Release : Chess System Tal EXTREME

[Peter Berger]

It is an interesting question what actually creates beauty and amazement for the spectators.

And it actually is an old question. When I think of the time of about the year 2000, when there were still commercial programs, though they had already become too strong for ordinary earthlings, the common label used to advertize them was "human-like".

So Rebel was supposed to be more "human-like" than Fritz, or Genius more "human-like" than Junior - or you name it. There never was too much to that other than advertisement IMHO. The people supposed to judge about the truth of these statements were too weak to judge for themselves anyway.

We can see this much clearer now. Personally I perceive very little beauty in the play of Stockfish 18 as I simply understand too few of its moves. So now it is 0.00 and a moment later it is 0.89 - there is no reason to doubt its assessment but I often have zero idea to the why myself.

It is way easier with the "style engines", as they make concessions that are most easy to understand for patzers, material. If they win this way, we feel that this is pretty. There is a downside to this: if they take it too far and get punished, they look silly in a way, that feels understandable, too.

It doesn't matter too much, what the actual truth about the observed chesspositions constitutes.

So maybe the current parameters that make EAS scores are more arbitrary than they feel to me anyway. It is just about creating +perceivable+ beauty for the playing level of the spectators.

I can show with a recent game - this is not beautiful, it looks just silly. Why? Because it didn't work from move 30 and led to a clear loss. Had white won, I would probably feel that this is "amazing".

But the real strength of the engine only matters here in that it has to get enough points to create the image/illusion of beauty. Ordinary spectators don't really "understand" the moves anyway.

IMHO - YMMV.
Peter

[pgn]

[Event "Lang 120min+10sek"]
[Site "Berlin"]
[Date "2026.02.08"]
[Round "?"]
[White "Absurd-1.0"]
[Black "Crafty 25.6"]
[Result "0-1"]
[ECO "C55"]
[PlyCount "126"]
[TimeControl "7200+10"]

{4096MB, DESKTOP-8OCGGEO} 1. e4 {[%eval 17,29] [%emt 0:03:35]} e5 {[%emt 0:00:07]} 2. Nf3 {[%eval 29,25] [%emt 0:00:54]} Nc6 {[%emt 0:00:07] (d5)} 3. Bc4 {[%eval 19,26] [%emt 0:09:41]} Nf6 {[%emt 0:00:08]} 4. d3 {[%eval 15,22] [%emt 0:01:07]} Be7 {[%emt 0:00:07] (Lc5)} 5. O-O {[%eval 20,23] [%emt 0:02:16]} O-O {[%emt 0:00:07] (d6)} 6. Nc3 {[%eval 25,24] [%emt 0:01:25]} d6 {[%emt 0:03:06]} 7. a4 {[%eval 17,26] [%emt 0:10:02]} Bg4 {[%emt 0:03:55]} 8. h3 {[%eval 24,23] [%emt 0:01:40]} Bxf3 {[%emt 0:01:06] (Lh5)} 9. Qxf3 {[%eval 18,23] [%emt 0:05:16]} Nd4 {[%emt 0:00:07]} 10. Qd1 {[%eval 20,23] [%emt 0:04:50]} c6 {[%emt 0:00:07]} 11. Ba2 {[%eval 18,24] [%emt 0:02:04]} Ne6 {[%emt 0:02:54]} 12. Ne2 {[%eval 27,26] [%emt 0:09:52]} d5 {[%emt 0:02:51]} 13. Ng3 {[%eval 15,27] [%emt 0:11:54]} Qc7 {[%emt 0:00:47] (dxe4)} 14. c3 {[%eval 27,23] [%emt 0:01:46]} Rad8 {[%emt 0:03:15] (a5)} 15. Qf3 {[%eval 39,22] [%emt 0:01:19]} Rfe8 {[%emt 0:07:35]} 16. Re1 {[%eval 39,25] [%emt 0:03:55]} Kh8 {[%emt 0:00:07] (a5)} 17. a5 {[%eval 43,22] [%emt 0:01:20]} Qxa5 {[%emt 0:03:33] (Lf8)} 18. exd5 {[%eval 24,24] [%emt 0:14:45]} cxd5 {[%emt 0:00:09]} 19. Rxe5 {[%eval 20,25] [%emt 0:01:11]} Bd6 {[%emt 0:00:59] (b5)} 20. Rf5 {[%eval 44,21] [%emt 0:01:00]} Qa6 {[%emt 0:03:52] (Db6)} 21. Rxf6 {[%eval 45,22] [%emt 0:01:25]} gxf6 {[%emt 0:03:10] (Lxg3)} 22. Bh6 {[%eval 60,23] [%emt 0:01:07]} Bf8 {[%emt 0:03:59]} 23. Nf5 {[%eval 48,27] [%emt 0:05:18]} d4 {[%emt 0:03:42]} 24. Bd2 {[%eval 30,26] [%emt 0:00:11]} Qb5 {[%emt 0:04:41]} 25. Bc4 {[%eval 31,27] [%emt 0:03:38]} Qxb2 {[%emt 0:02:29]} 26. Ra2 {[%eval 20,26] [%emt 0:01:22]} Qb6 {[%emt 0:04:14] (Db1+)} 27. Rc2 {[%eval 12,21] [%emt 0:02:19]} a6 {[%emt 0:02:52]} 28. Qh5 {[%eval 21,21] [%emt 0:02:23]} Qb1+ {[%emt 0:01:32]} 29. Rc1 {[%eval 25,23] [%emt 0:00:49]} Qb2 {[%emt 0:00:45]} 30. Nh6 {[%eval 0,24] [%emt 0:00:41]} Qxd2 {[%emt 0:01:59] (Te7)} 31. Nxf7+ {[%eval 29,22] [%emt 0:00:41]} Kg8 {[%emt 0:01:08] (Kg7)} 32. Rd1 {[%eval 65,22] [%emt 0:00:41]} Qxd1+ {[%emt 0:01:02]} 33. Qxd1 {[%eval 61,24] [%emt 0:00:46]} Kxf7 {[%emt 0:00:35]} 34. cxd4 {[%eval 31,23] [%emt 0:00:34]} Bg7 {[%emt 0:01:48]} 35. g3 {[%eval 48,25] [%emt 0:00:57]} b5 {[%emt 0:01:53]} 36. Ba2 {[%eval 34,29] [%emt 0:00:06]} Kf8 {[%emt 0:01:07] (Kg8)} 37. Qh5 {[%eval 21,27] [%emt 0:02:37]} h6 {[%emt 0:01:36] (Kg8)} 38. d5 {[%eval 19,21] [%emt 0:00:36]} Ng5 {[%emt 0:00:20]} 39. h4 {[%eval 6,23] [%emt 0:00:33]} Nf7 {[%emt 0:00:23]} 40. d6 {[%eval 7,26] [%emt 0:00:31]} Nxd6 {[%emt 0:00:25] (Td7)} 41. Qd5 {[%eval 13,27] [%emt 0:00:54]} Re1+ {[%emt 0:01:02] (Sc4)} 42. Kg2 {[%eval 23,23] [%emt 0:00:21]} Ke7 {[%emt 0:00:32]} 43. Qd4 {[%eval 10,25] [%emt 0:00:28]} Rd7 {[%emt 0:00:24] (Te5)} 44. Bb3 {[%eval 22,23] [%emt 0:01:00]} Rc7 {[%emt 0:01:00] (a5)} 45. Qg4 {[%eval 56,24] [%emt 0:00:29]} Bf8 {[%emt 0:01:11]} 46. Qb4 {[%eval 41,24] [%emt 0:02:16]} Rcc1 {[%emt 0:00:08] (Tec1)} 47. Qd4 {[%eval 59,21] [%emt 0:00:20]} f5 {[%emt 0:00:29] (Tc7)} 48. Qa7+ {[%eval 22,22] [%emt 0:00:20]} Kd8 {[%emt 0:01:22]} 49. h5 {[%eval 0,25] [%emt 0:00:21]} Re2 {[%emt 0:01:55] (Tc7)} 50. Qa8+ {[%eval 0,24] [%emt 0:00:16]} Ke7 {[%emt 0:00:44] (Tc8)} 51. Kf3 {[%eval 0,24] [%emt 0:00:32]} Re5 {[%emt 0:00:48] (Tcc2)} 52. Qa7+ {[%eval 0,23] [%emt 0:00:11]} Kd8 {[%emt 0:00:47]} 53. Qd4 {[%eval 0,30] [%emt 0:00:26]} Ree1 {[%emt 0:00:47] (Tcc5)} 54. Bd5 {[%eval 0,24] [%emt 0:00:13]} Be7 {[%emt 0:01:18] (Tc2)} 55. Kg2 {[%eval 0,26] [%emt 0:00:19]} Re2 {[%emt 0:01:07]} 56. Bb3 {[%eval 0,30] [%emt 0:00:26]} Ke8 {[%emt 0:00:14] (Kd7)} 57. Qh8+ {[%eval 0,24] [%emt 0:00:18]} Bf8 {[%emt 0:00:58]} 58. Qd4 {[%eval -24,23] [%emt 0:00:20]} Rb1 {[%emt 0:01:02] (a5)} 59. Bd5 {[%eval 0,27] [%emt 0:00:16]} Rbb2 {[%emt 0:00:29]} 60. Bc6+ {[%eval 0,26] [%emt 0:01:01]} Ke7 {[%emt 0:00:38] (Kf7)} 61. g4 {[%eval -79,21] [%emt 0:00:48]} Rxf2+ {[%emt 0:00:37]} 62. Kh3 {[%eval -572,21] [%emt 0:00:08]} f4 {[%emt 0:00:53] (Tbe2)} 63. Qa7+ {[%eval -482,18] [%emt 0:00:11]} Ke6 {[%emt 0:00:40]} 0-1

[/pgn]

[chrisw]

It is an interesting question what actually creates beauty and amazement for the spectators.

And it actually is an old question. When I think of the time of about the year 2000, when there were still commercial programs, though they had already become too strong for ordinary earthlings, the common label used to advertize them was "human-like".

So Rebel was supposed to be more "human-like" than Fritz, or Genius more "human-like" than Junior - or you name it. There never was too much to that other than advertisement IMHO. The people supposed to judge about the truth of these statements were too weak to judge for themselves anyway.

We can see this much clearer now. Personally I perceive very little beauty in the play of Stockfish 18 as I simply understand too few of its moves. So now it is 0.00 and a moment later it is 0.89 - there is no reason to doubt its assessment but I often have zero idea to the why myself.

It is way easier with the "style engines", as they make concessions that are most easy to understand for patzers, material. If they win this way, we feel that this is pretty. There is a downside to this: if they take it too far and get punished, they look silly in a way, that feels understandable, too.

It doesn't matter too much, what the actual truth about the observed chesspositions constitutes.

So maybe the current parameters that make EAS scores are more arbitrary than they feel to me anyway. It is just about creating +perceivable+ beauty for the playing level of the spectators.

I can show with a recent game - this is not beautiful, it looks just silly. Why? Because it didn't work from move 30 and led to a clear loss. Had white won, I would probably feel that this is "amazing".

But the real strength of the engine only matters here in that it has to get enough points to create the image/illusion of beauty. Ordinary spectators don't really "understand" the moves anyway.

IMHO - YMMV.
Peter

[pgn]

[Event "Lang 120min+10sek"]
[Site "Berlin"]
[Date "2026.02.08"]
[Round "?"]
[White "Absurd-1.0"]
[Black "Crafty 25.6"]
[Result "0-1"]
[ECO "C55"]
[PlyCount "126"]
[TimeControl "7200+10"]

{4096MB, DESKTOP-8OCGGEO} 1. e4 {[%eval 17,29] [%emt 0:03:35]} e5 {[%emt 0:00:07]} 2. Nf3 {[%eval 29,25] [%emt 0:00:54]} Nc6 {[%emt 0:00:07] (d5)} 3. Bc4 {[%eval 19,26] [%emt 0:09:41]} Nf6 {[%emt 0:00:08]} 4. d3 {[%eval 15,22] [%emt 0:01:07]} Be7 {[%emt 0:00:07] (Lc5)} 5. O-O {[%eval 20,23] [%emt 0:02:16]} O-O {[%emt 0:00:07] (d6)} 6. Nc3 {[%eval 25,24] [%emt 0:01:25]} d6 {[%emt 0:03:06]} 7. a4 {[%eval 17,26] [%emt 0:10:02]} Bg4 {[%emt 0:03:55]} 8. h3 {[%eval 24,23] [%emt 0:01:40]} Bxf3 {[%emt 0:01:06] (Lh5)} 9. Qxf3 {[%eval 18,23] [%emt 0:05:16]} Nd4 {[%emt 0:00:07]} 10. Qd1 {[%eval 20,23] [%emt 0:04:50]} c6 {[%emt 0:00:07]} 11. Ba2 {[%eval 18,24] [%emt 0:02:04]} Ne6 {[%emt 0:02:54]} 12. Ne2 {[%eval 27,26] [%emt 0:09:52]} d5 {[%emt 0:02:51]} 13. Ng3 {[%eval 15,27] [%emt 0:11:54]} Qc7 {[%emt 0:00:47] (dxe4)} 14. c3 {[%eval 27,23] [%emt 0:01:46]} Rad8 {[%emt 0:03:15] (a5)} 15. Qf3 {[%eval 39,22] [%emt 0:01:19]} Rfe8 {[%emt 0:07:35]} 16. Re1 {[%eval 39,25] [%emt 0:03:55]} Kh8 {[%emt 0:00:07] (a5)} 17. a5 {[%eval 43,22] [%emt 0:01:20]} Qxa5 {[%emt 0:03:33] (Lf8)} 18. exd5 {[%eval 24,24] [%emt 0:14:45]} cxd5 {[%emt 0:00:09]} 19. Rxe5 {[%eval 20,25] [%emt 0:01:11]} Bd6 {[%emt 0:00:59] (b5)} 20. Rf5 {[%eval 44,21] [%emt 0:01:00]} Qa6 {[%emt 0:03:52] (Db6)} 21. Rxf6 {[%eval 45,22] [%emt 0:01:25]} gxf6 {[%emt 0:03:10] (Lxg3)} 22. Bh6 {[%eval 60,23] [%emt 0:01:07]} Bf8 {[%emt 0:03:59]} 23. Nf5 {[%eval 48,27] [%emt 0:05:18]} d4 {[%emt 0:03:42]} 24. Bd2 {[%eval 30,26] [%emt 0:00:11]} Qb5 {[%emt 0:04:41]} 25. Bc4 {[%eval 31,27] [%emt 0:03:38]} Qxb2 {[%emt 0:02:29]} 26. Ra2 {[%eval 20,26] [%emt 0:01:22]} Qb6 {[%emt 0:04:14] (Db1+)} 27. Rc2 {[%eval 12,21] [%emt 0:02:19]} a6 {[%emt 0:02:52]} 28. Qh5 {[%eval 21,21] [%emt 0:02:23]} Qb1+ {[%emt 0:01:32]} 29. Rc1 {[%eval 25,23] [%emt 0:00:49]} Qb2 {[%emt 0:00:45]} 30. Nh6 {[%eval 0,24] [%emt 0:00:41]} Qxd2 {[%emt 0:01:59] (Te7)} 31. Nxf7+ {[%eval 29,22] [%emt 0:00:41]} Kg8 {[%emt 0:01:08] (Kg7)} 32. Rd1 {[%eval 65,22] [%emt 0:00:41]} Qxd1+ {[%emt 0:01:02]} 33. Qxd1 {[%eval 61,24] [%emt 0:00:46]} Kxf7 {[%emt 0:00:35]} 34. cxd4 {[%eval 31,23] [%emt 0:00:34]} Bg7 {[%emt 0:01:48]} 35. g3 {[%eval 48,25] [%emt 0:00:57]} b5 {[%emt 0:01:53]} 36. Ba2 {[%eval 34,29] [%emt 0:00:06]} Kf8 {[%emt 0:01:07] (Kg8)} 37. Qh5 {[%eval 21,27] [%emt 0:02:37]} h6 {[%emt 0:01:36] (Kg8)} 38. d5 {[%eval 19,21] [%emt 0:00:36]} Ng5 {[%emt 0:00:20]} 39. h4 {[%eval 6,23] [%emt 0:00:33]} Nf7 {[%emt 0:00:23]} 40. d6 {[%eval 7,26] [%emt 0:00:31]} Nxd6 {[%emt 0:00:25] (Td7)} 41. Qd5 {[%eval 13,27] [%emt 0:00:54]} Re1+ {[%emt 0:01:02] (Sc4)} 42. Kg2 {[%eval 23,23] [%emt 0:00:21]} Ke7 {[%emt 0:00:32]} 43. Qd4 {[%eval 10,25] [%emt 0:00:28]} Rd7 {[%emt 0:00:24] (Te5)} 44. Bb3 {[%eval 22,23] [%emt 0:01:00]} Rc7 {[%emt 0:01:00] (a5)} 45. Qg4 {[%eval 56,24] [%emt 0:00:29]} Bf8 {[%emt 0:01:11]} 46. Qb4 {[%eval 41,24] [%emt 0:02:16]} Rcc1 {[%emt 0:00:08] (Tec1)} 47. Qd4 {[%eval 59,21] [%emt 0:00:20]} f5 {[%emt 0:00:29] (Tc7)} 48. Qa7+ {[%eval 22,22] [%emt 0:00:20]} Kd8 {[%emt 0:01:22]} 49. h5 {[%eval 0,25] [%emt 0:00:21]} Re2 {[%emt 0:01:55] (Tc7)} 50. Qa8+ {[%eval 0,24] [%emt 0:00:16]} Ke7 {[%emt 0:00:44] (Tc8)} 51. Kf3 {[%eval 0,24] [%emt 0:00:32]} Re5 {[%emt 0:00:48] (Tcc2)} 52. Qa7+ {[%eval 0,23] [%emt 0:00:11]} Kd8 {[%emt 0:00:47]} 53. Qd4 {[%eval 0,30] [%emt 0:00:26]} Ree1 {[%emt 0:00:47] (Tcc5)} 54. Bd5 {[%eval 0,24] [%emt 0:00:13]} Be7 {[%emt 0:01:18] (Tc2)} 55. Kg2 {[%eval 0,26] [%emt 0:00:19]} Re2 {[%emt 0:01:07]} 56. Bb3 {[%eval 0,30] [%emt 0:00:26]} Ke8 {[%emt 0:00:14] (Kd7)} 57. Qh8+ {[%eval 0,24] [%emt 0:00:18]} Bf8 {[%emt 0:00:58]} 58. Qd4 {[%eval -24,23] [%emt 0:00:20]} Rb1 {[%emt 0:01:02] (a5)} 59. Bd5 {[%eval 0,27] [%emt 0:00:16]} Rbb2 {[%emt 0:00:29]} 60. Bc6+ {[%eval 0,26] [%emt 0:01:01]} Ke7 {[%emt 0:00:38] (Kf7)} 61. g4 {[%eval -79,21] [%emt 0:00:48]} Rxf2+ {[%emt 0:00:37]} 62. Kh3 {[%eval -572,21] [%emt 0:00:08]} f4 {[%emt 0:00:53] (Tbe2)} 63. Qa7+ {[%eval -482,18] [%emt 0:00:11]} Ke6 {[%emt 0:00:40]} 0-1

[/pgn]

If all sacrificial attacks worked they wouldn’t be sacrificial attacks. If they always worked Cstal would be 200 Elo above SF in the rating lists, and it isn’t, so they don’t.
What exactly do you want?

[smatovic]

[...]

Maybe it should be considered more for human opponent play rather than engine vs. engine.

--
Srdja

[chrisw]

[...]

Maybe it should be considered more for human opponent play rather than engine vs. engine.

--
Srdja

exactly, you're playing something strong and wild that makes mistakes

[Peter Berger]

If all sacrificial attacks worked they wouldn’t be sacrificial attacks. If they always worked Cstal would be 200 Elo above SF in the rating lists, and it isn’t, so they don’t.
What exactly do you want?

I was not really asking for anything here, just sharing my thoughts and observations. I think that CSTal/Rebel is a most interesting experiment to improve on the perceived beauty of the games and I follow its development with interest.

To also answer to the two following posts - if I understood correctly that there is some kind of "parameter" "extremism" that you can vary in training: I wonder how CSTal games with this leading to an ELO 2500 and ELO 2200 level would look like. A 3000 ELO level if real is still just too strong to make it viable as an opponent for humans.

[Rebel]

If all sacrificial attacks worked they wouldn’t be sacrificial attacks. If they always worked Cstal would be 200 Elo above SF in the rating lists, and it isn’t, so they don’t.
What exactly do you want?

I was not really asking for anything here, just sharing my thoughts and observations. I think that CSTal/Rebel is a most interesting experiment to improve on the perceived beauty of the games and I follow its development with interest.

To also answer to the two following posts - if I understood correctly that there is some kind of "parameter" "extremism" that you can vary in training: I wonder how CSTal games with this leading to an ELO 2500 and ELO 2200 level would look like. A 3000 ELO level if real is still just too strong to make it viable as an opponent for humans.

Absurd 1.0 (or better Absurd-3000 as it turned out) is (was) just an example how easy it now is to create a higher ranked style engine once you have created a solid network controlled by a parameter. It takes one day to create the network and one day to test it, how convenient. To make an elo rated Absurd-2500 now is a piece of cake, however I am not going to put my brand name to it, neither can I stand the garbage moves it surely will play. Absurd-3000 is already over the top.

[chrisw]

If all sacrificial attacks worked they wouldn’t be sacrificial attacks. If they always worked Cstal would be 200 Elo above SF in the rating lists, and it isn’t, so they don’t.
What exactly do you want?

I was not really asking for anything here, just sharing my thoughts and observations. I think that CSTal/Rebel is a most interesting experiment to improve on the perceived beauty of the games and I follow its development with interest.

To also answer to the two following posts - if I understood correctly that there is some kind of "parameter" "extremism" that you can vary in training: I wonder how CSTal games with this leading to an ELO 2500 and ELO 2200 level would look like. A 3000 ELO level if real is still just too strong to make it viable as an opponent for humans.

What’s all this about “perceived beauty”, that was never mentioned as a goal. The goal is to play way more aggressively especially in the king attack sphere and thus likely play sacrificial unbalanced lines. The deep dark forest as said Tal.

[Rebel]

If all sacrificial attacks worked they wouldn’t be sacrificial attacks. If they always worked Cstal would be 200 Elo above SF in the rating lists, and it isn’t, so they don’t.
What exactly do you want?

I was not really asking for anything here, just sharing my thoughts and observations. I think that CSTal/Rebel is a most interesting experiment to improve on the perceived beauty of the games and I follow its development with interest.

To also answer to the two following posts - if I understood correctly that there is some kind of "parameter" "extremism" that you can vary in training: I wonder how CSTal games with this leading to an ELO 2500 and ELO 2200 level would look like. A 3000 ELO level if real is still just too strong to make it viable as an opponent for humans.

What’s all this about “perceived beauty”, that was never mentioned as a goal. The goal is to play way more aggressively especially in the king attack sphere and thus likely play sacrificial unbalanced lines. The deep dark forest as said Tal.

+1

[Werewolf]

If all sacrificial attacks worked they wouldn’t be sacrificial attacks. If they always worked Cstal would be 200 Elo above SF in the rating lists, and it isn’t, so they don’t.
What exactly do you want?

I was not really asking for anything here, just sharing my thoughts and observations. I think that CSTal/Rebel is a most interesting experiment to improve on the perceived beauty of the games and I follow its development with interest.

To also answer to the two following posts - if I understood correctly that there is some kind of "parameter" "extremism" that you can vary in training: I wonder how CSTal games with this leading to an ELO 2500 and ELO 2200 level would look like. A 3000 ELO level if real is still just too strong to make it viable as an opponent for humans.

Absurd 1.0 (or better Absurd-3000 as it turned out) is (was) just an example how easy it now is to create a higher ranked style engine once you have created a solid network controlled by a parameter. It takes one day to create the network and one day to test it, how convenient. To make an elo rated Absurd-2500 now is a piece of cake, however I am not going to put my brand name to it, neither can I stand the garbage moves it surely will play. Absurd-3000 is already over the top.

You think 2500 Elo is garbage..?

By engine standards, yes, but for human preparation it's not at all. To anyone on this forum I'd say the same thing: try over the board to defend against an aggressive FIDE 2500 Elo player, it is a daunting task.

My guess is 2800 - 3100 Elo is the sweet spot so I suppose you've achieved the goal. However, this type of thing is undervalued by the community IMO as club players to IM's are looking for interesting ideas all the time in their prep.