeepMind division released a paper detailing an AI that had started with zero chess knowledge—only the rules of the game—and, after just 4 hours of playing against itself, had defeated Stockfish 8, the strongest engine in the world at the time.

This wasn't just a victory for a new program; it was a victory for a new kind of intelligence. This page explores how Neural Networks work and why they play chess more like "aliens" than machines.

Reinforcement Learning: Learning from Scratch

Traditional engines (like the older versions of Stockfish or Deep Blue) were hand-crafted by Grandmasters and programmers. They were told: "Knights are worth 3 points," "Control the center," and "King safety is important."

AlphaZero knew none of this. It used a technique called Reinforcement Learning.

Tabula Rasa: It started with random moves. It didn't know a Queen was better than a Pawn.
Self-Play: It played millions of games against itself at blitz speed.
Feedback Loop: When it won, it strengthened the neural pathways that led to that result. When it lost, it weakened them.

The "Alien" Style: Why AlphaZero Was Different

Because AlphaZero wasn't taught human principles, it didn't have human biases. It played with a style that stunned the chess world.

1. The Return of the Gambit

Computers were historically materialistic—they would grab a pawn and defend it perfectly. AlphaZero, however, frequently sacrificed pawns (and sometimes exchanges) for long-term positional compensation that wouldn't pay off for 30 moves. It proved that activity > material.

2. Harry the h-pawn

AlphaZero had a distinct fondness for pushing flank pawns (a-pawn and h-pawn) early in the game to cramp the opponent's space. This idea has since been adopted by every top human player, including Magnus Carlsen.

3. King Marches

In complex middlegames, AlphaZero would often walk its King up the board to safety or to support an attack, a concept previously thought to be suicidal.

The Legacy: NNUE and Leela

AlphaZero was never released to the public. However, its impact is everywhere today:

Leela Chess Zero (Lc0): An open-source project that replicated the AlphaZero paper. Thousands of volunteers donated their GPU power to train Leela, creating a publicly available "neural" engine that anyone can use.
Stockfish NNUE: Stockfish eventually adapted. It combined its brute-force calculation speed with "Efficiently Updatable Neural Networks" (NNUE). This created a hybrid monster: an engine that calculates fast and understands chess like AlphaZero.