Amazing results of LeelaKnightOdds-Dev

[Father]

Mr. Larry Kaufman.
This performance in the fight of man against machine with odds is really wonderful.
I consider that in man-machine duels with odds, that in some way, there are real probabilities on the part of the human being to acquire progress in shortening the distances.
If at any time you wish or consider that it is useful for you to participate in man-versus-machine duels, I would be attentive to an invitation, in exchange for the sole pleasure of playing chess, without a prize of any kind.

I would like to play against an odds in which the probabilities in a match at 20 games for me are 50%. What should this odds be like?

[lkaufman]

Mr. Larry Kaufman.
This performance in the fight of man against machine with odds is really wonderful.
I consider that in man-machine duels with odds, that in some way, there are real probabilities on the part of the human being to acquire progress in shortening the distances.
If at any time you wish or consider that it is useful for you to participate in man-versus-machine duels, I would be attentive to an invitation, in exchange for the sole pleasure of playing chess, without a prize of any kind.

I would like to play against an odds in which the probabilities in a match at 20 games for me are 50%. What should this odds be like?

Assuming you want to stick to blitz and bullet, the only option is queen odds. Probably you will score more than 50% at 3 min, but you will score much less than that at 1 min. Maybe 2 min or 1'1" or 2'1" might be around 50-50 for you at queen odds. Probably at queen for knight you could also score 50% at 15' + 10" if you really took the time, but I doubt that you want to do this. Color also makes a difference, normally odds giver gets White, but if you need a larger handicap you can play White, that probably adds 75 elo or so to the handicap.

[Father]

Thank you Mr. Larry Kaufman for allowing the participation of my points of view on this topic of odds.
These 4 algorithmic profiles are definitely superior to the same odds implemented inside the "powerful rigid algorithms" or previous to the AI ​​profiles (Alphazero and brothers) before I would say the year 2017.
I personally notice immediately the great or extraordinary work that goes into these odds.
Through the years of life that I have received as a gift, I have also had the opportunity to meet many players who are very in love with chess and dueling and study who dedicate important time of their lives to it, sheltered within the game of science. Many of these players have elo power markers from 1000-1200 and up. I consider it very important to open the umbrella of the algorithm so that everyone can take refuge from the barrage of moves that seek to checkmate them. The art of chess war poses the strategic and tactical impossibility of annihilation... makes compliance with the guiding principles of tactics and strategy imperative. With me the chess computers will continue to have the hunt... the duel will always stand with or without odds in the same way... it would be a beautiful gift from you to the community, to have an extensive menu for our breakfast...

[lkaufman]

Thank you Mr. Larry Kaufman for allowing the participation of my points of view on this topic of odds.
These 4 algorithmic profiles are definitely superior to the same odds implemented inside the "powerful rigid algorithms" or previous to the AI ​​profiles (Alphazero and brothers) before I would say the year 2017.
I personally notice immediately the great or extraordinary work that goes into these odds.
Through the years of life that I have received as a gift, I have also had the opportunity to meet many players who are very in love with chess and dueling and study who dedicate important time of their lives to it, sheltered within the game of science. Many of these players have elo power markers from 1000-1200 and up. I consider it very important to open the umbrella of the algorithm so that everyone can take refuge from the barrage of moves that seek to checkmate them. The art of chess war poses the strategic and tactical impossibility of annihilation... makes compliance with the guiding principles of tactics and strategy imperative. With me the chess computers will continue to have the hunt... the duel will always stand with or without odds in the same way... it would be a beautiful gift from you to the community, to have an extensive menu for our breakfast...

I expect that we will have other nets trained for specific handicaps in the future. Right now only the knight odds-dev bot is actually trained for the odds, the others rely on "Contempt" for good odds play. Perhaps eventually we'll have bots trained for rook odds, queen for knight, two knights, rook and knight, queen odds, and perhaps beyond (queen and knight, etc.), but we don't have the GPU resources for this right now. My belief is that training a net specifically for a handicap adds about a class (200 elo) to the level an opponent needs to score 50% against it, compared with just using Contempt.

[Uri Blass]

Since getting upgraded just prior to Sept. 28, LeelaKnightOdds-Dev (trained specifically for knight odds play) has the following results on LiChess, counting only games against 2400+ rated humans and only with a minimum of 3 min base AND 2 sec increment. 164 wins, five draws, ZERO losses!! Most of the opponents in the 2500s, which is typically somewhere between FM and IM level. This works out to a performance rating approaching 3300, way above the blitz rating of Magnus Carlsen on LiChess!! It's risky to extrapolate from this that the bot would actually win a 3'2" match giving Magnus Carlsen knight odds, but that would be my bet. Moreover, Leela did this despite being internally set to play 1'1" bullet speed or even faster, and despite the single-GPU computer sometimes playing other odds games simultaneously! Of course in Rapid it MIGHT only take a 2700 FIDE level opponent to win a match, perhaps we'll find out soon.
If anyone knows of any players near or above 2700 FIDE who might be interested in a match for moderate prize money, or if anyone has contact info for such a player who might be receptive, please send me a PM. We have one interested so far, but we would like to have two or three opponents this year. If and when we show that we can win a Rapid match from a 2700 ballpark player or two, we can start to approach the top ten players in the world. The goal is to be able to win a Rapid knight odds match with any human in the world by next year sometime. If we can't get Magnus Carlsen to play we'll have to show a performance rating against top-10 players that exceeds his rating to claim success.
I recall the discussions here about a decade ago, with the general conclusion that the above would be forever impossible, although I recall that Bob Hyatt disagreed with this conclusion. Hopefully we'll find out before too long.

The main question is how much elo can humans earn by a special preperation for these conditions.
Preperation can be both by memorizing opening lines and by learning from mistakes in previous games.

I think that we should talk about specific time control and not about blitz or rapid.
3+2 and 5+3 are both blitz but not the same blitz (there are also different rapid time controls).

[lkaufman]

Since getting upgraded just prior to Sept. 28, LeelaKnightOdds-Dev (trained specifically for knight odds play) has the following results on LiChess, counting only games against 2400+ rated humans and only with a minimum of 3 min base AND 2 sec increment. 164 wins, five draws, ZERO losses!! Most of the opponents in the 2500s, which is typically somewhere between FM and IM level. This works out to a performance rating approaching 3300, way above the blitz rating of Magnus Carlsen on LiChess!! It's risky to extrapolate from this that the bot would actually win a 3'2" match giving Magnus Carlsen knight odds, but that would be my bet. Moreover, Leela did this despite being internally set to play 1'1" bullet speed or even faster, and despite the single-GPU computer sometimes playing other odds games simultaneously! Of course in Rapid it MIGHT only take a 2700 FIDE level opponent to win a match, perhaps we'll find out soon.
If anyone knows of any players near or above 2700 FIDE who might be interested in a match for moderate prize money, or if anyone has contact info for such a player who might be receptive, please send me a PM. We have one interested so far, but we would like to have two or three opponents this year. If and when we show that we can win a Rapid match from a 2700 ballpark player or two, we can start to approach the top ten players in the world. The goal is to be able to win a Rapid knight odds match with any human in the world by next year sometime. If we can't get Magnus Carlsen to play we'll have to show a performance rating against top-10 players that exceeds his rating to claim success.
I recall the discussions here about a decade ago, with the general conclusion that the above would be forever impossible, although I recall that Bob Hyatt disagreed with this conclusion. Hopefully we'll find out before too long.

The main question is how much elo can humans earn by a special preperation for these conditions.
Preperation can be both by memorizing opening lines and by learning from mistakes in previous games.

I think that we should talk about specific time control and not about blitz or rapid.
3+2 and 5+3 are both blitz but not the same blitz (there are also different rapid time controls).

Yes, humans learn from mistakes, but the process is very slow with this bot, as it varies openings a lot and doesn't always play for the same strategy. One IM with a FIDE rating of 2420 played it about 140 games (3'2") before finally winning ONE game (got a few draws in the 140 games). If you want to limit the discussion to Rapid games (mostly 5'5" but some slower), since the last upgrade prior to Sep 28, against players rated over 2200 (up to 2700+), the score is 129 wins, six draws, ZERO losses. Over fifty games were by a single player rated over 2500. Sometimes unrated ("1500") players beat the bot with 98 or 99% accuracy, I don't need to tell you that they are using engines. It seems pretty clear that to win a Rapid match, let's say at 15'10", you would need an opponent above or at least near 2700 FIDE, and therefore to win a match at classical time controls (say 90' + 30") you need an opponent above or at least near 2500 FIDE (the requirement to earn the GM title normally). So if you know of either a 2500+ GM willing to play a classic TC match or a 2700+ GM willing to play a Rapid TC match for moderate prize money, let me know by PM and we can discuss details.

[Father]

Since getting upgraded just prior to Sept. 28, LeelaKnightOdds-Dev (trained specifically for knight odds play) has the following results on LiChess, counting only games against 2400+ rated humans and only with a minimum of 3 min base AND 2 sec increment. 164 wins, five draws, ZERO losses!! Most of the opponents in the 2500s, which is typically somewhere between FM and IM level. This works out to a performance rating approaching 3300, way above the blitz rating of Magnus Carlsen on LiChess!! It's risky to extrapolate from this that the bot would actually win a 3'2" match giving Magnus Carlsen knight odds, but that would be my bet. Moreover, Leela did this despite being internally set to play 1'1" bullet speed or even faster, and despite the single-GPU computer sometimes playing other odds games simultaneously! Of course in Rapid it MIGHT only take a 2700 FIDE level opponent to win a match, perhaps we'll find out soon.
If anyone knows of any players near or above 2700 FIDE who might be interested in a match for moderate prize money, or if anyone has contact info for such a player who might be receptive, please send me a PM. We have one interested so far, but we would like to have two or three opponents this year. If and when we show that we can win a Rapid match from a 2700 ballpark player or two, we can start to approach the top ten players in the world. The goal is to be able to win a Rapid knight odds match with any human in the world by next year sometime. If we can't get Magnus Carlsen to play we'll have to show a performance rating against top-10 players that exceeds his rating to claim success.
I recall the discussions here about a decade ago, with the general conclusion that the above would be forever impossible, although I recall that Bob Hyatt disagreed with this conclusion. Hopefully we'll find out before too long.

The main question is how much elo can humans earn by a special preperation for these conditions.
Preperation can be both by memorizing opening lines and by learning from mistakes in previous games.

I think that we should talk about specific time control and not about blitz or rapid.
3+2 and 5+3 are both blitz but not the same blitz (there are also different rapid time controls).

Good morning Mr. Uri Blass. My opinions are the result of my own perception, the product of the relationship between my being and the cybernetic being. The psychological profile of LeelaKnightOdds-Dev that I have been able to capture and interpret a little, is determined by a tendency and instruction of the algorithm in order to achieve greater influence and rapid spatial dominance markedly strategic, through the accumulation of a greater mobilization of troops to control the space where there is balance of material influence. This clever monster fights us with only one of his arms, while keeping the other in his pocket. I believe that the power of this beast can only be mitigated through the implementation of a strategic content plan that leads to the blind spot. You are fighting against a lightning occupation army, where the inventory of weapons is not so relevant compared to the speed of occupation, to the point that many raw materials and battle elements that will never participate in the war will remain unused. I see a minimal chance for the human being in this specific odds knight battle. On the other hand, the strategic intelligence of the human being could become definitively superior to that of the computer, but its implementation will be unlikely, due to the astronomical greater probability of incurring a tactical error on the part of the human being. The battle of man against machine with odds markers becomes spatially interesting to the extent that it makes reasonable the combination of conventional and anti-machine parameters by the human mind. The leaders could only be hurt through purely strategic ideas, this battle will not end. The essential weapon of the human against AI will always be his state of consciousness.

[Father]

Since getting upgraded just prior to Sept. 28, LeelaKnightOdds-Dev (trained specifically for knight odds play) has the following results on LiChess, counting only games against 2400+ rated humans and only with a minimum of 3 min base AND 2 sec increment. 164 wins, five draws, ZERO losses!! Most of the opponents in the 2500s, which is typically somewhere between FM and IM level. This works out to a performance rating approaching 3300, way above the blitz rating of Magnus Carlsen on LiChess!! It's risky to extrapolate from this that the bot would actually win a 3'2" match giving Magnus Carlsen knight odds, but that would be my bet. Moreover, Leela did this despite being internally set to play 1'1" bullet speed or even faster, and despite the single-GPU computer sometimes playing other odds games simultaneously! Of course in Rapid it MIGHT only take a 2700 FIDE level opponent to win a match, perhaps we'll find out soon.
If anyone knows of any players near or above 2700 FIDE who might be interested in a match for moderate prize money, or if anyone has contact info for such a player who might be receptive, please send me a PM. We have one interested so far, but we would like to have two or three opponents this year. If and when we show that we can win a Rapid match from a 2700 ballpark player or two, we can start to approach the top ten players in the world. The goal is to be able to win a Rapid knight odds match with any human in the world by next year sometime. If we can't get Magnus Carlsen to play we'll have to show a performance rating against top-10 players that exceeds his rating to claim success.
I recall the discussions here about a decade ago, with the general conclusion that the above would be forever impossible, although I recall that Bob Hyatt disagreed with this conclusion. Hopefully we'll find out before too long.

The main question is how much elo can humans earn by a special preperation for these conditions.
Preperation can be both by memorizing opening lines and by learning from mistakes in previous games.

I think that we should talk about specific time control and not about blitz or rapid.
3+2 and 5+3 are both blitz but not the same blitz (there are also different rapid time controls).

Yes, humans learn from mistakes, but the process is very slow with this bot, as it varies openings a lot and doesn't always play for the same strategy. One IM with a FIDE rating of 2420 played it about 140 games (3'2") before finally winning ONE game (got a few draws in the 140 games). If you want to limit the discussion to Rapid games (mostly 5'5" but some slower), since the last upgrade prior to Sep 28, against players rated over 2200 (up to 2700+), the score is 129 wins, six draws, ZERO losses. Over fifty games were by a single player rated over 2500. Sometimes unrated ("1500") players beat the bot with 98 or 99% accuracy, I don't need to tell you that they are using engines. It seems pretty clear that to win a Rapid match, let's say at 15'10", you would need an opponent above or at least near 2700 FIDE, and therefore to win a match at classical time controls (say 90' + 30") you need an opponent above or at least near 2500 FIDE (the requirement to earn the GM title normally). So if you know of either a 2500+ GM willing to play a classic TC match or a 2700+ GM willing to play a Rapid TC match for moderate prize money, let me know by PM and we can discuss details.

[[Event "Casual blitz game"]
[Site "https://lichess.org/JaouZPjQ"]
[Date "2024.10.13"]
[White "LeelaKnightOdds-dev"]
[Black "Catecan"]
[Result "1/2-1/2"]
[UTCDate "2024.10.13"]
[UTCTime "16:10:11"]
[WhiteElo "2000"]
[BlackElo "2082"]
[WhiteTitle "BOT"]
[Variant "From Position"]
[TimeControl "180+0"]
[ECO "?"]
[Opening "?"]
[Termination "Normal"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKB1R w KQkq - 0 1"]
[SetUp "1"]
[Annotator "lichess.org"]

1. d4 d5 2. Bf4 e6 3. Nc3 c6 4. e4 dxe4 5. Nxe4 Nf6 6. Nxf6+ Qxf6 7. Be5 Qd8 8. Qd2 Nd7 9. Bg3 Nf6 10. f3 Bd6 11. O-O-O Bxg3 12. hxg3 Qc7 13. Bd3 Bd7 14. Qg5 Rg8 15. Bxh7 Nxh7 16. Rxh7 Qd8 17. Qe5 Qb8 18. Qg5 Qd8 19. Qe5 Qb8 20. Qg5 Qd8 { The game is a draw. } 1/2-1/2[/pgn]

[Father]

[pgn][/pgn]

Since getting upgraded just prior to Sept. 28, LeelaKnightOdds-Dev (trained specifically for knight odds play) has the following results on LiChess, counting only games against 2400+ rated humans and only with a minimum of 3 min base AND 2 sec increment. 164 wins, five draws, ZERO losses!! Most of the opponents in the 2500s, which is typically somewhere between FM and IM level. This works out to a performance rating approaching 3300, way above the blitz rating of Magnus Carlsen on LiChess!! It's risky to extrapolate from this that the bot would actually win a 3'2" match giving Magnus Carlsen knight odds, but that would be my bet. Moreover, Leela did this despite being internally set to play 1'1" bullet speed or even faster, and despite the single-GPU computer sometimes playing other odds games simultaneously! Of course in Rapid it MIGHT only take a 2700 FIDE level opponent to win a match, perhaps we'll find out soon.
If anyone knows of any players near or above 2700 FIDE who might be interested in a match for moderate prize money, or if anyone has contact info for such a player who might be receptive, please send me a PM. We have one interested so far, but we would like to have two or three opponents this year. If and when we show that we can win a Rapid match from a 2700 ballpark player or two, we can start to approach the top ten players in the world. The goal is to be able to win a Rapid knight odds match with any human in the world by next year sometime. If we can't get Magnus Carlsen to play we'll have to show a performance rating against top-10 players that exceeds his rating to claim success.
I recall the discussions here about a decade ago, with the general conclusion that the above would be forever impossible, although I recall that Bob Hyatt disagreed with this conclusion. Hopefully we'll find out before too long.

The main question is how much elo can humans earn by a special preperation for these conditions.
Preperation can be both by memorizing opening lines and by learning from mistakes in previous games.

I think that we should talk about specific time control and not about blitz or rapid.
3+2 and 5+3 are both blitz but not the same blitz (there are also different rapid time controls).

Yes, humans learn from mistakes, but the process is very slow with this bot, as it varies openings a lot and doesn't always play for the same strategy. One IM with a FIDE rating of 2420 played it about 140 games (3'2") before finally winning ONE game (got a few draws in the 140 games). If you want to limit the discussion to Rapid games (mostly 5'5" but some slower), since the last upgrade prior to Sep 28, against players rated over 2200 (up to 2700+), the score is 129 wins, six draws, ZERO losses. Over fifty games were by a single player rated over 2500. Sometimes unrated ("1500") players beat the bot with 98 or 99% accuracy, I don't need to tell you that they are using engines. It seems pretty clear that to win a Rapid match, let's say at 15'10", you would need an opponent above or at least near 2700 FIDE, and therefore to win a match at classical time controls (say 90' + 30") you need an opponent above or at least near 2500 FIDE (the requirement to earn the GM title normally). So if you know of either a 2500+ GM willing to play a classic TC match or a 2700+ GM willing to play a Rapid TC match for moderate prize money, let me know by PM and we can discuss details.

[[Event "Casual blitz game"]
[Site "https://lichess.org/JaouZPjQ"]
[Date "2024.10.13"]
[White "LeelaKnightOdds-dev"]
[Black "Catecan"]
[Result "1/2-1/2"]
[UTCDate "2024.10.13"]
[UTCTime "16:10:11"]
[WhiteElo "2000"]
[BlackElo "2082"]
[WhiteTitle "BOT"]
[Variant "From Position"]
[TimeControl "180+0"]
[ECO "?"]
[Opening "?"]
[Termination "Normal"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKB1R w KQkq - 0 1"]
[SetUp "1"]
[Annotator "lichess.org"]

1. d4 d5 2. Bf4 e6 3. Nc3 c6 4. e4 dxe4 5. Nxe4 Nf6 6. Nxf6+ Qxf6 7. Be5 Qd8 8. Qd2 Nd7 9. Bg3 Nf6 10. f3 Bd6 11. O-O-O Bxg3 12. hxg3 Qc7 13. Bd3 Bd7 14. Qg5 Rg8 15. Bxh7 Nxh7 16. Rxh7 Qd8 17. Qe5 Qb8 18. Qg5 Qd8 19. Qe5 Qb8 20. Qg5 Qd8 { The game is a draw. } 1/2-1/2[/pgn]

[pgn] [Event "Casual blitz game"]
[Site "https://lichess.org/JaouZPjQ"]
[Date "2024.10.13"]
[White "LeelaKnightOdds-dev"]
[Black "Catecan"]
[Result "1/2-1/2"]
[UTCDate "2024.10.13"]
[UTCTime "16:10:11"]
[WhiteElo "2000"]
[BlackElo "2082"]
[WhiteTitle "BOT"]
[Variant "From Position"]
[TimeControl "180+0"]
[ECO "?"]
[Opening "?"]
[Termination "Normal"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKB1R w KQkq - 0 1"]
[SetUp "1"]
[Annotator "lichess.org"]

1. d4 d5 2. Bf4 e6 3. Nc3 c6 4. e4 dxe4 5. Nxe4 Nf6 6. Nxf6+ Qxf6 7. Be5 Qd8 8. Qd2 Nd7 9. Bg3 Nf6 10. f3 Bd6 11. O-O-O Bxg3 12. hxg3 Qc7 13. Bd3 Bd7 14. Qg5 Rg8 15. Bxh7 Nxh7 16. Rxh7 Qd8 17. Qe5 Qb8 18. Qg5 Qd8 19. Qe5 Qb8 20. Qg5 Qd8 { The game is a draw. } 1/2-1/2[/pgn]

[Father]

Since getting upgraded just prior to Sept. 28, LeelaKnightOdds-Dev (trained specifically for knight odds play) has the following results on LiChess, counting only games against 2400+ rated humans and only with a minimum of 3 min base AND 2 sec increment. 164 wins, five draws, ZERO losses!! Most of the opponents in the 2500s, which is typically somewhere between FM and IM level. This works out to a performance rating approaching 3300, way above the blitz rating of Magnus Carlsen on LiChess!! It's risky to extrapolate from this that the bot would actually win a 3'2" match giving Magnus Carlsen knight odds, but that would be my bet. Moreover, Leela did this despite being internally set to play 1'1" bullet speed or even faster, and despite the single-GPU computer sometimes playing other odds games simultaneously! Of course in Rapid it MIGHT only take a 2700 FIDE level opponent to win a match, perhaps we'll find out soon.
If anyone knows of any players near or above 2700 FIDE who might be interested in a match for moderate prize money, or if anyone has contact info for such a player who might be receptive, please send me a PM. We have one interested so far, but we would like to have two or three opponents this year. If and when we show that we can win a Rapid match from a 2700 ballpark player or two, we can start to approach the top ten players in the world. The goal is to be able to win a Rapid knight odds match with any human in the world by next year sometime. If we can't get Magnus Carlsen to play we'll have to show a performance rating against top-10 players that exceeds his rating to claim success.
I recall the discussions here about a decade ago, with the general conclusion that the above would be forever impossible, although I recall that Bob Hyatt disagreed with this conclusion. Hopefully we'll find out before too long.

Mr. Larry Kaufman. I thank you again for teaching me about odds. I want to tell you that I have begun to understand LeelaKnightOdds-Dev. As you can deduce from reading my last ten pieces of the dance, I have changed the strategic step, and the computer could soon be in trouble, to the point that if the probabilistic tendency, the blind spot, occurs, it could lead to a superior performance. to the estimated statistician to rate my playing strength. I will continue from my silence, building the network and the cage to contain, from the philosophy of a simple mortal and ordinary human being, the beast of AI. Thank you again Mr. Larry Kaufman. For me, the greatest reward received is that you have invested part of your time in discussing my concerns about the game of science and computers.

[pgn] [Event "Casual blitz game"]
[Site "https://lichess.org/L04V0eFa"]
[Date "2024.10.14"]
[White "LeelaKnightOdds-dev"]
[Black "Catecan"]
[Result "1/2-1/2"]
[UTCDate "2024.10.14"]
[UTCTime "03:16:26"]
[WhiteElo "2000"]
[BlackElo "2082"]
[WhiteTitle "BOT"]
[Variant "From Position"]
[TimeControl "180+0"]
[ECO "?"]
[Opening "?"]
[Termination "Normal"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKB1R w KQkq - 0 1"]
[SetUp "1"]
[Annotator "lichess.org"]

1. d4 d5 2. Bf4 e6 3. Nc3 c6 4. e4 dxe4 5. Nxe4 Nf6 6. Nxf6+ Qxf6 7. Be5 Qd8 8. Qd3 Nd7 9. Bg3 Nf6 10. O-O-O Bd6 11. Be2 Bxg3 12. hxg3 Qe7 13. Qc3 Bd7 14. Qa5 Qd8 15. Qe5 Qb8 16. Qg5 Rg8 17. Rd3 Qd8 18. Qe5 Qb8 19. Qe3 Nd5 20. Qd2 Nf6 21. Ra3 Qd8 22. g4 h6 23. f4 Qc7 24. g5 hxg5 25. fxg5 Nd5 26. Rf1 a6 27. Bc4 O-O-O 28. Qe2 Rgf8 29. Rb3 b5 30. Bxd5 exd5 31. Qd2 g6 32. a4 Kb7 33. Re3 Be6 34. Rc3 Qb6 35. b3 Rd7 36. Kb2 Rfd8 37. Rc5 Ka7 38. Ra1 Kb7 39. Qc3 Rc7 40. Qg3 Rcd7 41. Qc3 Rc7 42. Qg3 Rcd7 43. Qc3 { The game is a draw. } 1/2-1/2[/pgn]