Could an AI figure out the rules of chess from games?

I think so. AI has become very intelligent nowadays.

@Hagredion said in #30:

[...] Well you should ask yourself what is more likely that a super-talent like Capablanca simply noticed that both players constantly only ever move the bishops diagonally and made the logical conclusion that the piece only moves diagonally or that by some miraculous stroke of luck he was present at the exactly moment when one of them made an illegal move, which is a rare occurrence in slow games. The first option is infinitely more likely.

I guess you and I have different experiences regarding illegal moves in casual games played among friends. I used to move into check by accident all the time when I was learning the game, and embarrassingly I actually did it a few weeks ago at a club 50 years later!

I agree that the young Capablanca could easily have worked out how the bishop moves by watching, and in fact other very intelligent 4-year-olds could probably have done the same thing. But noticing that neither player ever made a move which left their king attacked is something else.

@Brian-E said in #32:

I guess you and I have different experiences regarding illegal moves in casual games played among friends. I used to move into check by accident all the time when I was learning the game, and embarrassingly I actually did it a few weeks ago at a club 50 years later!

But Capablanca's dad wasn't learning the game, he was by all accounts a strong amateur player and strong amateur players don't do that kind of mistake often. In any case whether or not the story about Capablanca is true isn't even that important here, the point is that it is at least theoretically possible that a very talented person could learn the rules of chess just by observing the pieces move.
And then the next questions are if a human could do it could an AI do it as well and does such an AI already exist? In my opinion the answer to the first question is obviously yes and the answer to the second question is as far as I know no.

@Hagredion said in #30:

We are not talking about some random boy but about Capablanca who many considered the biggest natural talent to ever play the game. Now I went to check your post #21 and the only example you gave there is the following: "You can't do that, you've moved into check from the bishop." Well you should ask yourself what is more likely that a super-talent like Capablanca simply noticed that both players constantly only ever move the bishops diagonally and made the logical conclusion that the piece only moves diagonally or that by some miraculous stroke of luck he was present at the exactly moment when one of them made an illegal move, which is a rare occurrence in slow games. The first option is infinitely more likely.
Throughout our history apprentices learned from their masters "just by observing", that includes "observing" the body language of their masters to get what they trying to say, at least I think succeeded only those who were getting information from many levels, not just from the plain current state.

@SquareTableKnight said in #34:

Throughout our history apprentices learned from their masters "just by observing", that includes "observing" the body language of their masters to get what they trying to say, at least I think succeeded only those who were getting information from many levels, not just from the plain current state.

Again Capablanca wasn't anybody apprentice his father wasn't even aware that he was learning the game by just observing, that's the whole point of the story. Apart from that you also can't learn how the pieces move by observing anybody's body language, that's just nonsense.

Not too long ago, I heard a story about an AI being told that 3 blankets would dry after about 2 hours of being hung in the sun. It was then asked for the drying time for 6 blankets, hung in the sun. The machine's answer did not make the AI people proud.

Thanks for all the interesting discussion about this.

For the record it doesn't have to be 1 million, it could be however many, trillions or whatever

Maybe someday, but it would take a greatly superhuman AI. Pgns give no hint that it is a game, or that the game involves moving pieces, or that the pieces are on a square board. Deducing all of that from some strings of text would be amazing.

if the question was "could AI generate valid PNGs?"

if you give AI a bunch of books to learn from, it can generate words that are understandable and follows spelling and grammar rules. You can give it the start of a sentence and it will be able to generate words to make it a complete sentence that's valid. I don't think it understands english, though. My stone age understanding is It's generating words based on probability.

therefore

If you give AI a bunch of PGNs to learn from, it should be able to generate valid PNGs. If you give it a partial PNG it should be able to generate new entries to complete it to a win/lose/draw. In theory you could play white and feed it your move and it will generate a valid move for black based on the PNGs you fed it and probabilities.

does that mean it understands the rules of chess? idk. I don't think so.

@ProgrammerAngrim said in #38:

Maybe someday, but it would take a greatly superhuman AI. Pgns give no hint that it is a game, or that the game involves moving pieces, or that the pieces are on a square board. Deducing all of that from some strings of text would be amazing.

It has already been done.
https://thegradient.pub/othello/

They trained an LLM to predict othello moves, E3, D3, C4 etc. The model wasn't given any information about rules or that it is a game played on a board.

After training they probed the network and found that it has an internal representation of the current state of the board when predicting moves. So yes, it figured out that othello is played on a square 8x8 board just from looking at text moves.