The Nostradamus experiment

[David Carteau]

Hi everyone !

I have recently been working on trying to train small language models to help me to understand the "magic" behind the transformer architecture which is used in large language models like GPT, Llama, etc.

I started by trying to replicate the work of this excellent paper ("TinyStories: How Small Can Language Models Be and Still Speak Coherent English?"), to answer THE ultimate question: "Can cows fly ?" (see paper... ) Then I switched to building a translation model (Greek to English) to better understand the tokenisation process. Finally (and this was the real goal), I tried to train such a small language model to... play chess

Three weeks ago, I released a new experimental UCI chess engine, Nostradamus:
* unlike other engines, it doesn't rely on a traditional search tree to find the best combination of moves
* it uses instead a language model to predict the best move given a specific position



At this stage, the engine is very weak (depth = 1), but there's some kind of magic happening here: the transformer architecture seems to be a real game changer ! I will continue my experiments, and if anyone is interested or wants to contribute, do not hesitate to contact me !

David

[Jim Ablett]

Here is a Windows executable download package for anyone who wants to try it out > https://mega.nz/file/yEVWwR7K#uzAmAuwdp ... Wi7TMJqCE4

It is a large download (1.6gb) as it also contains the Torch backend dlls. You may get an antivirus warning with this. This is something
that happens when you convert Python to executable - nothing I can do about it.

Jim.

[Eelco de Groot]

Thanks guys for the engine that can see the future! Most things seem to work fine and no virus detection with Malwarebytes. I could not get analysis to work very well in Shredder GUI but that may be the implementation, reminded me a bit CSTal in Shredder (in Analysis mode I mean). It could give a move here: viewtopic.php?t=84638 (See the diagram below)

You may have to restart the GUI after installation also to give the GUI enough time to load all of the program, it takes a while.

From command prompt I check 'uci' command:

uci
id name Nostradamus 0.2
id author David Carteau
option name Hash type spin default 0 min 0 max 0
uciok



[d]4rrk1/p1p2p1p/1pbq1b2/4N2Q/3Pp3/8/PPP4P/R1B1K2R w KQ -

Engine: Nostradamus 0.2 (1024 MB)
gemaakt door David Carteau

beste zet: Th1-g1 tijd: 0:06.422 min

Correct move but I did not get it to play any next move not with Black or any with White (in Shredder Analysis mode, maybe in Arena it would work different?)

Nostradamus vs Nostradamus, there were some illegal moves but they were stopped, I do not know if Shredder did not accept them or the other Nostradamus, with comments like "Are you kidding Nostradamus?" "That's an illegal move".

[pgn][Event "Chess in 2035"]
[Site "Ye olde Dell computer from 2014"]
[Date "2024.12.23"]
[Round "?"]
[White "Nostradamus 0.2"]
[Black "Nostradamus 0.2"]
[Result "1/2-1/2"]

1. e4 {Boek 0s} e6 {Boek 0s} 2. d4 {Boek 0s} d5 {Boek 0s}
3. Nc3 {Boek 0s} Bb4 {Boek 0s} 4. Nge2 {Boek 0s} Nf6 {Boek
0s} 5. e5 {Boek 0s} Nfd7 {Boek 0s} 6. a3 {Boek 0s} Be7
{Boek 0s} 7. f4 {Boek 0s} O-O {Boek 6s} 8. Ng3 {Boek 4s} c5
{Boek 4s} 9. dxc5 {Boek 4s} Bxc5 {Boek 4s} 10. Bd3 {Boek
4s} Nc6 {Boek 4s} 11. Kd2 {Boek 5s} Nd4 {Boek 4s} 12. b4
{Boek 4s} Bb6 {Boek 4s} 13. Bb2 {Boek 4s} f6 {Boek 4s}
14. exf6 {Boek 4s} Nxf6 {Boek 4s} 15. Qe1 {Boek 4s} a5
{Boek 4s} 16. Nh5 {Boek 5s} axb4 {Boek 4s} 17. Nxf6+ {Boek
4s} Qxf6 {Boek 4s} 18. axb4 {Boek 4s} Rxa1 {Boek 4s}
19. Bxa1 {Boek 5s} Bd7 {Boek 4s} 20. g3 {Boek 4s} Ra8 {Boek
4s} 21. Bb2 {Boek 4s} Nf3+ {Boek 4s} 22. Kd1 {Boek 4s} Nxe1
{Boek 4s} 23. Rxe1 {Boek 4s} Ra1+ {Boek 4s} 24. Bxa1 {Boek
4s} Bf2 {Boek 4s} 25. Re2 {Boek 4s} Bg1 {Boek 4s} 26. Bb2
{Boek 4s} Bc6 {Boek 4s} 27. b5 {Boek 4s} Bd7 {Boek 4s}
28. Na4 {Boek 4s} Bxh2 {Boek 4s} 29. Nc5 {Boek 4s} Bxg3
{Boek 4s} 30. Nxd7 {Boek 4s} Qxf4 {Boek 4s} 31. Bc3 {Boek
4s} Qf3 {Boek 4s} 32. Kd2 {Boek 4s} h5 {Boek 4s} 33. Bd4
{Boek 4s} h4 {Boek 4s} 34. Re3 {Boek 4s} Qf4 {Boek 5s}
35. c4 {Boek 5s} Qxd4 {Boek 4s} 36. cxd5 {Boek 4s} exd5
{Boek 4s} 37. Re8+ {Boek 4s} Kf7 {Boek 4s} 38. Rf8+ {Boek
4s} Ke6 {Boek 4s} 39. Re8+ {Boek 4s} Kxd7 {Boek 4s} 40. Re2
{Boek 4s} Bf4+ {Boek 4s} 41. Kc2 {Boek 4s} h3 {Boek 4s}
42. Re1 {Boek 4s} Be3 {Boek 4s} 43. Rh1 {Boek 4s} h2 {Boek
4s} 44. Rxh2 {Boek 4s} Qc5+ {Boek 4s} 45. Kd1 {Boek 4s} Bd4
{Boek 4s} 46. Rh7 {Boek 4s} Qc3 {Boek 4s} 47. Rxg7+ {Boek
4s} Kd6 {Boek 4s} 48. Rg6+ {Boek 4s} Ke5 {Boek 4s} 49. Ke2
{Boek 4s} Qb2+ {Boek 4s} 50. Kf3 {Boek 4s} Qf2+ {Boek 4s}
51. Kg4 {Boek 4s} Qf4+ {Boek 4s} 52. Kh5 {Boek 4s} Qh2+
{Boek 4s} 53. Kg4 {Boek 4s} Qg2+ {Boek 4s} 54. Kh5 {Boek
4s} Qh3+ {Boek 4s} 55. Kg5 {Boek 4s} Qg3+ {Boek 4s} 56. Kh5
{Boek 4s} Qh3+ {Boek 4s} 57. Kg5 {Boek 4s} Qg3+ {Boek 4s}
58. Kh5 {Boek 4s} Qh3+ {Boek 4s} *
[/pgn]

[Tibono]

there were some illegal moves but they were stopped, I do not know if Shredder did not accept them or the other Nostradamus, with comments like "Are you kidding Nostradamus?" "That's an illegal move".

Hi,
I also wanted to test some games and it is definitely fun. Thanks David for the interesting work and sharing, thanks Jim for the compile.
I was much surprised most (if not each and every) move make sense, a priori. It does look to play chess! The weakness is mainly due to blunders: moves that look good at first glance, but are easily refuted.
About illegal moves: I understand the program controls the compliancy and if necessary, answers a fallback move. I would be curious to know how is the fallback move chosen? Is it outside the language/translation model?

[Sylwy]

Interesting experiment ! Thanks to both: the author-David Carteau and Maestro Jim !

[Sylwy]

[pgn]
[Event "EXPERIMENT"]
[Site "ISR 3"]
[Date "2024.12.23"]
[Round "3"]
[White "Nostradamus_02_x64_ja"]
[Black "Zoe-01-ja"]
[Result "1/2-1/2"]
[BlackElo "2200"]
[ECO "A45"]
[Opening "Indian"]
[Time "19:56:30"]
[Variation "Omega Gambit"]
[WhiteElo "2000"]
[TimeControl "240+2"]
[Termination "normal"]
[PlyCount "108"]
[WhiteType "program"]
[BlackType "program"]

1. d4 Nf6 2. e4 Nxe4 {(# pv: f6e4 b1d2 e4d2 c1d2 b8c6 d4d5 e7e5 115)
+1.15/7 18} 3. Bd3 d5 {(# pv: d7d5 b1d2 e4f6 d3b5 c8d7 b5d7 b8d7 130)
+1.30/7 12} 4. Nf3 Nc6 {(# pv: b8c6 c2c4 c8g4 c4d5 g4f3 d3e4 -50) -0.50/6
5} 5. O-O Nb4 {(# pv: c6b4 c2c4 b7b5 d3e4 d5e4 c4b5 -25) -0.25/6 3} 6. Bxe4
dxe4 {(# pv: d5e4 f3d2 c8f5 d1h5 f5e6 d2e4 -55) -0.55/6 19} 7. Ne5 c5 {(#
pv: c7c5 d1e2 c5d4 e2b5 c8d7 b5b4 -85) -0.85/6 4} 8. dxc5 Qxd1 {(# pv: d8d1
f1d1 b4c2 e5f7 e8f7 b1c3 280) +2.80/6 4} 9. Rxd1 Nxc2 {(# pv: b4c2 d1d8
e8d8 e5f7 d8c7 f7h8 55) +0.55/6 3} 10. Nc3 Bg4 {(# pv: c8g4 e5g4 c2a1 c3e4
a1c2 c5c6 105) +1.05/6 10} 11. Nxg4 Nxa1 {(# pv: c2a1 g4e5 a1c2 e5d7 e7e5
d7f8 60) +0.60/6 9} 12. Nxe4 f5 {(# pv: f7f5 c1g5 a1c2 d1d2 c2b4 g5e7 f5e4
440) +4.40/7 4} 13. Ne5 fxe4 {(# pv: f5e4 c1d2 a8d8 c5c6 a1c2 e5d7 b7c6
d7f8 315) +3.15/8 8} 14. Be3 Nc2 {(# pv: a1c2 e3g5 a8c8 c5c6 b7c6 g5f4 c2d4
590) +5.90/7 2} 15. Bf4 e6 {(# pv: e7e6 c5c6 b7c6 e5c6 f8c5 c6e5 c5f2 620)
+6.20/7 3} 16. Rd7 Bxc5 {(# pv: f8c5 d7c7 b7b6 b2b4 c2b4 c7g7 c5f2 715)
+7.15/7 3} 17. Rxb7 O-O {(# pv: e8g8 g2g3 c5d4 e5c4 e6e5 c4e5 460) +4.60/6
2} 18. Bg3 Rfc8 {(# pv: f8c8 e5f7 c2b4 f7g5 b4a2 g5e4 465) +4.65/6 2} 19.
Nd7 Bd4 {(# pv: c5d4 h2h3 c8d8 d7e5 e4e3 e5d3 e3f2 680) +6.80/7 3} 20. Rc7
Bxb2 {(# pv: d4b2 c7c8 a8c8 g3e5 c8c4 e5g3 e6e5 715) +7.15/7 1} 21. Rxc8+
Rxc8 {(# pv: a8c8 h2h4 c2b4 g3e5 c8c1 g1h2 b2e5 d7e5 685) +6.85/8 2} 22.
Nc5 Rxc5 {(# pv: c8c5 g3b8 c5a5 a2a3 a5a3 b8f4 e6e5 1145) +11.45/7 2} 23.
h4 Ra5 {(# pv: c5a5 g3d6 a5a2 d6c5 e6e5 h4h5 c2d4 1180) +11.80/7 1} 24. Bc7
Rxa2 {(# pv: a5a2 h4h5 c2d4 g1h2 d4f5 h2g1 b2d4 c7g3 1140) +11.40/8 3} 25.
Kf1 Bf6 {(# pv: b2f6 h4h5 c2d4 f1g1 a2a1 g1h2 a1a2 h2g1 1155) +11.55/8 4}
26. h5 Rb2 {(# pv: a2b2 f1g1 b2b5 g2g4 b5b1 g1g2 c2d4 g2g1 1160) +11.60/8
4} 27. h6 gxh6 {(# pv: g7h6 c7f4 f6g5 f4e5 b2b1 f1e2 g5e7 e2d1 1240)
+12.40/8 3} 28. Ke2 Nd4+ {(# pv: c2d4 e2e1 e6e5 g2g4 b2b7 g4g5 b7c7 g5f6
1280) +12.80/8 5} 29. Ke3 Rc2 {(# pv: b2c2 c7g3 c2e2 e3f4 f6h4 f4g4 h4g3
g4g3 1300) +13.00/8 6} 30. Kxe4 Rxc7 {(# pv: c2c7 e4d3 c7g7 g2g3 d4f5 d3c2
e6e5 c2b1 1485) +14.85/8 4} 31. g3 Rc2 {(# pv: c7c2 f2f4 c2c3 g3g4 d4c6
f4f5 e6f5 g4f5 1505) +15.05/8 3} 32. f4 Rg2 {(# pv: c2g2 e4d3 g2g3 d3e4
g3g2 e4f5 f6g7 f5e6 1540) +15.40/8 3} 33. g4 Rxg4 {(# pv: g2g4 e4f5 g4g6
f5e4 d4c2 e4f3 c2d4 f3g2 1565) +15.65/8 2} 34. Ke3 Rg3+ {(# pv: g4g3 e3e4
f6g7 e4e5 g3c3 e5e4 c3c2 e4e3 1610) +16.10/8 2} 35. Kf2 Bh4 {(# pv: f6h4
f4f5 g3g4 f2e3 e6e5 e3d2 g4g1 d2c3 d4f5 1720) +17.20/9 8} 36. Kf1 Rg4 {(#
pv: g3g4 f4f5 e6e5 f5f6 h4f6 f1f2 e5e4 f2g1 1700) +17.00/8 1} 37. f5 e5 {(#
pv: e6e5 f5f6 h4f6 f1e1 g4g2 e1f2 g2h2 f2g1 1700) +17.00/8 1} 38. f6 Rf4+
{(# pv: g4f4 f1g2 f4f3 g2g1 h4f6 g1f2 f3b3 f2g1 1695) +16.95/8 2} 39. Kg2
Rf3 {(# pv: f4f3 f6f7 g8g7 g2g1 f3f7 g1f2 h4f6 f2g1 1685) +16.85/8 3} 40.
f7+ Kg7 {(# pv: g8g7 f7f8q f3f8 g2g3 h4g5 g3f2 g5f4 f2g1 f4g3 g1f1 g3f4
f1g2 g7g8 1700) +17.00/9 4} 41. f8=Q+ Rxf8 {(# pv: f3f8 g2f3 d4c6 f3e4 g7g8
e4e5 f8f2 e5e4 h4f6 1625) +16.25/9 3} 42. Kh3 Bf2 {(# pv: h4f2 h3g4 f2e3
g4h5 g7g8 h5h6 f8f2 h6h7 1555) +15.55/8 3} 43. Kg2 Be1 {(# pv: f2e1 g2f3
e1c3 f3e4 d4c6 e4e5 g7g8 e5f4 1600) +16.00/8 2} 44. Kh3 Bd2 {(# pv: e1d2
h3g4 d2e3 g4h5 g7g8 h5h6 f8f2 h6h7 1555) +15.55/8 6} 45. Kg2 Bc3 {(# pv:
d2c3 g2f3 d4c6 f3e4 g7g8 e4e5 f8f2 e5f4 1610) +16.10/8 6} 46. Kg3 h5 {(#
pv: h6h5 g3f4 h5h4 f4e5 h4h3 e5e4 h3h2 1665) +16.65/7 1} 47. Kg2 a5 {(# pv:
a7a5 g2f3 a5a4 f3g4 f8f5 g4h3 f5f2 1725) +17.25/7 1} 48. Kg3 a4 {(# pv:
a5a4 g3f4 a4a3 f4e5 a3a2 e5e4 a2a1q 2455) +24.55/7 1} 49. Kg2 a3 {(# pv:
a4a3 g2f3 a3a2 f3e4 d4c6 e4d4 a2a1q d4c3 2230) +22.30/8 3} 50. Kg3 a2 {(#
pv: a3a2 g3f4 a2a1q f4e5 a1b1 e5d5 b1d3 2490) +24.90/7 1} 51. Kg2 a1=Q {(#
pv: a2a1q g2f1 f8a8 f1g2 a8a2 g2h3 a1b2 2590) +25.90/7 1} 52. Kh2 e4 {(#
pv: e5e4 h2g3 e4e3 g3g4 e3e2 g4h5 e2e1q 3315) +33.15/7 4} 53. Kg2 e3 {(#
pv: e4e3 g2f1 e3e2 f1g1 a1b2 g1h1 e2e1q 3385) +33.85/7 6} 54. Kg3 Qh1 {(#
pv: a1h1 g3g2 d4e2 g2h1 f8f2 h1g1 e2g1 4760) +47.60/7 4 Stalemate} 1/2-1/2
[/pgn]

[Sylwy]

NICE ?!?

[pgn]
[Event "EXPERIMENT"]
[Site "ISR 3"]
[Date "2024.12.23"]
[Round "7"]
[White "Nostradamus_02_x64_ja"]
[Black "Monchester_0.99_x64"]
[Result "1-0"]
[BlackElo "2200"]
[ECO "C00"]
[Opening "French"]
[Time "21:06:47"]
[Variation "2.d4"]
[WhiteElo "2200"]
[TimeControl "240+2"]
[Termination "normal"]
[PlyCount "65"]
[WhiteType "program"]
[BlackType "program"]

1. d4 e6 {(4 7020 0 e7e6 b1c3 g8f6 d1d3) -0.42/4 0} 2. e4 Nc6 {(4 6124 0
b8c6 b1c3 g8f6 g1f3) -0.28/4 0} 3. Nf3 Nf6 {(4 8151 0 g8f6 c1g5 f8b4
b1d2) -0.37/4 0} 4. Bd3 Bb4+ {(4 0 0 f8b4 c2c3 b4a5 e1f1) -0.28/4 0} 5. c3
Ke7 {(4 6756 0 e8e7 c3b4 c6b4 d3f1) -7.28/4 0} 6. cxb4 Nxb4 {(4 7431 0
c6b4 c1g5 b4d3 d1d3) -8.42/4 0} 7. Nc3 Ne8 {(4 6112 0 f6e8 f3g5 h7h6
g5e6) -7.94/4 0} 8. Be2 Nc6 {(4 5647 0 b4c6 d1a4 c6b8 a4a7) -8.34/4 0} 9.
O-O a5 {(4 6355 0 a7a5 c1g5 f7f6 g5f6) -8.02/4 0} 10. d5 exd5 {(4 6677 0
e6d5 d1d5 c6b8 d5b7) -8.40/4 0} 11. Nxd5+ Kf8 {(4 6554 0 e7f8 e2b5 c6a7
d5c7) -8.22/4 0} 12. Bg5 f6 {(4 6973 0 f7f6 a1c1 c6b8 d5f6) -8.60/4 0} 13.
Be3 Nb8 {(4 23510 0 c6b8 d1d2 c7c6 d5f6) -8.28/4 0} 14. Rc1 g6 {(4 7455 0
g7g6 d1d3 h7h5 d5c7) -8.31/4 0} 15. Bh6+ Kf7 {(4 7854 0 f8f7 f1e1 d7d6
d5f6) -8.34/4 0} 16. e5 c6 {(4 6564 0 c7c6 d5f6 e8f6 e5f6) -8.00/4 0} 17.
exf6 cxd5 {(4 7202 0 c6d5 h6f4 e8f6 f4b8) -6.77/4 0} 18. Qxd5+ Kxf6 {(4 0
0 f7f6 e2c4 a5a4 d5e5) -M143/4 0} 19. Qe5+ Kf7 {(4 5861 0 f6f7 e5h8 d7d5
h8e8) -33.42/4 0} 20. Qxh8 Ke6 {(4 5469 0 f7e6 h8f8 d7d6 f8e8) -33.57/4 0}
21. Qxh7 Kf5 {(4 5165 0 e6f5 h7g8 b7b5 g8e8) -34.45/4 0} 22. Qf7+ Qf6 {(4
4151 0 d8f6 f7d5 f6e5 d5e5) -M89/4 0} 23. Qxe8 Na6 {(4 5404 0 b8a6 e2d3
f5g4 d3a6) -44.20/4 0} 24. Bd3+ Kg4 {(4 5182 0 f5g4 d3g6 f6f3 g2f3) -M51/4
0} 25. h3+ Kh5 {(4 5649 0 g4h5 h6g5 f6g5 f3g5) -M13/4 0} 26. g4+ Kxh6 {(4
6235 0 h5h6 g4g5 f6g5 f3g5) -M100/4 0} 27. g5+ Qxg5+ {(4 3788 0 f6g5 f3g5
h6g5 d3a6) -M94/4 0} 28. Nxg5 Kxg5 {(4 6025 0 h6g5 e8g6 g5f4 g6f6) -M143/4
0} 29. h4+ Kf6 {(4 0 0 g5f6 e8g6 f6e7 g6a6) -M50/4 0} 30. Bxg6 Rb8 {(4
5905 0 a8b8 f2f4 f6g7 e8c8) -M107/4 0} 31. Qf7+ Ke5 {(4 4765 0 f6e5 c1d1
a5a4 d1d5) -M143/4 0} 32. Qe7+ Kd5 {(4 0 0 e5d5 f1e1 a5a4 e1d1) -M143/4 0}
33. Rfd1# 1-0
[/pgn]

[Tibono]

A short won game against a very weak chess computer. But what is interesting is the way it does play and win.
Nostradamus didn't spot the mate in 1 on move #17 and #19 yet played rather well. No opening book. Only average 10cpl, according to HIARCS analysis. I think it can be a good sparring partner for beginners!
[pgn][Event "Arena Tournament.at"]
[Site "Arena"]
[Date "2024.12.22"]
[Round "1"]
[White "Nostradamus 0.2"]
[Black "MC Novag CC MK I"]
[Result "1-0"]
[WhiteElo "689"]
[BlackElo "723"]
[ECO "D00"]
[PlyCount "39"]
[BlackType "program"]
[Opening "Colle"]
[Termination "normal"]
[Time "16:46:20"]
[TimeControl "40/600:40/600:40/600"]
[Variation "3...Bg4"]
[WhiteType "program"]

1.d4 d5 2.Nc3 Nc6 $6 ( {Theory prefers} 2...Nf6 $1 ) 3.e3 $6 {Deep HIARCS 15.3.1: -0.09/20} ( 3.e4 {Deep HIARCS 15.3.1: +0.42/16} 3...dxe4 4.d5 Nb8 5.Bf4 Nf6 6.f3 c6 7.dxc6 Nxc6 8.Qxd8+ Nxd8 9.fxe4 Ne6 10.Be5 a6 11.Nf3 g6 12.O-O-O Bg7 ) 3...Nf6 4.Nf3 Bg4 5.Bb5 Bxf3 6.Qxf3 e6 7.O-O h6 8.Bd2 Qd7 9.e4 dxe4 10.Nxe4 Nxe4 11.Qxe4 Qxd4 $4 {Deep HIARCS 15.3.1: +16.78/20} ( 11...O-O-O {Deep HIARCS 15.3.1: +0.55/22} 12.c3 a6 13.Ba4 Qd5 14.Qxd5 exd5 15.b4 Be7 16.Rae1 Bf6 17.Bd1 b6 18.g3 Kb7 19.a4 a5 20.b5 Ne7 21.Bg4 g6 22.Kg2 Nf5 ) 12.Qxd4 Rd8 13.Bxc6+ bxc6 14.Qxa7 $1 Rxd2 15.Qxc7 Rxc2 16.Rad1 Rxb2 17.Qxc6+ $6 {Deep HIARCS 15.3.1: +29.01/17} ( 17.Rd8# {Deep HIARCS 15.3.1: #1/121} ) 17...Ke7 18.Rd7+ Ke8 19.Rfd1 Rxa2 20.Qc8# {White loses 0.1 pawn per move (1 inaccuracy), Black loses 0.46 pawn per move (1 blunder).} 1-0

[/pgn]

[mclane]

Yes it could be a good partner for beginners or dedicated chess computers.

[David Carteau]

Here is a Windows executable download package for anyone who wants to try it out > https://mega.nz/file/yEVWwR7K#uzAmAuwdp ... Wi7TMJqCE4
[...]

Thanks Jim for making these binaries !

Thanks guys for the engine that can see the future!
[...]
[d]4rrk1/p1p2p1p/1pbq1b2/4N2Q/3Pp3/8/PPP4P/R1B1K2R w KQ -
[...]
Correct move but I did not get it to play any next move not with Black or any with White (in Shredder Analysis mode, maybe in Arena it would work different?)

Thank you Eelco for this feedback: I'm sorry, currently the engine does not support 'position fen ... moves ...' commands (that's why in your example it only responded to 'position fen ...', but not after 'position fen ... moves ...'). It's just one line of code to add, but I'm far from home for the holidays, so I'll fix it in early January

Nostradamus vs Nostradamus, there were some illegal moves but they were stopped, I do not know if Shredder did not accept them or the other Nostradamus, with comments like "Are you kidding Nostradamus?" "That's an illegal move".

Yes, the comments are from the engine, a kind of "Easter-Egg"

I also wanted to test some games and it is definitely fun. Thanks David for the interesting work and sharing, thanks Jim for the compile.

Hi Eric, thanks for your positive feedback!

About illegal moves: I understand the program controls the compliancy and if necessary, answers a fallback move. I would be curious to know how is the fallback move chosen? Is it outside the language/translation model?

That's right. If the predicted move is illegal, the engine uses a different selection method and randomly chooses a legal move. For sure, it might be better to use the second best prediction if the first one is irrelevant, and so on... (I will add this in the next version).

Interesting experiment ! Thanks to both: the author-David Carteau and Maestro Jim !

Thanks Selwy! If you play several games against Zoe, I will be curious to see the results!

Yes it could be a good partner for beginners or dedicated chess computers.

Hi Mclane, I hope you have managed to get the engine installed and running, thank you for your interest!

Merry Christmas

David