I toast to LeelaqueenOdds and its creators. I raise my wine glass high and shout: Long live the queen. We have a queen!!

[M ANSARI]

I have to say I have been hooked on this BOT for the last week or so. I have had good success with slower time controls, but really at 1 minute it is ridiculously difficult as all it needs is one error and you are screwed. It certainly needs a completely different type of play than normal chess. You should readily accept to trade a bishop for a Knight in the opening (something you absolutely don't do in normal chess) even if it is not advantageous positionally. Also you can readily place your rook to have it exchanged for a bishop (usually it will not take it). Also it will always try to trap your queen and you should accept to give up your queen for a material disadvantage if it still gives you a simplified winning position. One thing is clear is that this BOT will NOT play the strongest most predictable move in the position ... it will play the move that will complicate the position the most ... especially as time control goes down. It has "learned" that as soon as the human hears that sound which tells your time is running out ... they will almost immediately fall for a one move "cheap shot" or will not be able to calculate quickly enough the danger of a sudden mating net. It is remarkable that this type of learning via Neural Networks can be so extremely successful against humans.

I feel you need to play a different type of chess to beat it, I have found that I have become tactically much much more aware chesswise. I appreciate much more that sometimes you need to just simplify the winning position and go for a weaker move that prevents any possible counter play rather than try to play what you think is the strongest move and try to calculate and punish what you know is a dubious move. I don't think humans can get anywhere near the strength tactically as this BOT, but in many situations where a strong human chess player finds himself in a losing position, he will look for tactically problematic tricks and will come up with similar type of play as the BOT does. I think a Queen is a huge material disadvantage to give up ... yet somehow the BOT will still beat the crap out of you with it. I cannot even imagine how a similar BOT with Knight or even Rook handicap would do. However the BOT would need to create a NN with that handicap. Remember that playing a normal full strength SF with a Queen handicap is actually quite easy to beat, but once a customized NN is created that understand human weaknesses ... that is just a totally different ball game!

An interesting thing is that when you do an analysis of the game you won, on many occasions it will show a 99% accuracy for both the BOT and you. That is 99% even though you have played very obvious low quality moves just to force exchanges or allowed your queen to get trapped just because it simplifies the position. Obviously the engine doing the game analysis is SF with a normal NN ... but it is still an interesting observation.

[lkaufman]

I have to say I have been hooked on this BOT for the last week or so. I have had good success with slower time controls, but really at 1 minute it is ridiculously difficult as all it needs is one error and you are screwed. It certainly needs a completely different type of play than normal chess. You should readily accept to trade a bishop for a Knight in the opening (something you absolutely don't do in normal chess) even if it is not advantageous positionally. Also you can readily place your rook to have it exchanged for a bishop (usually it will not take it). Also it will always try to trap your queen and you should accept to give up your queen for a material disadvantage if it still gives you a simplified winning position. One thing is clear is that this BOT will NOT play the strongest most predictable move in the position ... it will play the move that will complicate the position the most ... especially as time control goes down. It has "learned" that as soon as the human hears that sound which tells your time is running out ... they will almost immediately fall for a one move "cheap shot" or will not be able to calculate quickly enough the danger of a sudden mating net. It is remarkable that this type of learning via Neural Networks can be so extremely successful against humans.

I feel you need to play a different type of chess to beat it, I have found that I have become tactically much much more aware chesswise. I appreciate much more that sometimes you need to just simplify the winning position and go for a weaker move that prevents any possible counter play rather than try to play what you think is the strongest move and try to calculate and punish what you know is a dubious move. I don't think humans can get anywhere near the strength tactically as this BOT, but in many situations where a strong human chess player finds himself in a losing position, he will look for tactically problematic tricks and will come up with similar type of play as the BOT does. I think a Queen is a huge material disadvantage to give up ... yet somehow the BOT will still beat the crap out of you with it. I cannot even imagine how a similar BOT with Knight or even Rook handicap would do. However the BOT would need to create a NN with that handicap. Remember that playing a normal full strength SF with a Queen handicap is actually quite easy to beat, but once a customized NN is created that understand human weaknesses ... that is just a totally different ball game!

An interesting thing is that when you do an analysis of the game you won, on many occasions it will show a 99% accuracy for both the BOT and you. That is 99% even though you have played very obvious low quality moves just to force exchanges or allowed your queen to get trapped just because it simplifies the position. Obviously the engine doing the game analysis is SF with a normal NN ... but it is still an interesting observation.

Your comments suggest that you are unaware that there are indeed also LeelaRookOdds and LeelaKnightOdds bots, which have played countless matches with strong players that have been widely reported both here and elsewhere, with many videos of the games. LeelaRookOdds in blitz (3'2") is probably about the same strength as Hikaru or Magnus, though so far there haven't been any games with top GMs. It did beat GM Danny Gormally by 29 to 1 at 5'3"!! As for knight odds, it almost never loses a game (except to cheaters or due to internet outage or the like). One strong GM, David Paravyan, has won 5 times, but that's out of several hundred. Top fifty GMs including Caruana, Nepo, MVL, Giri, Dubov, Jorden Van Foreest, Awonder Liang, Ray Robson, and Sarana have played about 180 blitz games (mostly 3'2", some 5'3") at knight odds, none of them winning even a single game (about 15% draws)!! In Rapid it's not quite as overwhelming, but still crushing against GMs. There was one classical (one hour plus 30") match with GM Joel Benjamin, which he won by 2 to 1 with 5 draws, but that was before a major improvement on Feb. 27.

Regarding the 99% Accuracy, I believe all positions with an eval beyond 10 pawns are ignored, which means that a queen odds game that goes well for the human won't have many moves even considered. So Accuracy is only useful for knight odds and rook odds, not queen odds.

[Father]

I have to say I have been hooked on this BOT for the last week or so. I have had good success with slower time controls, but really at 1 minute it is ridiculously difficult as all it needs is one error and you are screwed. It certainly needs a completely different type of play than normal chess. You should readily accept to trade a bishop for a Knight in the opening (something you absolutely don't do in normal chess) even if it is not advantageous positionally. Also you can readily place your rook to have it exchanged for a bishop (usually it will not take it). Also it will always try to trap your queen and you should accept to give up your queen for a material disadvantage if it still gives you a simplified winning position. One thing is clear is that this BOT will NOT play the strongest most predictable move in the position ... it will play the move that will complicate the position the most ... especially as time control goes down. It has "learned" that as soon as the human hears that sound which tells your time is running out ... they will almost immediately fall for a one move "cheap shot" or will not be able to calculate quickly enough the danger of a sudden mating net. It is remarkable that this type of learning via Neural Networks can be so extremely successful against humans.

I feel you need to play a different type of chess to beat it, I have found that I have become tactically much much more aware chesswise. I appreciate much more that sometimes you need to just simplify the winning position and go for a weaker move that prevents any possible counter play rather than try to play what you think is the strongest move and try to calculate and punish what you know is a dubious move. I don't think humans can get anywhere near the strength tactically as this BOT, but in many situations where a strong human chess player finds himself in a losing position, he will look for tactically problematic tricks and will come up with similar type of play as the BOT does. I think a Queen is a huge material disadvantage to give up ... yet somehow the BOT will still beat the crap out of you with it. I cannot even imagine how a similar BOT with Knight or even Rook handicap would do. However the BOT would need to create a NN with that handicap. Remember that playing a normal full strength SF with a Queen handicap is actually quite easy to beat, but once a customized NN is created that understand human weaknesses ... that is just a totally different ball game!

An interesting thing is that when you do an analysis of the game you won, on many occasions it will show a 99% accuracy for both the BOT and you. That is 99% even though you have played very obvious low quality moves just to force exchanges or allowed your queen to get trapped just because it simplifies the position. Obviously the engine doing the game analysis is SF with a normal NN ... but it is still an interesting observation.

Your comments suggest that you are unaware that there are indeed also LeelaRookOdds and LeelaKnightOdds bots, which have played countless matches with strong players that have been widely reported both here and elsewhere, with many videos of the games. LeelaRookOdds in blitz (3'2") is probably about the same strength as Hikaru or Magnus, though so far there haven't been any games with top GMs. It did beat GM Danny Gormally by 29 to 1 at 5'3"!! As for knight odds, it almost never loses a game (except to cheaters or due to internet outage or the like). One strong GM, David Paravyan, has won 5 times, but that's out of several hundred. Top fifty GMs including Caruana, Nepo, MVL, Giri, Dubov, Jorden Van Foreest, Awonder Liang, Ray Robson, and Sarana have played about 180 blitz games (mostly 3'2", some 5'3") at knight odds, none of them winning even a single game (about 15% draws)!! In Rapid it's not quite as overwhelming, but still crushing against GMs. There was one classical (one hour plus 30") match with GM Joel Benjamin, which he won by 2 to 1 with 5 draws, but that was before a major improvement on Feb. 27.

Regarding the 99% Accuracy, I believe all positions with an eval beyond 10 pawns are ignored, which means that a queen odds game that goes well for the human won't have many moves even considered. So Accuracy is only useful for knight odds and rook odds, not queen odds.

Good evening, Mr. Larry Kaufman. I just played my first game against the bot "LeelaRookOdds," and on my first try, I hit the target. I played on a 1-minute time control with no increment. Here's my game. I hope it's helpful. Best regards.

[pgn] Event "Casual bullet game"]
[Site "https://lichess.org/aBLkGcPd"]
[Date "2025.04.24"]
[White "Catecan"]
[Black "LeelaRookOdds"]
[Result "1/2-1/2"]
[GameId "aBLkGcPd"]
[UTCDate "2025.04.24"]
[UTCTime "02:18:39"]
[WhiteElo "2030"]
[BlackElo "2000"]
[BlackTitle "BOT"]
[Variant "From Position"]
[TimeControl "60+0"]
[ECO "?"]
[Opening "?"]
[Termination "Normal"]
[FEN "1nbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQk - 0 1"]
[SetUp "1"]
[Annotator "lichess.org"]

1. d4 f5 2. f4 Nf6 3. e3 h6 4. Nf3 g5 5. c3 e6 6. Bd3 g4 7. Nh4 Kf7 8. g3 Be7 9. O-O d5 10. Bd2 b6 11. Be1 c5 12. Qe2 Bb7 13. Rf2 Nc6 14. Rg2 Rg8 15. Nd2 h5 16. Nf1 Qd7 17. a3 a5 18. Rc1 Rc8 19. Rd1 c4 20. Bc2 b5 21. Ra1 Ra8 22. Kh1 Na7 23. Bd1 Nc8 24. Kg1 Nd6 25. Kh1 Rc8 26. Kg1 Bc6 27. Kh1 Bb7 28. Kg1 Bc6 29. Kh1 Nde4 30. Kg1 Qb7 31. Kh1 Nd7 32. Kg1 Ndf6 33. Kh1 Nd7 34. Kg1 Nb6 35. Kh1 Nf6 36. Kg1 Kg7 37. Kh1 Kf7 38. Kg1 Kg7 39. Kh1 Nbd7 40. Kg1 Ra8 41. Kh1 Rb8 42. Kg1 Ra8 43. Kh1 Rb8 44. Kg1 Rc8 45. Kh1 Kf7 46. Kg1 Kg7 47. Kh1 Nb6 48. Kg1 Ne8 49. Kh1 Nf6 50. Kg1 Ne8 51. Kh1 Kf7 52. Kg1 Nf6 53. Kh1 Ne8 54. Kg1 Nf6 55. Kh1 Be8 56. Kg1 Ne4 57. Kh1 Nd6 58. Kg1 Bc6 59. Kh1 Be8 60. Kg1 Bc6 61. Kh1 Kg7 62. Kg1 Qa8 63. Kh1 Qb7 64. Kg1 Be8 65. Kh1 Bf7 66. Kg1 Bf6 67. Kh1 Be7 68. Kg1 a4 69. Kh1 Nd7 70. Kg1 Nf6 71. Kh1 Rg8 72. Kg1 Kf8 73. Kh1 Ke8 74. Kg1 Kd8 75. Kh1 Kc7 76. Kg1 Qa8 77. Kh1 Nb7 78. Kg1 Na5 79. Kh1 Nb3 80. Rb1 Na5 81. Kg1 Nb7 82. Kh1 Nd6 83. Kg1 Nfe4 84. Kh1 Nf6 85. Kg1 Nfe4 86. Kh1 Nf6 { The game is a draw. } 1/2-1/2[/pgn]

[Father]

I have to say I have been hooked on this BOT for the last week or so. I have had good success with slower time controls, but really at 1 minute it is ridiculously difficult as all it needs is one error and you are screwed. It certainly needs a completely different type of play than normal chess. You should readily accept to trade a bishop for a Knight in the opening (something you absolutely don't do in normal chess) even if it is not advantageous positionally. Also you can readily place your rook to have it exchanged for a bishop (usually it will not take it). Also it will always try to trap your queen and you should accept to give up your queen for a material disadvantage if it still gives you a simplified winning position. One thing is clear is that this BOT will NOT play the strongest most predictable move in the position ... it will play the move that will complicate the position the most ... especially as time control goes down. It has "learned" that as soon as the human hears that sound which tells your time is running out ... they will almost immediately fall for a one move "cheap shot" or will not be able to calculate quickly enough the danger of a sudden mating net. It is remarkable that this type of learning via Neural Networks can be so extremely successful against humans.

I feel you need to play a different type of chess to beat it, I have found that I have become tactically much much more aware chesswise. I appreciate much more that sometimes you need to just simplify the winning position and go for a weaker move that prevents any possible counter play rather than try to play what you think is the strongest move and try to calculate and punish what you know is a dubious move. I don't think humans can get anywhere near the strength tactically as this BOT, but in many situations where a strong human chess player finds himself in a losing position, he will look for tactically problematic tricks and will come up with similar type of play as the BOT does. I think a Queen is a huge material disadvantage to give up ... yet somehow the BOT will still beat the crap out of you with it. I cannot even imagine how a similar BOT with Knight or even Rook handicap would do. However the BOT would need to create a NN with that handicap. Remember that playing a normal full strength SF with a Queen handicap is actually quite easy to beat, but once a customized NN is created that understand human weaknesses ... that is just a totally different ball game!

An interesting thing is that when you do an analysis of the game you won, on many occasions it will show a 99% accuracy for both the BOT and you. That is 99% even though you have played very obvious low quality moves just to force exchanges or allowed your queen to get trapped just because it simplifies the position. Obviously the engine doing the game analysis is SF with a normal NN ... but it is still an interesting observation.

Your comments suggest that you are unaware that there are indeed also LeelaRookOdds and LeelaKnightOdds bots, which have played countless matches with strong players that have been widely reported both here and elsewhere, with many videos of the games. LeelaRookOdds in blitz (3'2") is probably about the same strength as Hikaru or Magnus, though so far there haven't been any games with top GMs. It did beat GM Danny Gormally by 29 to 1 at 5'3"!! As for knight odds, it almost never loses a game (except to cheaters or due to internet outage or the like). One strong GM, David Paravyan, has won 5 times, but that's out of several hundred. Top fifty GMs including Caruana, Nepo, MVL, Giri, Dubov, Jorden Van Foreest, Awonder Liang, Ray Robson, and Sarana have played about 180 blitz games (mostly 3'2", some 5'3") at knight odds, none of them winning even a single game (about 15% draws)!! In Rapid it's not quite as overwhelming, but still crushing against GMs. There was one classical (one hour plus 30") match with GM Joel Benjamin, which he won by 2 to 1 with 5 draws, but that was before a major improvement on Feb. 27.

Regarding the 99% Accuracy, I believe all positions with an eval beyond 10 pawns are ignored, which means that a queen odds game that goes well for the human won't have many moves even considered. So Accuracy is only useful for knight odds and rook odds, not queen odds.

Good evening, Mr. Larry Kaufman. I just played my first game against the bot "LeelaRookOdds," and on my first try, I hit the target. I played on a 1-minute time control with no increment. Here's my game. I hope it's helpful. Best regards.

[pgn] Event "Casual bullet game"]
[Site "https://lichess.org/aBLkGcPd"]
[Date "2025.04.24"]
[White "Catecan"]
[Black "LeelaRookOdds"]
[Result "1/2-1/2"]
[GameId "aBLkGcPd"]
[UTCDate "2025.04.24"]
[UTCTime "02:18:39"]
[WhiteElo "2030"]
[BlackElo "2000"]
[BlackTitle "BOT"]
[Variant "From Position"]
[TimeControl "60+0"]
[ECO "?"]
[Opening "?"]
[Termination "Normal"]
[FEN "1nbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQk - 0 1"]
[SetUp "1"]
[Annotator "lichess.org"]

1. d4 f5 2. f4 Nf6 3. e3 h6 4. Nf3 g5 5. c3 e6 6. Bd3 g4 7. Nh4 Kf7 8. g3 Be7 9. O-O d5 10. Bd2 b6 11. Be1 c5 12. Qe2 Bb7 13. Rf2 Nc6 14. Rg2 Rg8 15. Nd2 h5 16. Nf1 Qd7 17. a3 a5 18. Rc1 Rc8 19. Rd1 c4 20. Bc2 b5 21. Ra1 Ra8 22. Kh1 Na7 23. Bd1 Nc8 24. Kg1 Nd6 25. Kh1 Rc8 26. Kg1 Bc6 27. Kh1 Bb7 28. Kg1 Bc6 29. Kh1 Nde4 30. Kg1 Qb7 31. Kh1 Nd7 32. Kg1 Ndf6 33. Kh1 Nd7 34. Kg1 Nb6 35. Kh1 Nf6 36. Kg1 Kg7 37. Kh1 Kf7 38. Kg1 Kg7 39. Kh1 Nbd7 40. Kg1 Ra8 41. Kh1 Rb8 42. Kg1 Ra8 43. Kh1 Rb8 44. Kg1 Rc8 45. Kh1 Kf7 46. Kg1 Kg7 47. Kh1 Nb6 48. Kg1 Ne8 49. Kh1 Nf6 50. Kg1 Ne8 51. Kh1 Kf7 52. Kg1 Nf6 53. Kh1 Ne8 54. Kg1 Nf6 55. Kh1 Be8 56. Kg1 Ne4 57. Kh1 Nd6 58. Kg1 Bc6 59. Kh1 Be8 60. Kg1 Bc6 61. Kh1 Kg7 62. Kg1 Qa8 63. Kh1 Qb7 64. Kg1 Be8 65. Kh1 Bf7 66. Kg1 Bf6 67. Kh1 Be7 68. Kg1 a4 69. Kh1 Nd7 70. Kg1 Nf6 71. Kh1 Rg8 72. Kg1 Kf8 73. Kh1 Ke8 74. Kg1 Kd8 75. Kh1 Kc7 76. Kg1 Qa8 77. Kh1 Nb7 78. Kg1 Na5 79. Kh1 Nb3 80. Rb1 Na5 81. Kg1 Nb7 82. Kh1 Nd6 83. Kg1 Nfe4 84. Kh1 Nf6 85. Kg1 Nfe4 86. Kh1 Nf6 { The game is a draw. } 1/2-1/2[/pgn]

… In no way, shape or form do my results mean that the LeelaRookOdds computer lacks the playing power that mathematics concludes it has when evaluating the pulse of man against machine... all this reminds me of when at 19 years old, "a Master of life taught me": "Paul in response to your question: In life, 2 + 2 often do not equal 4."

[lkaufman]

Good evening, Mr. Larry Kaufman. I just played my first game against the bot "LeelaRookOdds," and on my first try, I hit the target. I played on a 1-minute time control with no increment. Here's my game. I hope it's helpful. Best regards.

[pgn] Event "Casual bullet game"]
[Site "https://lichess.org/aBLkGcPd"]
[Date "2025.04.24"]
[White "Catecan"]
[Black "LeelaRookOdds"]
[Result "1/2-1/2"]
[GameId "aBLkGcPd"]
[UTCDate "2025.04.24"]
[UTCTime "02:18:39"]
[WhiteElo "2030"]
[BlackElo "2000"]
[BlackTitle "BOT"]
[Variant "From Position"]
[TimeControl "60+0"]
[ECO "?"]
[Opening "?"]
[Termination "Normal"]
[FEN "1nbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQk - 0 1"]
[SetUp "1"]
[Annotator "lichess.org"]

1. d4 f5 2. f4 Nf6 3. e3 h6 4. Nf3 g5 5. c3 e6 6. Bd3 g4 7. Nh4 Kf7 8. g3 Be7 9. O-O d5 10. Bd2 b6 11. Be1 c5 12. Qe2 Bb7 13. Rf2 Nc6 14. Rg2 Rg8 15. Nd2 h5 16. Nf1 Qd7 17. a3 a5 18. Rc1 Rc8 19. Rd1 c4 20. Bc2 b5 21. Ra1 Ra8 22. Kh1 Na7 23. Bd1 Nc8 24. Kg1 Nd6 25. Kh1 Rc8 26. Kg1 Bc6 27. Kh1 Bb7 28. Kg1 Bc6 29. Kh1 Nde4 30. Kg1 Qb7 31. Kh1 Nd7 32. Kg1 Ndf6 33. Kh1 Nd7 34. Kg1 Nb6 35. Kh1 Nf6 36. Kg1 Kg7 37. Kh1 Kf7 38. Kg1 Kg7 39. Kh1 Nbd7 40. Kg1 Ra8 41. Kh1 Rb8 42. Kg1 Ra8 43. Kh1 Rb8 44. Kg1 Rc8 45. Kh1 Kf7 46. Kg1 Kg7 47. Kh1 Nb6 48. Kg1 Ne8 49. Kh1 Nf6 50. Kg1 Ne8 51. Kh1 Kf7 52. Kg1 Nf6 53. Kh1 Ne8 54. Kg1 Nf6 55. Kh1 Be8 56. Kg1 Ne4 57. Kh1 Nd6 58. Kg1 Bc6 59. Kh1 Be8 60. Kg1 Bc6 61. Kh1 Kg7 62. Kg1 Qa8 63. Kh1 Qb7 64. Kg1 Be8 65. Kh1 Bf7 66. Kg1 Bf6 67. Kh1 Be7 68. Kg1 a4 69. Kh1 Nd7 70. Kg1 Nf6 71. Kh1 Rg8 72. Kg1 Kf8 73. Kh1 Ke8 74. Kg1 Kd8 75. Kh1 Kc7 76. Kg1 Qa8 77. Kh1 Nb7 78. Kg1 Na5 79. Kh1 Nb3 80. Rb1 Na5 81. Kg1 Nb7 82. Kh1 Nd6 83. Kg1 Nfe4 84. Kh1 Nf6 85. Kg1 Nfe4 86. Kh1 Nf6 { The game is a draw. } 1/2-1/2[/pgn]

… In no way, shape or form do my results mean that the LeelaRookOdds computer lacks the playing power that mathematics concludes it has when evaluating the pulse of man against machine... all this reminds me of when at 19 years old, "a Master of life taught me": "Paul in response to your question: In life, 2 + 2 often do not equal 4."
[/quote]

Both LeelaKnightOdds and LeelaRookOdds were trained together against the same grandmaster-level bot opponent. As a consequence, Leela will be far more likely to go for a draw down a rook than down a knight. It is indeed MUCH easier to draw with Leela Rook odds than with Leela knight odds. But it is very difficult to beat either of them. We could potentially fix this by training for rook odds against a weaker opponent than the one used for knight odds, perhaps that may happen in the future. The queen odds bot is trained against a much weaker opponent, so it doesn't go for draws as readily as the rook odds bot.

[M ANSARI]

Oh I was very aware that Knight and Rook odds BOTS were out there! What I really didn't quite appreciate is that these BOTS had NN specifically designed to take advantage of human weaknesses when playing Knight or Rook odds. I had thought that this was just a super powerful engine that could play with this handicap. So in essence these BOTS were NOT playing the strongest chess move in the position, but rather the chess move that had the best chance of succeeding against a human. That is something I just did not appreciate the effectiveness of.

[lkaufman]

Oh I was very aware that Knight and Rook odds BOTS were out there! What I really didn't quite appreciate is that these BOTS had NN specifically designed to take advantage of human weaknesses when playing Knight or Rook odds. I had thought that this was just a super powerful engine that could play with this handicap. So in essence these BOTS were NOT playing the strongest chess move in the position, but rather the chess move that had the best chance of succeeding against a human. That is something I just did not appreciate the effectiveness of.

Well, there are no "strongest chess moves" in a lost position, unless your definition is the move that postpones the checkmate as long as possible with perfect play. The human definition is the move that gives the best practical chances of saving the game against an imperfect opponent, and this is precisely what the bots do. The imperfect opponent is a bot that is a simulation of a human GM (for the smaller odds) or master (for queen odds), but it is still a bot, not an actual human, so perhaps it's best to think of the Leela bots as optimized against both strong humans and weak engines, ones it can actually draw or beat at the odds.

[M ANSARI]

Yes, it is not easy to assess the "strongest chess move" in a losing position. But in general you would think that a chess engine would try to avoid as much as possible a forcing mate. Analyzing some of the games I played against the Queen odds BOT I noticed that quite often, it will play a move that allows a forced mate, even in the middle game. I am not sure that this happens due to low search as the BOT is being shared to play more people (thus not at full strength), or if it has "learned" that some forced mating sequences will not be found by a human, especially when time is low, and it is worth the risk to gamble. If it is "learned" behavior, then that would really be interesting. Engines will generally rate obscure forced mates the same as simple forced mates, and will avoid them at all costs.

[lkaufman]

Yes, it is not easy to assess the "strongest chess move" in a losing position. But in general you would think that a chess engine would try to avoid as much as possible a forcing mate. Analyzing some of the games I played against the Queen odds BOT I noticed that quite often, it will play a move that allows a forced mate, even in the middle game. I am not sure that this happens due to low search as the BOT is being shared to play more people (thus not at full strength), or if it has "learned" that some forced mating sequences will not be found by a human, especially when time is low, and it is worth the risk to gamble. If it is "learned" behavior, then that would really be interesting. Engines will generally rate obscure forced mates the same as simple forced mates, and will avoid them at all costs.

Yes, it is usually due to “search contempt”, which assumes that the human will only see lines that leela would find with just a fixed number of nodes, 40 for queen odds.

[chrisw]

Oh I was very aware that Knight and Rook odds BOTS were out there! What I really didn't quite appreciate is that these BOTS had NN specifically designed to take advantage of human weaknesses when playing Knight or Rook odds. I had thought that this was just a super powerful engine that could play with this handicap. So in essence these BOTS were NOT playing the strongest chess move in the position, but rather the chess move that had the best chance of succeeding against a human. That is something I just did not appreciate the effectiveness of.

Well, there are no "strongest chess moves" in a lost position,

engines with non-infinite search and an evaluation function would beg to differ

unless your definition is the move that postpones the checkmate as long as possible with perfect play. The human definition is the move that gives the best practical chances of saving the game against an imperfect opponent, and this is precisely what the bots do. The imperfect opponent is a bot that is a simulation of a human GM (for the smaller odds) or master (for queen odds), but it is still a bot, not an actual human, so perhaps it's best to think of the Leela bots as optimized against both strong humans and weak engines, ones it can actually draw or beat at the odds.