Skynet awakens

[Werewolf]

ChatGPT is a LLM and as such cannot play a game of chess. Well now it can.

I’ve tried this numerous times over the last year only to be told chess was too hard for it. But now it plays maybe 1300 Elo.
I was white below and with a stinking cold.

But I imagine it’ll get better.

Also we were chatting as we were playing.

1.e4 e5 2.Nf3 Nc6 3.Bb5 a6 4.Bxc6 dxc6 5.O‑O Nf6 6.Nxe5 Bd6 7.Nc4 Be7 8.d3 O‑O 9.Nc3 Be6 10.Ne3 Bc5 11.Qf3 Nd7 12.Qg3 Kh8 13.Kh1 Bd6 14.f4 f5 15.Qh3 Qf6 16.e5 Nxe5 17.fxe5 Qxe5 18.Nc4 Qc5 19.Be3 Qb4 20.a3 Qxc4 21.dxc4 Bxc4 22.Rxf5 Rxf5 23.Qxf5 Re8 24.Bd4 c5 25.Bxc5 Bxc5 26.Qxc5 Re1+ 27.Rxe1

The machine, totally lost, had tracking errors here.

[AdminX]

Things that might start to keep me up at night ...

[Werewolf]

What I can’t understand is how it plays. There can’t be any search going on.
Maybe it has a basic evaluation function and plays accordingly without any look ahead?

[smatovic]

...my guess is (currently) it has no internal board representation and just predicts the next move in line as a string completion, maybe trained on PGN data in the web, so there might be some kind of knowledge between lines/moves present, or alike.

--
Srdja

[dkappe]

[pgn] [Event "?"] [Site "?"] [Date "2023.09.28"] [Round "1"] [White "dkappe"] [Black "ChatGPT-4"] [Result "*"] 1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 a6 6. Be2 e5 7. Nb3 Be7 8. g4 Be6 9. g5 Nfd7 10. h4 Nc6 11. Be3 Nb6 12. Qd2 * [/pgn]


But then it came with the master stroke:

You've played 12. Qd2, connecting your rooks. I'll castle queenside with 12... O-O-O.

Your move!

How did it get so good?

[dkappe]

[pgn] [Event "?"] [Site "?"] [Date "2023.09.28"] [Round "1"] [White "dkappe"] [Black "ChatGPT-4"] [Result "*"] 1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 a6 6. Be2 e5 7. Nb3 Be7 8. g4 Be6 9. g5 Nfd7 10. h4 Nc6 11. Be3 Nb6 12. Qd2 Qc7 13. O-O-O Rd8 14. Kb1 d5 15. Bxb6 Qxb6 16. exd5 Bf5 17. Bd3 Bxd3 18. Qxd3 g6 19. Qg3 Bd6 20. dxc6 Qxc6 21. h5 gxh5 22. Nd5 Qa4 23. Nf6+ Ke7 24. Rxd6 Kxd6 25. Rd1+ Ke6 26. Nc5+ Kf5 27. Re1 e4 28. Nxa4 h4 *[/pgn]

After a few more illegal move, I corrected the AI until it got boring.

[AdminX]

[pgn] [Event "?"] [Site "?"] [Date "2023.09.28"] [Round "1"] [White "dkappe"] [Black "ChatGPT-4"] [Result "*"] 1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 a6 6. Be2 e5 7. Nb3 Be7 8. g4 Be6 9. g5 Nfd7 10. h4 Nc6 11. Be3 Nb6 12. Qd2 Qc7 13. O-O-O Rd8 14. Kb1 d5 15. Bxb6 Qxb6 16. exd5 Bf5 17. Bd3 Bxd3 18. Qxd3 g6 19. Qg3 Bd6 20. dxc6 Qxc6 21. h5 gxh5 22. Nd5 Qa4 23. Nf6+ Ke7 24. Rxd6 Kxd6 25. Rd1+ Ke6 26. Nc5+ Kf5 27. Re1 e4 28. Nxa4 h4 *[/pgn]

After a few more illegal move, I corrected the AI until it got boring.

Ahh, but tomorrow (Near Future) it will be full on Alpha/ LC0 like.

[dkappe]

I play chess by analyzing the board position, calculating potential moves, and evaluating the resulting positions. Here's a breakdown of the process:

1. **Board Representation**: I internally represent the chessboard and the positions of the pieces. When you provide a move, I update this representation.

2. **Move Generation**: I generate all legal moves for the position. This considers the specific rules for each piece and additional rules like castling and en passant.

3. **Evaluation**: For each move, I assess the resulting position. The evaluation is based on various factors such as material balance, piece activity, pawn structure, king safety, and other positional considerations.

4. **Search**: I use a depth-first search algorithm, specifically the minimax algorithm with alpha-beta pruning, to explore possible sequences of moves (or "lines"). This helps me anticipate both my and my opponent's best responses.

5. **Time Management**: Depending on the context (like in our casual game), I might not dive very deep into the search tree, ensuring quick but reasonable moves. In more competitive contexts or when given more time, I can search deeper.

6. **Selection**: After evaluating a certain depth of moves and counter-moves, I pick the move that I believe leads to the best position for me.

It's worth noting that while I can play chess at a high level due to my computational abilities, I don't "intuit" or "feel" the game like humans do. I don't have emotions or intuition; I rely on calculation, pattern recognition, and positional evaluation algorithms.

When asked why it made illegal moves, it was a little vague. But it looks like ChatGPT is just responsible for parsing the move and passing it to a conventional, albeit weak, chess engine.

[Werewolf]

[pgn] [Event "?"] [Site "?"] [Date "2023.09.28"] [Round "1"] [White "dkappe"] [Black "ChatGPT-4"] [Result "*"] 1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 a6 6. Be2 e5 7. Nb3 Be7 8. g4 Be6 9. g5 Nfd7 10. h4 Nc6 11. Be3 Nb6 12. Qd2 Qc7 13. O-O-O Rd8 14. Kb1 d5 15. Bxb6 Qxb6 16. exd5 Bf5 17. Bd3 Bxd3 18. Qxd3 g6 19. Qg3 Bd6 20. dxc6 Qxc6 21. h5 gxh5 22. Nd5 Qa4 23. Nf6+ Ke7 24. Rxd6 Kxd6 25. Rd1+ Ke6 26. Nc5+ Kf5 27. Re1 e4 28. Nxa4 h4 *[/pgn]

After a few more illegal move, I corrected the AI until it got boring.

But it's quite incredible you can correct it.

I wonder how much longer you and I will be able to beat it for...?

[Werewolf]

After I asked it many questions on how it plays:

I don't have a traditional chess engine's "search" capability built into me, nor do I use a specific chess evaluation function like dedicated chess engines such as Stockfish or AlphaZero do. My capability to play chess stems from:

Knowledge: I've been trained on vast amounts of text data, including information about chess, its strategies, openings, middlegame, and endgame. I can generate moves based on patterns and knowledge I've seen.

Pattern Recognition: My architecture is good at recognizing patterns, so I can "see" positions and suggest moves that seem to align with standard chess principles.

General Computation: I can do basic "if-then" style logic to think a few moves ahead, but this is nowhere near as deep or sophisticated as a real chess engine's search tree.

That said, while I can provide general advice, simulate games, and discuss chess, I am not optimized for high-level play or deep analysis. For serious analysis or play, dedicated chess software is recommended.

To be clear, if chatGPT spends longer per move it does not get stronger, because there is no iterative search.