The Evolution of Chess Engines: From Shannon to Stockfish 18

Chapter I

The Timeline

From a theoretical paper to superhuman intelligence in 76 years. Every milestone that mattered.

1950

Shannon's Paper

Claude Shannon publishes "Programming a Computer for Playing Chess," proposing two strategies: Type A (brute-force exhaustive search) and Type B (selective, human-like pruning). He estimates the game-tree complexity at ~10¹²⁰ positions, proving brute-force solution is impossible. This paper becomes the foundational blueprint for all chess programming.

Foundational Theory

1951

Turing's Algorithm

Alan Turing creates the first chess-playing algorithm, though no computer is powerful enough to run it. He hand-simulates it on paper. Dietrich Prinz programs the Ferranti Mark 1 at Manchester to solve mate-in-two problems via retrograde analysis.

First Algorithm

1957

Bernstein's Engine

Alex Bernstein creates the first fully functional chess engine that can play a complete game from start to finish. Each move takes approximately 8 minutes to compute. The program uses Shannon's Type B selective search.

First Complete Engine

1966-67

Mac Hack VI & Soviet Programs

MIT's Richard Greenblatt writes Mac Hack VI, the first program to play in a human tournament (1966 Massachusetts Amateur, rated 1243). In 1967, it becomes the first program to beat a human in tournament play. Meanwhile, the Soviet ITEP Program defeats the American Kotok-McCarthy program 3-1 in a historic East-West computer chess match played over 9 months.

Tournament Play Begins

1970-79

Chess 4.x & Kaissa

Northwestern University's Chess 4.x series (Atkin, Slate, Gorlen) dominates the US Computer Chess Championships, winning 1970-1973 and 1975. Chess 4.0 (1973) pioneered full-width Type A search, abandoning selective search. In 1974, the Soviet program Kaissa becomes the first World Computer Chess Champion. By 1977, Chess 4.6 defeats US Champion GM Walter Browne in a simultaneous exhibition.

Competitive Era Begins

1972-83

Belle: First Chess Master

Ken Thompson and Joe Condon at Bell Labs build Belle, pioneering custom chess hardware with dedicated boards for move generation, position evaluation, and microcode alpha-beta pruning. Belle wins 5 ACM championships and the 1980 World Computer Chess Championship. In 1983, Belle achieves a USCF Master rating of 2250 — the first machine to reach that level. Its architecture directly inspires the designs that become Deep Blue.

Hardware Revolution

1985-89

ChipTest & Deep Thought

At Carnegie Mellon, Feng-hsiung Hsu, Thomas Anantharaman, and Murray Campbell build ChipTest (1986) and evolve it into Deep Thought (1988). Deep Thought becomes the first computer to defeat a grandmaster (Bent Larsen, 1988) and wins the 1989 World Computer Chess Championship 5-0. However, it falls to Kasparov in a two-game exhibition match, proving the world champion is still beyond reach.

Grandmaster Barrier Broken

1996-97

Deep Blue vs. Kasparov

IBM sponsors the team and builds Deep Blue: an RS/6000 SP supercomputer with 30 PowerPC 604e processors and 480 custom VLSI chess chips, evaluating 200 million positions/second. In 1996, Kasparov wins the match 4-2, but Deep Blue takes Game 1 — the first time a reigning world champion loses to a computer under tournament conditions. In the 1997 rematch, an upgraded Deep Blue wins 3.5-2.5, changing history forever.

World Champion Falls

2004-08

Stockfish Origins

Tord Romstad creates Glaurung, an open-source chess engine (2004). In 2008, Marco Costalba forks it and creates Stockfish, named because the code was "produced in Norway and cooked in Italy." The first version is released November 2, 2008. Stockfish adopts the alpha-beta search with hand-crafted evaluation — the gold standard of conventional engines.

Open Source Revolution

2013

Fishtest Distributed Testing

Stockfish launches Fishtest, a distributed testing framework where volunteers donate CPU time. Using sequential probability ratio testing across thousands of games, the community can rigorously test every proposed improvement. In its first year, Fishtest helps Stockfish gain 120 ELO points, propelling it to the top of all major rating lists. By 2026, Fishtest will accumulate 19,900+ years of CPU time and 9.9 billion games played.

Community-Driven Development

Dec 2017

AlphaZero Shakes the World

DeepMind unveils AlphaZero, which taught itself chess from scratch in 4 hours using self-play reinforcement learning — no opening book, no endgame tables, no human knowledge. Using MCTS + deep neural networks on 5,000 TPUs, it defeats Stockfish 8 with +28 -0 =72. Its "alien" playing style — sacrifice-heavy, positionally profound, and deeply creative — stuns the chess world. Published in Science (Dec 2018).

Paradigm Shift

Jan 2018

Leela Chess Zero (Lc0)

Gary Linscott (a Stockfish developer) launches Leela Chess Zero, an open-source reimplementation of AlphaZero's approach. The project uses distributed volunteer computing: thousands of contributors generate self-play games to train a shared neural network. By 2026, Lc0 has played over 2.5 billion self-play games. In 2022, Lc0 transitions from convolutional networks to a transformer architecture, gaining ~300 ELO in raw policy strength.

Open-Source AlphaZero

2018

NNUE Invented for Shogi

Yu Nasu publishes the NNUE (Efficiently Updatable Neural Network) architecture for the shogi engine YaneuraOu (developed by Motohiro Isozaki). The key insight: a neural network where inputs change minimally between evaluations (moving one piece only changes 2-3 entries), enabling CPU-efficient incremental computation at millions of evaluations per second. The name is a Japanese wordplay on "Nue," a mythical chimera.

NNUE Born

Sep 2020

Stockfish 12: The NNUE Merger

After Hisayori "Nodchip" Noda ports NNUE to Stockfish in early 2020, Stockfish 12 officially incorporates it. The result: +100 ELO in one month — roughly two years of normal improvement compressed into a single architectural change. Stockfish now combines its battle-tested alpha-beta search with a neural network evaluation, creating a hybrid that outperforms both pure classical and pure neural approaches.

Hybrid Revolution

Nov 2020

Komodo Dragon: Another Hybrid

Komodo releases Dragon 1.0, integrating NNUE with both alpha-beta and MCTS search modes. Designed by Larry Kaufman and Mark Lefler (continuing Don Dailey's legacy after his 2013 death), Dragon uses a neural network trained supervised by GM Kaufman plus reinforcement learning from billions of positions. It offers a unique hybrid that can switch between traditional and Monte Carlo search.

MCTS + NNUE Hybrid

2023-25

Stockfish Dominance

Stockfish 16 (+50 ELO over SF15), 16.1, 17 (+46 ELO over SF16), and 17.1 continue the relentless climb. The engine wins every TCEC and CCC superfinal from Season 20 onward. Neural network training shifts to automated, reproducible pipelines using over 100 billion positions of Lc0 evaluation data. CCRL ratings push past 3,700.

Unstoppable

Jan 2026

Stockfish 18

Released January 31, 2026. Introduces SFNNv10 architecture with "Threat Inputs" for natural threat perception, "Shared Memory" for multi-process efficiency, "Correction History" for dynamic evaluation adjustment, and removal of the 1024-thread limit. +46 ELO over Stockfish 17, winning 4x more game pairs than it loses. The strongest chess entity ever created.

Current State of the Art

Chapter III

The AlphaZero Revolution

In December 2017, DeepMind changed everything. A program that knew nothing about chess — not even the value of a queen — taught itself to play at a superhuman level in four hours.

"It's like chess from another dimension."

Demis Hassabis, CEO of DeepMind

Training Time

Zero to superhuman

44M

Self-Play Games

During training

80K

Positions/sec

vs Stockfish's 70M

+28

Wins vs Stockfish

0 losses in 100 games

5,000

Gen-1 TPUs

For self-play generation

Gen-2 TPUs

For neural net training

AlphaZero's approach to chess was fundamentally different from any engine before it. Rather than searching millions of positions per second using brute force, it evaluated just 80,000 positions per second (compared to Stockfish's 70 million) but made each evaluation count through deep positional understanding.

Key characteristics of its play:

Material sacrifices for initiative: AlphaZero routinely sacrificed pawns, pieces, and even queens for long-term positional compensation — something traditional engines would never do because their evaluation functions assign fixed material values.
Rook pawn advances: It frequently pushed h-pawns as battering rams to open lines against the opponent's king, even when this appeared structurally dubious.
Long-range planning: While Stockfish relies on deep tactical calculation, AlphaZero's play showed strategic vision spanning 20-30+ moves — slowly restricting the opponent until the position collapsed.
King safety awareness: Its understanding of king safety went far beyond any hand-crafted evaluation function, leading to creative attacking play reminiscent of the great romantics.

Danish GM Peter Heine Nielsen compared it to "a superior alien species." Norwegian GM Jon Ludvig Hammer called it "insane attacking chess" with "profound positional understanding." Garry Kasparov said AlphaZero's style reminded him of his own bold, attacking play — and that watching it was "a pleasure."

GM Hikaru Nakamura offered a more measured take, noting that Stockfish was running on standard hardware while AlphaZero had Google's TPU infrastructure — an asymmetry that made direct strength comparisons questionable.

AlphaZero's architecture consists of three tightly integrated components:

1. Deep Neural Network: A convolutional neural network that takes a board position as input and outputs two things: (a) a policy vector — a probability distribution over all legal moves indicating which moves are most promising, and (b) a value estimate — a single number representing the expected outcome from that position.

2. Monte Carlo Tree Search (MCTS): Instead of alpha-beta's exhaustive search, AlphaZero uses MCTS to build a search tree selectively. The neural network's policy guides which branches to explore (focusing on promising moves), and its value estimate determines how to evaluate leaf nodes. This allows AlphaZero to "think deeply" about promising lines while ignoring obviously bad moves.

3. Self-Play Reinforcement Learning: The neural network is trained entirely through self-play. Starting from random play, AlphaZero plays millions of games against itself, using MCTS to select moves. The outcomes of these games (win/loss/draw) provide the training signal. Over 700,000 training steps with mini-batches of 4,096, the network learns to predict both good moves and accurate position evaluations.

The key insight is that no human chess knowledge is provided — only the rules of the game. AlphaZero rediscovered centuries of chess theory (and then went beyond it) in just four hours of training, producing approximately 3,000 games per second on Google's TPU hardware.

Leela Chess Zero (Lc0) launched January 9, 2018 — barely a month after AlphaZero's paper — as an open-source attempt to replicate AlphaZero's results. Founded by Gary Linscott (himself a Stockfish developer), the project uses distributed volunteer computing: anyone can contribute GPU cycles to generate self-play games that train the shared neural network.

Key milestones:

2019: Lc0 wins TCEC Season 15 and Season 17, defeating Stockfish in both superfinals — proving that the AlphaZero approach works at the highest level.
2022: Transition from convolutional neural networks to a transformer architecture (designed by Daniel Monroe and Philip Chalmers). This addresses the inherent weakness of CNNs: small receptive fields requiring 7+ hops for information to travel from corner to corner. Transformers see the entire board at once.
2024: The transformer model BT4 is ~300 ELO stronger in raw policy than the best convolutional model, with fewer parameters and less computation.
2025-26: Lc0 has played over 2.5 billion self-play games, adding ~1 million games per day. Its CCRL rating reaches 3,713.

Unlike AlphaZero (which required Google's TPU farm), Lc0 runs on consumer GPUs, making neural network chess accessible to anyone. The community-driven model — where distributed contributors replace Google's massive infrastructure — has been remarkably successful, demonstrating that open-source can compete with corporate AI research.

Chapter IV

The NNUE Revolution

How a neural network architecture from Japanese chess (shogi) transformed Stockfish overnight and created the dominant hybrid approach.

"NNUE gave Stockfish the equivalent of two years of improvement in a single month."

Stockfish Development Team, September 2020

NNUE (Efficiently Updatable Neural Network) was invented by Yu Nasu in 2018 for the shogi engine YaneuraOu. The name is a Japanese wordplay on "Nue" (鵼), a mythical chimera. It was first implemented by Motohiro Isozaki.

The core insight is that chess positions change minimally between evaluations. Moving a single piece typically changes only 2-3 entries in the network's input. NNUE exploits this by using incremental computation — instead of recalculating the entire neural network for each position, it only updates the parts affected by the last move.

Three design principles:

Low number of non-zero inputs (~0.1% input sparsity)
Minimal change between consecutive evaluations
Simple enough for low-precision integer arithmetic (int8/int16)

This lets NNUE run at millions of evaluations per second on CPU — no GPU required. It uses SIMD (Single Instruction Multiple Data) instructions available on modern CPUs to process 256-element vectors in parallel.

The architecture was inspired by Kunihito Hoki's "Bonanza method" for shogi, which used piece-square tables indexed by king location. NNUE generalizes this with a small neural network that fuses sparse input construction with the first layer's affine transformation.

The story of NNUE in Stockfish begins with Hisayori "Nodchip" Noda, a Japanese programmer who ported NNUE to a development branch of Stockfish 10 in early 2020. The results were immediately dramatic.

Timeline:

June 2020: NNUE evaluation introduced to Stockfish, initially as an experimental feature alongside the classical hand-crafted evaluation.
August 2020: Testing shows NNUE version gains 80+ ELO over classical Stockfish.
September 2, 2020: Stockfish 12 released with NNUE fully integrated, winning 10x more game pairs than it loses vs. Stockfish 11.
2021-2026: Continuous refinement of NNUE architecture through versions SFNNv1 through SFNNv10.

What makes Stockfish + NNUE special is the hybrid approach: the neural network handles position evaluation (replacing hand-crafted heuristics), while the battle-tested alpha-beta search handles move ordering and tactical calculation. This combines the pattern-recognition strengths of neural networks with the tactical precision of deep search — the best of both worlds.

By Stockfish 18, the neural network training pipeline uses over 100 billion positions of Lc0 evaluation data, creating a fascinating symbiosis where the MCTS-based engine's positional judgments are distilled into the alpha-beta engine's evaluation function.

Chapter V

Technical Comparison

Three paradigms, one game. How the major approaches to chess engine design differ in philosophy and implementation.

♙

Alpha-Beta + NNUE

Stockfish, Komodo Dragon

Search Alpha-Beta

Evaluation NNUE (neural)

Positions/sec 50-100M

Hardware CPU (multi-thread)

Search Depth 30-50+ plies

Style Tactical precision

Learning Supervised + RL

♘

MCTS + Neural Net

AlphaZero, Leela Chess Zero

Search MCTS

Evaluation Deep NN (value head)

Positions/sec 40-80K

Hardware GPU / TPU

Search Depth Variable (selective)

Style Positional / creative

Learning Pure self-play RL

♖

Classical (Historical)

Deep Blue, Pre-NNUE Stockfish

Search Alpha-Beta

Evaluation Hand-crafted heuristics

Positions/sec 10-200M

Hardware CPU / custom ASIC

Search Depth 6-20+ plies

Style Materialistic / tactical

Learning Human-engineered

Alpha-Beta Pruning asks: "What is the best move assuming my opponent plays perfectly?" It exhaustively explores the game tree to a fixed depth, pruning branches that can't possibly affect the outcome. This finds the objectively best move within the search horizon but is blind to possibilities beyond it.

Monte Carlo Tree Search asks: "Which move gives me the best chance of winning?" It selectively builds a search tree by simulating many possible continuations, guided by a neural network. Promising branches receive more attention; unpromising ones are ignored. This mirrors how humans think about chess — focusing on what's likely, not what's possible.

Key tradeoffs:

Alpha-beta is provably optimal within its search depth; MCTS is probabilistic and can miss forced tactical sequences.
MCTS handles long-term positional concepts better because it evaluates positions holistically rather than feature-by-feature.
Alpha-beta scales linearly with compute; MCTS benefits from larger neural networks and more GPU power.
In head-to-head testing (250 games), alpha-beta won 209, MCTS won 7, with 34 draws. But with more thinking time and larger networks, MCTS closes the gap significantly.

Modern Stockfish represents a convergence: alpha-beta search for tactical precision, NNUE (a simple neural network) for positional evaluation, and now Lc0's evaluation data feeding into NNUE training. The boundaries between paradigms are blurring.

Chapter VII

The Fat Fritz Controversy

Open source vs. commercial interests: a cautionary tale about what happens when companies try to monetize community work.

The controversy centers on ChessBase (a major chess software company) and developer Albert Silver.

Fat Fritz 1 (2019): Silver trained a custom neural network for Leela Chess Zero and bundled it as a commercial product through ChessBase. The open-source community objected to monetizing a project built on volunteer contributions, but the GPL license technically allowed it since Lc0's source code was included.

The DeusX Deception: In July 2018, Silver secretly submitted his Lc0 variant to TCEC under the name "DeusX," concealing its Leela Chess Zero origins. When discovered, this was seen as a breach of trust within the community.

Fat Fritz 2 (2021): The real storm. ChessBase released Fat Fritz 2, which was essentially Stockfish with a different neural network and minimal code changes. ChessBase marketed it as "the new number one" based on a selective 1,552-game match against Stockfish 12. The Stockfish and Lc0 development teams issued a joint statement condemning it as false advertising.

The GPL Lawsuit: The Stockfish developers filed a lawsuit against ChessBase for violating the GNU GPL v3 license. Fat Fritz 2 was sold commercially without making the source code publicly available as required by the license. ChessBase's position was that they had a different interpretation of the GPL. The case highlighted the tensions between open-source licensing and commercial exploitation.

The Outcome: Fat Fritz 2 was eventually released as free software under GPL 3.0. On independent rating lists, it failed to demonstrate any meaningful improvement over standard Stockfish, validating the community's claims that it offered "very little added value for money."

Chapter VIII

Hardware: The Compute Question

From room-sized supercomputers to smartphone apps. How much compute does a top chess engine actually need?

CPU Engines

Stockfish, Komodo Dragon

Optimal hardware: High-core-count x86 CPUs with AVX-512 or AVX-VNNI support. Stockfish 18 removes the 1024-thread limit.

Consumer level: A modern 16-core AMD Ryzen or Intel Core i9 runs Stockfish at full strength (~3,600+ ELO). Even a laptop with 4 cores plays at superhuman level.

Competition level: TCEC uses high-end server hardware (128+ threads). Fishtest volunteers collectively provide thousands of CPU cores.

NNUE efficiency: Uses int8/int16 SIMD instructions, requiring zero GPU. The neural network weights are small enough to share between processes via Stockfish 18's "Shared Memory" feature.

GPU Engines

Leela Chess Zero

Optimal hardware: High-end NVIDIA GPUs (RTX 4090, A100, H100) with CUDA/cuDNN support. Performance scales roughly linearly with GPU compute.

Consumer level: An RTX 3080 or better provides competitive play (~3,500+ ELO). Older or smaller GPUs still work but with significantly reduced strength.

Competition level: TCEC provides Lc0 with an NVIDIA A100 80GB. Google's AlphaZero used 4 TPUs for inference (5,000 TPUs for training).

Transformer networks: Lc0's BT4 transformer model benefits from larger GPU memory and higher throughput, scaling better than older convolutional architectures.

Deep Blue was a physical monument to brute-force computing:

Size: Two 2-meter-tall towers
Processors: 30 PowerPC 604e CPUs + 480 custom VLSI chess chips
Speed: 200 million positions/second (avg. 126M in actual play)
Search depth: 6-8 moves typically, up to 20+ in endgames
Performance: 11.38 GFLOPS
Cost: Millions of dollars in hardware and development
Estimated ELO: ~2,800

Today, Stockfish 18 running on a single modern smartphone would demolish Deep Blue. A 2026 flagship phone has more raw compute than Deep Blue's entire cabinet, and Stockfish's algorithms are orders of magnitude more efficient. The ~800 ELO gap between Deep Blue (~2,800) and Stockfish 18 (~3,650+) is equivalent to the gap between a strong club player and a world champion.

The lesson: algorithms matter more than hardware. Deep Blue won through custom silicon. Today's engines win through neural network evaluation and decades of search refinements, running on commodity hardware that costs a few hundred dollars.

Chapter IX

Is Chess Solved?

With engines reaching ELO 3,700+, are we approaching perfect play? The answer is more complex than you'd think.

10¹²⁰

Shannon Number

Possible game variations

10⁵⁰

Legal Positions

Estimated by Allis

10⁸⁰

Atoms in Universe

For comparison

Pieces Solved

Endgame tablebases

Chess is not solved. Not even close.

To "solve" chess means determining the theoretical outcome (win for White, win for Black, or draw) from the starting position, assuming perfect play by both sides. There are two levels:

Weakly solved: The outcome from the starting position is known, with a strategy that achieves it.
Strongly solved: The optimal move is known for every possible position.

What we have solved:

Endgame tablebases provide perfect play for all positions with 7 or fewer pieces (including both kings). The Syzygy tablebases (Ronald de Man and Bojun Guo, 2018) reduced the 7-piece dataset from 140 TB to 18.4 TB. The earlier Lomonosov tablebases were generated on a supercomputer at Moscow State University but have been offline since a 2021 ransomware attack.
The longest known forced win in 7-piece endgames requires 549 moves — far beyond any practical game.
Work on 8-piece tablebases has begun but the storage requirements are enormous.

Why full solution is likely impossible:

The Shannon number (10¹²⁰ game variations) exceeds the number of atoms in the observable universe (10⁸⁰) by 40 orders of magnitude.
Even at 1 trillion positions per second, solving chess would take billions of years.
Storing all legal positions would require more storage than all global data storage combined.

What engines tell us: The extremely high draw rate at the top level (Stockfish vs. Lc0 in TCEC often draws 50-70% of games) suggests that perfect play likely results in a draw. This is the majority opinion among experts, but it remains unproven.

The gap between "superhuman" and "solved" is astronomical. Engines at ELO 3,700 are vastly beyond any human, but they still make errors — just errors that are invisible to humans and only detectable by other engines at even deeper analysis depths.

Rank	Engine	Type	ELO (CCRL)
1	Stockfish 18	AB + NNUE	~3,759
2	Leela Chess Zero	MCTS + NN	~3,713
3	Obsidian	AB + NNUE	~3,686
4	Caissa	AB + NNUE	~3,641
5	Komodo Dragon	AB/MCTS + NNUE	~3,634
6	PlentyChess	AB + NNUE	~3,623
7	Berserk	AB + NNUE	~3,615

The Evolution of
Chess Engines

The Timeline

Shannon's Paper

Turing's Algorithm

Bernstein's Engine

Mac Hack VI & Soviet Programs

Chess 4.x & Kaissa

Belle: First Chess Master

ChipTest & Deep Thought

Deep Blue vs. Kasparov

Stockfish Origins

Fishtest Distributed Testing

AlphaZero Shakes the World

Leela Chess Zero (Lc0)

NNUE Invented for Shogi

Stockfish 12: The NNUE Merger

Komodo Dragon: Another Hybrid

Stockfish Dominance

Stockfish 18

ELO Progression

The AlphaZero Revolution

The NNUE Revolution

Technical Comparison

The Arena: TCEC & CCC

TCEC Superfinal Winners (Recent Seasons)

The Fat Fritz Controversy

Hardware: The Compute Question

CPU Engines

GPU Engines

Is Chess Solved?

The Current Landscape

What It All Means

Sources

The Evolution ofChess Engines

The Timeline

Shannon's Paper

Turing's Algorithm

Bernstein's Engine

Mac Hack VI & Soviet Programs

Chess 4.x & Kaissa

Belle: First Chess Master

ChipTest & Deep Thought

Deep Blue vs. Kasparov

Stockfish Origins

Fishtest Distributed Testing

AlphaZero Shakes the World

Leela Chess Zero (Lc0)

NNUE Invented for Shogi

Stockfish 12: The NNUE Merger

Komodo Dragon: Another Hybrid

Stockfish Dominance

Stockfish 18

ELO Progression

The AlphaZero Revolution

The NNUE Revolution

Technical Comparison

The Arena: TCEC & CCC

TCEC Superfinal Winners (Recent Seasons)

The Fat Fritz Controversy

Hardware: The Compute Question

CPU Engines

GPU Engines

Is Chess Solved?

The Current Landscape

What It All Means

Sources

The Evolution of
Chess Engines