Machine Unique Chess Learn Test Research Sign in

Magnus Carlsen would lose about 950 of 1,000 games to Stockfish, a chess engine.

Can we teach the patterns that are unique to chess engines?

123,405positions from real games, analysed twice: by a strong engine and by a model of human play
4 in 100contain a move the engine rates decisively best that virtually no human plays, at any rating
8patterns those moves cluster into. Each card below is one of them, from a real game

Pick a pattern, study four positions, then drill 36 you haven't seen.

No signup needed to play. Register with an email when you want your progress everywhere.

Pattern 1 29.9%

Mostly middlegames; the answer is usually a quiet rook move.

924 positions 85% quiet −328cp average error
not started
Pattern 2 9.5%

Mostly middlegames; the answer is usually a quiet queen move.

444 positions 88% quiet −533cp average error
not started
Pattern 3 23.6%

Mostly middlegames; the answer is usually a quiet pawn move.

1149 positions 89% quiet −224cp average error
not started
Pattern 4 22.3%

Mostly openings; the answer is usually a quiet pawn move.

431 positions 89% quiet −219cp average error
not started
Pattern 5 26.0%

Mostly middlegames; the answer is usually a quiet rook move.

258 positions 77% quiet −440cp average error
not started
Pattern 6 24.1%

Mostly endgames; the answer is usually a quiet rook move.

829 positions 93% quiet −300cp average error
not started
Pattern 7 20.3%

Mostly middlegames; the answer is usually a quiet queen move.

861 positions 85% quiet −434cp average error
not started
Pattern 8 27.8%

Mostly middlegames; the answer is usually a quiet rook move.

259 positions 82% quiet −313cp average error
not started

The percentage is how often players who actually reached these positions over the board found the engine's move. Lower is harder.

Where this comes from

Every position was played in a real game. A position qualifies when a strong engine's move is decisively better than the alternatives and a neural model of human play gives that move under 5% probability at every rating from 1100 to 2600. Of 123,405 positions analysed, 5,155 qualified; in 3,956 the player at the board missed it too.

The eight groups above come from clustering the first 1,745 of those by how the engine represents them internally. The full method and findings →

12 players · 5 positions attempted · 0.0% solved.