From endgame tablebases to transformer architectures, from the draw problem to creative puzzle generation — a comprehensive investigation into where chess AI is heading and what remains unsolved.
Chess remains one of the most complex unsolved games in existence. Despite engines that play far beyond human capacity, the game itself resists complete solution by enormous margins.
Claude Shannon estimated in 1950 that solving chess would require a "dictionary" mapping optimal moves for approximately 4.8 × 1044 possible board positions. This is the number of legal positions reachable from the starting position — a number so vast it exceeds the number of atoms in the known universe (roughly 1080) by a factor that itself dwarfs comprehension.
The Shannon number — the conservative lower bound on chess's game-tree complexity — stands at 10120. Based on roughly 103 possibilities per pair of moves across a typical 40-move game, this figure represents more possible chess games than there are particles in the observable universe, by a margin of roughly 1040.
"A computer operating at 1 megahertz that could evaluate a terminal position in 1 microsecond would need 1090 years just to make its first move."
Strong solution: Finding the optimal strategy for every possible position. This requires exhaustively mapping the entire game tree — computationally impossible with any foreseeable technology.
Weak solution: Determining which outcome (White wins, Black wins, or draw) results from perfect play by both sides from the starting position, without necessarily revealing the full optimal strategy.
Even a weak solution remains far beyond current reach. For comparison, checkers was weakly solved in 2007 by Jonathan Schaeffer's team after 18 years of computation — and checkers has roughly the square root of chess's number of positions. Schaeffer himself stated that a breakthrough like quantum computing would be needed before solving chess could even be attempted.
Hans-Joachim Bremermann argued in 1965 that fundamental physical barriers — including the speed of light, quantum mechanics, and thermodynamics — place absolute limits on any computer's ability to examine chess's complete game tree.
Chess (generalized to an n×n board) is PSPACE-complete, meaning it belongs to the class of problems solvable with polynomial space but potentially requiring exponential time. This complexity class sits above BQP (bounded-error quantum polynomial time) — the class of problems efficiently solvable by quantum computers — which means even quantum computing won't provide a shortcut to solving chess completely.
No known algorithm, classical or quantum, can collapse the exponential search space of chess into something tractable. Brute force enumeration remains the only path to a complete solution, and the numbers make this permanently infeasible with physics as we understand it.
Tablebases represent chess's only domain of absolute, provably perfect play. Every position with 7 or fewer pieces has been solved. The frontier is pushing toward 8 pieces — but the road from 8 to 32 is effectively infinite.
The Syzygy tablebase format, developed by Ronald de Man, compressed all 7-piece endgames from 140 TB down to 18.4 TB, making them practical to deploy on consumer hardware. Completed in August 2018, these tables give engines perfect play in any position with 7 or fewer pieces on the board, including knowledge of the exact number of moves to checkmate.
Marc Bourzutschky has completed approximately 15% of pawnless 8-piece endgames. The longest forced win discovered requires exactly 400 moves (measured in depth-to-conversion, ending in a forced capture). Interestingly, the 7-piece record of 517 moves is actually longer, suggesting the growth pattern may plateau rather than follow the exponential doubling predicted by Haworth's Law.
Ronald de Man estimated in 2020 that complete 8-piece tablebases would be economically feasible within 5-10 years, requiring roughly 2 PB of disk space and a server with 64 TB of RAM. Lichess has already deployed partial 8-piece tables (op1 positions with only one opposing pair), covering slightly more than half of 8-piece endgames reached on the platform.
Each additional piece multiplies the search space by orders of magnitude. Conservative estimates suggest:
The tablebase approach works by exhaustive retrograde analysis — starting from every possible checkmate position and working backward. Each additional piece doesn't just add positions; it multiplies them combinatorially. There is no shortcut, no compression algorithm, and no architectural innovation that can bridge this gap. The full 32-piece tablebase will never exist.
The chess engine landscape has undergone a fundamental architectural shift — from hand-crafted evaluation functions to neural networks, and now to transformers and generative models that learn chess in entirely new ways.
Chessformer treats the 64 squares of a chessboard as tokens in a transformer encoder, paired with Geometric Attention Bias (GAB) — a novel positional encoding that adapts attention to the geometry of chess rather than using generic distance metrics. A bishop on d4 attends to diagonal squares, not just nearby ones. The result: Chessformer outperforms AlphaZero at 8x fewer FLOPs and matches grandmaster-level engines at 30x fewer FLOPs. The largest variant (CF-240M) has 243 million parameters, 15 encoder layers, and was trained on 500 million games.
The open-source heir to AlphaZero's approach, Lc0 switched to a transformer-based architecture in 2022, using a custom "smolgen" position encoding. After Stockfish became the first engine to cross 3700 Elo in April 2025, Lc0 followed close behind. It remains the strongest GPU-based engine and a proving ground for neural network chess research.
Released January 2026, Stockfish 18 introduces SFNNv10 — including "Threat Inputs" for recognizing threatened pieces — with a +46 Elo gain over v17. Its "Correction History" dynamically adjusts evaluations during search, enabling better fortress and stalemate detection. Trained on 100B+ positions with automated, reproducible pipelines. Estimated Elo: 3,653.
| Architecture | Approach | Strength | Weakness |
|---|---|---|---|
| Classical (pre-2020) | Hand-tuned eval + alpha-beta search | Fast, deterministic | No positional intuition |
| NNUE (Stockfish 12+) | Efficiently updatable neural net + search | Best of both worlds | CPU-bound architecture |
| Pure NN (Lc0, AlphaZero) | Self-play RL, MCTS | Deep positional understanding | GPU-dependent, slower nodes/sec |
| Transformer (Chessformer) | Board-as-tokens, attention-based eval | Compute efficient, interpretable | Still catching up in raw play |
| Diffusion/Generative | Learn distributions over positions | Creative, human-like output | Not competitive as engines |
Researchers have trained Auto-Regressive Transformers, Discrete Diffusion, and MaskGit models on 4 million chess puzzles from Lichess, using FEN notation as the sequence representation. These models don't play chess — they generate chess positions, learning the distribution of valid, interesting positions rather than optimal play.
When further trained with reinforcement learning, these generative models produce puzzles that are more creative, enjoyable, and counter-intuitive than human-composed book puzzles, according to expert evaluations. The diffusion approach also enables learning different trajectory spaces for specific Elo ranges, modeling human-like play at targeted skill levels.
Large language models are increasingly being evaluated through chess as a reasoning benchmark. Key findings from 2025-2026:
LLMs don't rival specialized engines, but chess serves as a powerful probe of general reasoning and multi-step planning capabilities.
DeepMind's chess research has evolved from a single superhuman agent to ensemble systems that incorporate diversity, creativity, and generative puzzle composition.
DeepMind's latest chess system, AZdb, combines multiple AlphaZero agents into a "league" structure. Two diversity techniques — Behavioral Diversity (maximizing differences in average piece positions) and Response Diversity (exposing agents to varied opponents) — produce agents with distinct playing styles: different opening preferences, varied pawn structures, and unique piece survivability approaches.
Through "sub-additive planning" (selecting the best agent for each opening type), AZdb achieves a +50 Elo increase over standard AlphaZero and solves twice as many challenging chess puzzles.
"Incorporating human-like creativity and diversity into AlphaZero can improve its ability to generalize."
In October 2025, DeepMind published a landmark 75-page study on generating creative chess puzzles using three generative architectures trained on 4 million Lichess puzzles. Expert reviewers — GM Matthew Sadler, GM Jonathan Levitt, and FM Amatzia Avni — evaluated the results:
DeepMind concluded: "Creativity and beauty in chess remain deeply subjective" — but AI can now generate positions that experts find genuinely novel and aesthetically pleasing.
AI hasn't just played chess better — it has discovered strategic concepts that humans missed in 1,500 years of play.
AlphaZero taught itself chess from scratch in four hours and developed a playing style that GM Matthew Sadler described as "discovering secret notebooks of some great player from the past." Its key contributions to theory include:
"It's like chess from another dimension."
Classical engines like pre-NNUE Stockfish excelled at tactics but struggled with long-term strategic assessment. Neural network engines revealed a strategic dimension previously invisible:
Magnus Carlsen explicitly cited AlphaZero as inspiration for his dominant 2019 performance, saying AlphaZero made him "a different player" by revealing new high-risk sacrificial strategies.
Elite chess preparation has been transformed from creative exploration into an engine-assisted science — with some grandmasters now pushing back against the very tools that made them stronger.
Opening theory now runs 20-30 moves deep in critical lines, with grandmasters arriving at the board having memorized engine-approved variations. Top players prepare using both Stockfish (for tactical precision) and neural network engines like Lc0 (for positional understanding), creating a hybrid approach that covers both dimensions.
The result: the early game has become "a feat of memory rather than creativity," transforming the first phase of elite games into what one writer called "trench warfare — a fight to avoid losing ground."
A striking counter-trend emerged in March 2026: grandmasters are now winning by making less optimal moves. After AI pushed chess toward perfect play and increasing draws, top players discovered that intentionally deviating from engine recommendations creates practical advantages against opponents expecting "book" moves.
This paradox — engines make preparation so thorough that the best strategy becomes not following engines — represents a fascinating evolution in the human-AI relationship.
"When I was a kid, I was preparing on paper-based material. Later on, in the early 2000s, it was already very clear that if you don't use the engine for help or advice, you're going to be falling behind."
"You cannot rely too much on it. If you're used to looking at computer lines, your brain will not switch on when it's time to play."
As engines surpass human capability by 1,000+ Elo points, the temptation and mechanics of cheating have become chess's most urgent existential challenge.
In September 2022, Magnus Carlsen accused Hans Niemann of cheating at the Sinquefield Cup, igniting chess's biggest scandal in decades. Chess.com's investigation found Niemann had cheated in 100+ online games, though no evidence of over-the-board cheating was confirmed. The lawsuit was privately settled, but the rivalry endures — Carlsen told Joe Rogan in 2025 that "there's still something off" about Niemann.
A Netflix documentary, Untold: Chess Mates, premieres April 7, 2026, bringing the scandal to a mainstream audience.
Chess.com's anti-cheating system operates on multiple layers:
As detection improves, so do cheating methods. Sophisticated cheaters don't copy engine moves directly — they use engines selectively in critical positions, play sub-optimal engine moves to avoid suspicion, or receive binary signals (better/worse) rather than specific moves. The gap between a strong grandmaster's natural play and engine-assisted play can be as small as a few centipawn per move — within the noise of natural variation.
Neural network approaches to detection are advancing, using trained models that analyze entire game trajectories rather than individual moves, but the fundamental challenge remains: a human playing 95% of their own moves and consulting an engine for the other 5% is extremely difficult to distinguish from a player simply having a good day.
Chess analysis has evolved from raw engine lines into AI systems that explain strategy in human language, personalize coaching, and generate real-time commentary.
Transforms Stockfish's numerical evaluations into natural language explanations — the "gold standard" for bridging the gap between engine precision and human comprehension. Explains not just the best move, but why it's best, using concepts like pawn structure, piece activity, and king safety.
Microsoft/U of T's human-like engine doesn't play the best move — it plays the move a human of a specific rating would play, with 52%+ accuracy. Personalized versions achieve 65% accuracy, and can identify individual players from 10 games with 86% reliability. Released as MaiaChess v1 in July 2025.
LLM-Powered Commentary: Frameworks combining lightweight CNN event prediction with open-source LLMs (LLaMA 3.3) generate dynamic, contextual commentary in real time. Chess.com's partnership with Perplexity signals the convergence of AI search and chess analysis.
Concept-Guided Generation: New research aligns textual analysis with engine signals through concept-guided scoring protocols, ensuring commentary is both accurate and accessible.
Rating-Aware Feedback: Rather than showing the objectively "best" move, next-gen tools show the best move for your level, identifying patterns in your mistakes and suggesting improvements calibrated to your skill.
Seventy years after Shannon's original paper, chess remains one of AI's most productive research environments — now serving as a benchmark for general-purpose reasoning, not just game-playing.
"Chess stands as a model system for studying how people can collaborate with AI, or learn from AI, just as chess has served as a leading indicator of many central questions in AI throughout the field's history."
Quantum computing is often cited as the technology that could finally "solve" chess. The reality is more nuanced — and more interesting — than the headlines suggest.
Chess (generalized) is PSPACE-complete, which sits above BQP — the class of problems quantum computers can solve efficiently. Grover's algorithm provides a quadratic speedup for search problems, reducing 10120 to roughly 1060 — still astronomically beyond feasibility. Quantum computing cannot collapse chess's exponential complexity into polynomial time.
While quantum won't solve chess outright, it could contribute in targeted ways:
Jonathan Schaeffer noted that quantum computing breakthroughs would be necessary before chess could be solved, but cautioned against underestimating future technology. The current consensus among experts is that chess will likely remain unsolved regardless of quantum advances.
TCEC and CCC pit the world's strongest engines against each other in continuous tournament play, revealing the cutting edge of engine development and the narrowing gap between architectural approaches.
Stockfish dev-20250824 defeated Leela Chess Zero to claim its 18th TCEC title. The superfinal saw the first 3700+ engine in competitive action, with the top-10 field averaging over 3621 Elo — more than 800 points above the best humans. TCEC Swiss 8 featured 44 engines, surpassing all previous participation records.
AlphaZero didn't just play strong chess — it played beautiful chess, reigniting a debate about whether creativity is exclusively human.
"AlphaZero's gameplay exhibits unconventional patterns that surprised chess analysts worldwide, employing counterintuitive tactics including queen sacrifices to secure positional advantages, revealing novel strategic dimensions previously unexplored in competitive chess."
"In positions requiring 'feeling,' 'insight,' or 'intuition,' AlphaZero plays like a human on fire."
Before AlphaZero, chess engines were associated with cold, mechanical play — tactically perfect but aesthetically dead. AlphaZero changed this completely. It won games through creative pawn sacrifices, speculative piece offers, and long-term strategic visions that human grandmasters found genuinely inspiring. Carlsen studied its games and adopted elements of its style.
The key insight: AlphaZero's neural network evaluation allowed it to make moves that looked "wrong" by traditional engine metrics but were deeply correct strategically. This is precisely what humans experience as beauty in chess — the surprising move that turns out to be right for reasons that aren't immediately visible.
University of Toronto researchers studied how humans perceive creativity in chess, finding that a move is perceived as "brilliant" when it breaks traditional rules in ultimately productive ways — sacrificing a piece that looks terrible but creates an unstoppable attack. This aligns precisely with what makes AlphaZero's play feel beautiful: it violates established heuristics to achieve deeper objectives.
DeepMind's puzzle composition research further demonstrates that AI can generate positions experts find aesthetically pleasing. The RL-trained generators produce puzzles that are rated as more creative, enjoyable, and counter-intuitive than composed book puzzles — suggesting AI has crossed a threshold from mere competence into genuine aesthetic production.
The philosophical question remains open: Is an AI generating beauty, or is it generating patterns that humans perceive as beautiful because they activate the same surprise-then-understanding response? Either way, the practical effect on chess culture is the same.
The draw rate is climbing, preparation is deeper than ever, and engines have surpassed humans by 1,000 Elo. Yet chess is more popular than it has ever been. The tension between these facts defines the game's future.
In the 2018 World Championship, all twelve classical games between Carlsen and Caruana ended in draws — a first in championship history dating to 1886. Draws have increased up to three times since 1850. Computer preparation has made the opening phase formulaic, with the first 15+ moves often pre-calculated. As one analyst wrote: "Computers have helped to flatten chess, increasing pure understanding at the expense of creativity, mystery, and dynamism."
But something unexpected happened. By 2026, grandmasters began winning by deliberately deviating from engine recommendations. The counter-revolution against AI homogenization has produced some of the most creative elite chess in decades. Meanwhile, chess viewership has exploded: Hikaru Nakamura and Magnus Carlsen lead streaming viewership rankings, Chess.com's user base continues to grow, and AI-powered analysis makes the game accessible to beginners who can now understand why grandmaster moves are brilliant.
"For the fans, it's been amazing. You can follow the games easier and get instant feedback when you play and learn faster."
"AI was extremely exciting at first because it presented a little bit of a different way to play chess, in more of a hybrid human-engine way."
"I've made my peace with it. 1997 was an unpleasant experience, but it helped me understand the future of human-machine collaboration."
Garry Kasparov developed what he calls "Kasparov's Law": a human of average intelligence and an AI system working together in harmony is more effective than either working alone, and even more advantageous than a brilliant human working with a system poorly.
This framework, born from his 1997 loss to Deep Blue, led to "Advanced Chess" (also called "Centaur Chess"), where human-engine teams compete. Kasparov advocates for "augmented intelligence" rather than artificial intelligence, arguing that AI tools should make us smarter rather than replace human judgment.
At a March 2025 conference, Kasparov urged over 2,000 attendees to "not be afraid of machines," asserting that "if we misuse them, we cannot blame technology." He emphasized that Deep Blue "didn't understand chess, or even know it was playing chess" — it performed brilliantly within narrow boundaries while possessing no understanding. This distinction, he argues, remains true of all current AI.
Bobby Fischer's answer to the memorization problem — randomize the starting position — has gained institutional legitimacy as the antidote to engine-driven opening preparation.
Chess960 (Fischer Random) randomizes the back-rank pieces among 960 possible starting configurations, making pre-computed opening theory irrelevant. In 2025, the Freestyle Chess Grand Slam Tour launched with five major tournaments, and Magnus Carlsen won the inaugural tour title. The FIDE Freestyle Chess World Championship followed in February 2026, with Carlsen defeating Fabiano Caruana in the final.
This format has found its moment precisely because of AI: as engines make standard opening preparation so thorough that it reduces creativity, Chess960 forces players back into pure calculation and intuition from move one.
Chess960 doesn't eliminate AI's advantage — engines are still vastly stronger than humans in any starting position. But it changes the human-vs-human dynamic fundamentally:
Chess960v2, a newer variant with double randomization, further increases unpredictability by randomizing both sides independently, creating thousands of additional starting positions.
From brute force to neural networks to transformers — a selected history of the key moments that shaped chess AI.