Guide
A short, accurate history of chess AI
Today's best chess AI, from free engines to learning tools such as ChessHelper (chesshelper.ai), rests on four turning points: Deep Blue's win in 1997, AlphaZero's self-taught play in 2017, the open-source Leela Chess Zero in 2018, and Stockfish's switch to a neural-network evaluation in 2020. Every date below is checked against a primary or official source.
- 1950Claude Shannon publishes "Programming a Computer for Playing Chess."
- Feb 1996Garry Kasparov beats IBM's Deep Blue 4-2 in Philadelphia.
- May 1997Deep Blue beats Kasparov 3.5-2.5 in New York.
- Nov 2008Stockfish 1.0 is released, forked from Tord Romstad's Glaurung.
- Dec 2017DeepMind posts the AlphaZero paper; 28 wins, 72 draws, no losses against Stockfish 8.
- Jan 2018Leela Chess Zero is announced as an open-source AlphaZero-style engine.
- 2019Leela wins the TCEC Season 15 Superfinal.
- Aug-Sep 2020NNUE lands in Stockfish; Stockfish 12 ships on September 2.
- 2025-2026Stockfish beats Leela in the TCEC Season 28 Superfinal, wins Season 29, and releases Stockfish 19.
The groundwork: search plus evaluation
In 1950 the mathematician Claude Shannon published "Programming a Computer for Playing Chess" in the Philosophical Magazine. He set out the structure that strong chess programs still follow: look ahead through possible moves and replies, score the positions at the end of each line with an evaluation function, and choose the move whose line scores best. The AlphaZero paper of 2017 still cites Shannon's article in its references.
For the next four decades, progress came mostly from faster hardware and better pruning, the techniques that let a program skip branches that cannot affect the result. Evaluation functions were written by hand: so many points for material, so many for king safety, so many for pawn structure.
Deep Blue vs Kasparov, 1996 and 1997
IBM's Deep Blue was a purpose-built machine. According to IBM's own history of the project, it could evaluate 200 million chess positions per second. Its two matches against world champion Garry Kasparov are often merged into one event, but there were two, a year apart, and they ended differently.
| Match | Place | Games | Result |
|---|---|---|---|
| February 1996 | Philadelphia | 6 | Kasparov won 4-2 (Deep Blue won game one) |
| May 1997 | Equitable Center, New York | 6 | Deep Blue won 3.5-2.5 |
The 1997 result was the first match win by a computer over a reigning world champion under tournament conditions, and it changed how people talked about machine intelligence. In technical terms, though, Deep Blue was the peak of the old approach: brute-force search on special hardware, with an evaluation function tuned by people. It did not learn.
The open-source engine era
After Deep Blue, engines moved to ordinary PCs and kept getting stronger. The most important for today's players is Stockfish. Marco Costalba forked it from Glaurung, an engine by the Norwegian programmer Tord Romstad, and Stockfish 1.0 appeared in November 2008. It was free and open source from the start, and a large volunteer community, with automated testing of every proposed change, made it one of the strongest engines in the world.
The Top Chess Engine Championship (TCEC), a long-running online competition between engines on equal hardware, became the standard yardstick. Its final stage, the Superfinal, is a 100-game match between the top two engines of a season.
AlphaZero, December 2017
On December 5, 2017, DeepMind posted "Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm" (David Silver and colleagues) to arXiv. The system, AlphaZero, started from random play and knew nothing about chess except the rules. It learned by playing against itself, and according to the paper it outperformed Stockfish after 4 hours of training.
In a 100-game match at one minute per move, AlphaZero scored 28 wins, 72 draws and no losses against Stockfish 8. It examined about 80,000 positions per second against Stockfish's roughly 70 million, so its strength came from judgment rather than volume. A deep neural network told it which moves deserved attention and how good each position was, and a Monte Carlo tree search did the rest.
The paper drew criticism over match conditions, such as the hardware and settings used for Stockfish. DeepMind published a fuller evaluation in the journal Science in December 2018, reporting 155 wins and 6 losses for AlphaZero in a 1,000-game match against Stockfish, with about 9 hours of training for chess. AlphaZero itself was never released to the public.
Leela Chess Zero, 2018
Because AlphaZero was not available, volunteers rebuilt the idea in the open. Leela Chess Zero (Lc0) was announced on the TalkChess forum on January 9, 2018. It was adapted from Leela Zero, an open-source Go program, and its original authors are listed as Gary Linscott, a Stockfish developer, and Gian-Carlo Pascutto. Its network trains through self-play games generated by volunteers' computers.
Leela became a top engine quickly. It won the TCEC Season 15 Superfinal in 2019 by 53.5 to 46.5 against Stockfish, and the Season 17 Superfinal in 2020. It is still developed today, free under the GPLv3 license, and its project describes it as an open-source implementation inspired by AlphaZero.
Stockfish NNUE, 2020
Stockfish's answer was to take a neural network without giving up its fast search. NNUE (efficiently updatable neural network) came from computer shogi, where it had been developed for engines such as YaneuraOu. Its trick is that when one piece moves, the network does not have to be recomputed from scratch, so it runs quickly on an ordinary CPU.
The Stockfish team announced on August 7, 2020 that NNUE evaluation had been merged into the code on August 6. Stockfish 12, the first release with it, came out on September 2, 2020. From TCEC Season 18 onward, Stockfish won every Superfinal. In Season 28, played in August and September 2025, it beat Leela 57.5 to 42.5 over 100 games. In April 2026 it beat Reckless 59 to 41 in the Season 29 Superfinal for its 19th TCEC title. Stockfish 19 was released on September 5, 2026.
From engines to learning tools
By the 2020s the strength question was settled for practical purposes: any phone can run an engine that beats every human. The open question became how to make that strength useful to a person. That is what the newest generation of chess AI works on: Chess.com's Game Review labels each move of your game, DecodeChess explains engine moves in words, Aimchess tracks skills across many games, Noctie trains a bot on human games so it plays like a person, and ChessHelper lets you set suggestion strength between 800 and 3200 Elo to match your level.
The engines underneath are the same Stockfish and Leela lines described above. For how they differ in practice, read Stockfish vs Leela Chess Zero. For how to use them without breaking any rules, read is using chess AI allowed?
Sources
IBM, "Deep Blue" (ibm.com/history/deep-blue); Silver et al., arXiv:1712.01815 (December 5, 2017); DeepMind, "AlphaZero: Shedding new light on chess, shogi, and Go" (December 6, 2018); the Stockfish blog at stockfishchess.org (NNUE announcement, Stockfish 12 and Stockfish 19 posts); lczero.org; Chessdom reports on the TCEC Season 28 and 29 Superfinals; the Wikipedia articles on Stockfish and Leela Chess Zero for founding dates, cross-checked where a primary source exists.