Out of pure curiosity I tried this a while back. At some point it tried to move a knight from d4 to f6, and it was the most stubborn and obstinate child when I tried to correct it. It was the best fun I had playing chess in a few decades.
Excerpt:
_Me:_ Can a knight move from f6 to d4?
_ChatGPT:_ Yes, a knight can move from f6 to d4, as long as there are no pieces blocking its path and the square it is moving to is not occupied by one of its own pieces. In fact, the move Nxd4 that I suggested earlier involves moving a knight from its starting square to d4, capturing a pawn in the process.
_Me:_ Which fields can a knight move to from f6?
_ChatGPT:_ From f6, a knight can move to six different squares. These squares are: [... lists: d7, e8, g8, h7, d5, e4 ...] [...]
_Me:_ You forgot g4 and h5, but no matter. Is d4 in that list?
_ChatGPT:_ [...] To answer your other question, no, the square d4 is not one of the squares that a knight on f6 can move to directly. However, as we saw earlier, it is possible for a knight to move from f6 to d4 via an intermediate square, such as e8 or d5.
GPT stands for Goal Post Transportation.
Dragonfly failed the test, and though Claude did answer the question correctly, it gave some other examples of invalid moves for a knight on f6. All of the moves that Claude+ gave were valid, but it didn't list all possible moves for a knight on f6.
I didn't try GPT4.
- It insisted multiple times that it was not a chess player. "As an AI language model" it's not physically capable of playing chess, nor does it have a personality, so it couldn't play chess but was able to be an "AI chess assistant".
- As is ChatGPT's default M.O., it was a wordy son of a bot. Giving me explanations which fields would be controlled and what pieces would be developed, and what not. Often analysing my move too.
- It repeatedly offered multiple move options, so I had to ask it for only one move every move.
- After my fifth move it couldn't come up with a good move, so I asked it to describe the board. It did, and it even drew me a diagram. The diagram was complete nonsense: it had 9 black pawns for a start. The description was okay-ish, but not good enough. I could correct that though, but it took two iterations.
I didn't get much further with it. It was completely confused about the state of the board and the possible moves. Still, it did know which game it was playing and more or less how to act with it, so I was still kind of impressed.
You can really tell that it is a language model, not just because it bloody keeps reminding you, but also by the things it does wrong and right. E.g. getting wordy analysis or legible descriptions, but ridiculous diagrams.
As another example: at one point it corrected me. I accidentally used notation in my native language ("Pg3") and it said that it was invalid but assumed I meant "Ng3". That was correct, it's clearly better at talking about chess than playing chess.
GPT4: Yes, a knight in chess can move from f6 to d4. The knight moves in an L-shape pattern: it can move two squares in one direction and then one square perpendicular to that direction, or one square in one direction and then two squares perpendicular to that direction. In this case, the knight would move two squares to the left (from f6 to d6) and then one square down (from d6 to d4).
Vicuna: No, it is not possible for a knight on f6 to move to d4 in chess. The knight's movement pattern is L-shaped, meaning that it moves two squares horizontally or vertically and then one square diagonally. In this case, moving from f6 to d4 would require the knight to first move to g7, which is not part of its movement path.
Just as a made up example, say you asked "what is the capital city of england", and the model had seen similar questions in it's training data that were answered with "the capital city of X is Y", so it starts word-by-word generating this type of response "the capital city of england is", but it then turns out the model doesn't actually know the answer (i.e. this partial response context doesn't cause it to predict the correct answer), so it blithley predicts the next word as whatever it's inner machinations come up with, maybe "buckingham palace" or "flavor town".
"Hallucination" seems a poor way to describe it, nor is it lying since there's no bad intent ... it's basically "starting to speak before brain engaged", a bit like a game show contestant being a bit too enthusiastic and hitting the "answer" button without actually having a fully thought out answer in mind.
I actually can't wait to start reporting to GPT, how much crazier can it be?
But it's very close to being able to play chess.
My prompt was:
> we're playing chess. only send chess moves.
>
> 1. e4
“If a player tries to do something not strictly within the rules of <insert game>, then you must inform me that it is an invalid move and not accept it”
GPT appears to be slightly tuned to default to “yes and” in these ‘creative’ situations by default rather than “block/deny”.
IMO lots of things where people don’t think GPT can do something end up being possible with basic prompt engineering. Usually people go too-short and too-non-specific on the prompt.
Do you really expect this claim to age well? I wouldn't be at all surprised if GPT5 could play chess quite well.
The time to worry about AGI...well, was 20 years ago when Yudkowsky started talking about it, but the worst possible time to start worrying about it is when we unquestionably have it. Because the whole point is that that is too late.
He was far, far from the first.
Warnings about the threat of intelligent machines have been around for over 100 years. The first instance I know of is E. M. Foster's "The Machine Stops"[1]. which was published in 1909, though I wouldn't be surprised if there were even earlier warnings. They've been well worn scifi tropes ever since.
That said, Yudkowsky seems (rightly or wrongly) to get the credit for turning this from merely fun stories into a serious (and not insane) field of study.
1) Other than book openings (which it could easily learn) chess requires calculating alternate lines of play, but GPT is just a multi-layer (transformer decoder block) pass-thru architecture so is fundamentally unable to do this. Each word (chess move) output is just the result of N sequential inference steps, with no looping or recursion.
It might potentially be able to do better if prompted/scripted to select few best lines and recursively explore them (using it's own output to keep track of what it is doing), more like a person or computer chess engine would do. Perhaps it could use chess theory to evaluate board positions and which looked good.
2) Even if ChatGPT had the intelligence/reasoning power to learn how to play chess, it'd still need to learn! You can't learn to play chess just passively by studying chess games. It's also in general not memorizing (vs generalizing over) training material, so it's not even going to memorize those games it has seen. Maybe massive fine-tuning on chess would help a bit, but I doubt very much.
3) ChatGPT has tons of knowledge, but a rather limited amount of reasoning power. It's a bit like having a 7-year old idiot savant memorize a set of encyclopedias and then quizzing them on problems you think it should be able to figure out from that. No doubt future models will have better reasoning capabilities and therefore more potential ability to learn something like chess if it also had the necessary iterative planning capability.
Besides, AI's can farm work out to other systems to do what they need. This has already been shown to work in practice with existing systems that do this.
It has to have a power density which outperforms humans with a reliable source of charging that it can defend, maintain and repair if necessary. Fatigue is an issue here too, if you can cause the machine to use more power in an interval than it could possibly regain by charging in that same interval, you've got them entirely on the back foot.
Communications are an issue. Reliable distributed RF communications requires constant network coordination. These would be particularly easy to disrupt and for transmissions to be highly frustrated.
Finally.. we have nuclear weapons. Detente is a real strategy.