Chess Puzzles that Computers Can’t Solve! | Tactics Time!

Nick Risko proposes compositions that chess engines cannot solve. The positions may be impractical or impossible, but the “tree” of potential moves is just too much for engines to solve. However, human players can solve them! See the puzzle that Mikhail Tap famously solved in his head after a walk.

Puzzles by
Sir Roger Penrose
Frederic Friedel
Surya Ganguly


  1. For a long time, I've been opening up the last puzzle with Kb1, I found I actually like to open it with Kd2 better. It's a mate in 45 but Kd2 gives me possibility of mate in 43. When the knight crosses to black's side of the board, there's 2 tempos the white king may make. When the knight is on a4, if the black king moves to c6, the tempo Kd1 puts black in zugzwang to move to d6, then white can play Nb6. Black must always prevent the knight from getting across or it's a much sooner win. When white plays Nd5 most likely check, if the black king moves to e6, not only after b4xNa3can the pawn march to a Queen, but also the black king can travel over to defend the pawn on a3, the black king can break into the white side of the board with Kb3, black will sweep the white pawns and it's a win for black. If the white king is on the c file, it's close enough over, but on the d file, if the black king is on e6, a tempo to the other d square puts black in zugzwang and forces Kf7. The black king makes any other move and the knight gets across and gets the pawn on h6 and the knight and white king can do something on the king side of the board. The black king on f7 is far enough over that if b4xNa3, the black king cannot get back in time to defend the pawn from the white king. And do play the king on g2-h3 before the knight on g2-h4 because it saves a king move.

  2. 8/p7/kpP5/qrp1b3/rpP2b2/pP4b1/P3K3/8 w – – 0 1, computers can't lose this one but have wrong eval. Best eval with lc0 + BT4 after a few minutes > -3. I suppose she can reach > -2.5 with more time. 2.5 is generally considered as the winning zone limit. But I don't know if she can reach -0.5 that is considered as a draw for her.3B4/1r2p3/r2p1p2/bkp1P1p1/1p1P1PPp/p1P1K2P/PPB5/8 w – – 0 1, lc0 + BT4 find the solution at depth 8 (all the sequence of move in PV) but with wrong eval, SF 17 can not.8/3P3k/n2K3p/2p3n1/1b4N1/2p1p1P1/8/3B4 w – – 0 1, lc0 + BT4 + 6 men TB find immediately (on RTX 2060) the solution (right move, winning eval). SF 17 dev 4 threads + 8GB hash + 6 men TB has bad eval and bad first move after several minutes and reached depth 38.3k4/8/7p/2p1p1pP/1pPpPpP1/1P1P1P2/N7/2K5 w – – 0 1, SF 17 dev find the win at depth 75 after a few minutes, after more time lc0 has with a +1.10 eval all the knights moves in PV until Nh4 and then Nf5.

  3. Kind of late to the game, but "some" engines now can solve this type of position.3B4/1r2p3/r2p1p2/bkp1P1p1/1p1P1PPp/p1P1K2P/PPB5/8 w – – 0 1 A variant of Stockfish called BlackDiamond (v13.17) finds the solution at depth 24, with the correct moves and a score of 0 (draw).

  4. the last puzzle has exactly 3 solutions to mate in 45 moves and 1.Kb2 is none of them (while Kd2!! actually is)…the other 2 perfect moves are: 1.Kb1!! and 1.Kc2!!…1.Kd1? and Kb2? are blocking the knight on his way to a4 and are worse therefore (mate in 46 moves)…better one knows his stuff ;-)…Gijs van Breukelen's "Plaskett's Puzzle" is of course the highlight of this collection and a tiny bit too hard for this format imo…worth mentioning that even Kasparov couldnt solve it (4…Kg4 is indeed the best move but sad and hopeless anyway)

  5. Great video
    Great puzzles
    Great use of time

  6. Wtf, I literally just searched "Positions stockfish doesn't understand" a few minutes ago

  7. You can even make a dark square king move because three dark squared bishops will never checkmate a king without a light square bishop.

  8. @54:53 Kc6 is a blunder because white can sacrifice Nxg5. Ke6 is better move.

  9. At 14:03, c4? Loses to axb2 at the end. I think Ba4+ repeating is correct.

  10. At 10:52 that is not correct, you can only play c7 if you manage to hide your king on a8 first, otherwise c7?? Kb7 -+ nevertheless.

  11. @10:46 pushing the pawn to c7 does not win the game for white but loses it! After Kb7 the pawn is lost and the black's queen is free….

  12. @12:10 the position is shown, but what's the puzzle? Who 's to move? Would be helpful…

  13. The second puzzle was featured on chess talks channel

  14. A lot of chess software isn't programmed to value a draw over a loss. Because it keeps looking for a win or not to get beat, it will keep evaluating trying to find a win.

  15. i was just wondering if a machine couldn't solve a mate in one puzzle. weird

  16. Crystal NNUE solves all puzzles in this video in less than 30 seconds.

  17. It seems that the Breukelen puzzle (or "Plaskett's Puzzle" as the wikipedia article calls it) does indeed have a flaw on move 4 (with …Kg4 instead of …Nf7+), but it can be fixed by moving the black knight from g5 to either e5 or h8 in the starting position. The myth behind this puzzle is fascinating. Gijs van Breukelen claims to have composed it in the mid 1970's, which Dutch puzzle guru Harold van den Heiden confirms, but in "The Bobby Fischer I Knew" by Arnold Denker, he claims that he and Ossip Bernstein discussed this problem back in 1953 and it was falsely attributed to a game between Bernstein and Capablanca, which seemingly annoyed Bernstein, and he himself did not know where the puzzle was from.

    NB: Van Breukelen actually realised the position had a flaw, and worked out that the knight position had to be changed. It is possible that Tal was shown a corrected version at the 1987 tournament in Brussels.

  18. It should be possible to train a lc0 network on puzzles and it would solve them all eventually. For example, in the game of Go there's an old puzzle called Hatsuyoron 120, with carefully crafted positions and is considered to be the hardest one that even professionals and couldn't solve for decades. Once AlphaGo clones became available people started attempting to solve the puzzle using them, initially to no avail. In late 2019 the NN AI KataGo was trained specifically on it and it eventually came up with a much better solution than all of the previous ones. The solution was so different, in fact, that despite the original description of the puzzle being "Black to play and win", KataGo managed to find a very convincing way for white to win the game

  19. 13:56 Wait, 1. Ba4+ Kc4 2. Bb3+ Kb5 3. c4+ Kc6 4. d5+ Kd7 5. e6+ Kxd8 and now white has to let either axb2 or gxf4 happen, with either one providing an opening for black.

  20. I LOVE the final puzzle. I want to say, as white could get the knight on a4 in 9 moves if the king moves to a square that is not in the knight's path, and also with the plan to move the knight from a4 to b6 or capture c5, black can prepare for it by having the king on c6 to capture the knight before it moves to b6 or c5. However I do see that white has 2 movable pieces, black only has 1. White would not have to move the knight back, white can keep the knight there and use up it's turns by moving its king, black can't guard both squares because the king has to move, so the knight would break into black's side of the board, and get to that goal black pawn. So with white playing carefully, black cannot draw.

  21. Puzzle would have been cool except that those were NOT best moves by black in the main puzzle. Instead of forking the newly promoted queen, best move is for black King to move to g4. Disappointing puzzle!

  22. Why is the first problem hard for an engine? With 63 white pieces on the board there are no more than I think 424 (or obviously less than 8*63=508 anyway) legal moves to evaluate. Isn't the engine instructed that it is a mate in one problem? Doesn't engines routinely first evaluate all legal moves, one move deep, since there is no computation cost at all to it?

  23. you missed a second solution for White in the first puzzle. D7 x e8 = Q+, black N retakes and then the other pawn takes and queens with mate!

  24. In the final puzzle, if the knight goes to a3 or h4 hoping to get captured, the capture is not forced so it would result in a draw by 50 moves.

    On black's side, the knight would have to capture an unprotected pawn, then a protected pawn, if the knight is captured by a pawn, as long as the 2 white free pawns are far enough from each other, the king can't stop both from queening, and black resigns or hopes for stalemate.

  25. Final puzzle, I set this position up on my old electronic board, the Fidelity Marauder, and beat it by moving the knight across black's side to the goal on h6, however I found black could prevent it halfway across. Going back to white's side of the board via a3 with king defending. If black captures, the king can break through and sweep the pawns, from h4 to f5, the king cannot keep defending h6, the knight has g7 checked, the king has to move and the knight would capture h6. Across under left, back right, over, under back across left is the solution. I would prepare the white king to the right to save a move. If the king prevents the h pawn from passing through, the knight can capture another black pawn, and with 2 white pawns freed, black can't prevent both from queening.

  26. I actually found studying the last puzzle further in depth, you can get the knight to h4 on the 24th move. You really don't have to defend the knight when it's on a3. If the knight is even retrieving back to the queen side, the black king was on e6 or e8 before knight from d5 to c7, so if PxN on a3, the white king can capture the pawn in 2 or 3 moves and the black king won't have enough moves to defend it. For the 1st move, move the king to b1, if PxN, closer to capture the pawn is better, and the square b1 is clear of all 3 knight trips on white's half. When the knight is on a4, if the black king is on c6, use up the tempo before Kd6 by moving the white king back to c1, that will keep the path clear for the knight's last trip across the board and if PxN, the white king is still close enough to capture the black pawn on a, prevent it from queening or capture the queen, then get in Black's half of the board and win.

    Another thing, usually when the white knight gets to black's half of the board, the knight should go from b6 to d5, then b7, finally b5, a3. You can sometimes, only if the black king lets you, go from b6 to c8, then a7, and then b5, a3, but if the white king is on c1, black's king may be closer to defend it's pawn, so be more careful. Most of the time, Nd5 is better, the black king goes to e6 or e8 to prevent the white knight from traveling on the king side to capture the h pawn, then the white king can head back to capture the a pawn if PxN.

  27. Dear Friends, today a chess engine running on my mobile phone (Huawei Mate 9 from 2017) is able to solve such problems. For example SugaR_AI_ICCF 1.80 for Android is able to understand the problem of "fortress building", and therefore finds such moves as 1.La4+!!
    My point is, you should never say never…
    And, by the way, the problem with 1.La4+ is not from Frederick Friedel (Founder of the Chessbase Company). The author is W. E. Rudolph, the problem was presented in "La Strategie" in 1912.
    So you should, please, make a little more research, before you present such puzzles.
    And be very careful with statements like "can never be solved".
    My best regards, Torsten from Germany

  28. And regarding the Breukelen Puzzle, my mobile phone, Huawei Mate 9, which is by far not the fastest today, and SugaR_AI_ICCF 1 80 finds the correct way to win within some seconds!

  29. And you really should prepare yourself better for your presentations!
    The tournament where the Breukelen Study was presented to the grandmasters was not AVRO, nor 1978 or 1979, it was the SWIFT Tournament in Brussels in 1987.

  30. On Chess For Charity, someone else made a point what if white could play g6 en passant, white would win. Usually puzzles are with the assumption a pawn didn't previously move 2 squares. When setting up for en passant, the previous pawn move has to be moved unless it's chess software that has the en passant option like the castling option.

Leave a Reply

Your email address will not be published.