i get 3 times the same screenshot in sucession (with the twitter text).
thanks for your blogs!
thanks for your blogs!
Getting the variation wrong... Well i would not say it is just a statistical artefact of the internet.....
Good questions as always (often enough).
Good questions as always (often enough).
Of course it doesn't understand chess positions. Alphazero and other ML chess algorithms, put a reward function into the model and it starts playing against itself.
The first million of moves, the model moves randomly pieces and pawns on the board. From that point on, it starts winning some games using random tactics and strategy. At that point the reward function guides the model to repeat tactics and strategy concepts which enable it to win in the previous games. From then onwards, several hundred million moves are played for the model to refine, invent and repeat winning tactics.
That process is a lot different than just reading png files on the internet.
The first million of moves, the model moves randomly pieces and pawns on the board. From that point on, it starts winning some games using random tactics and strategy. At that point the reward function guides the model to repeat tactics and strategy concepts which enable it to win in the previous games. From then onwards, several hundred million moves are played for the model to refine, invent and repeat winning tactics.
That process is a lot different than just reading png files on the internet.
It's interesting to see that it gets the first few moves right. I was playing with it and by asking it to divide say 2.48728 by 0.38278 it would say 6.4982342393, when the actual answer is 6.49793615... So it got the first few decimals right, but completely made up the rest. The same logic might apply here.
I think it has a lot more examples of opening theory but by the middle-game, it is probably a unique position so it has no reference and goes on to make up annotations/evaluations from whatever few games have similar moves (even in completely different positions).
I think it has a lot more examples of opening theory but by the middle-game, it is probably a unique position so it has no reference and goes on to make up annotations/evaluations from whatever few games have similar moves (even in completely different positions).
@RickFeynman said in #5:
>
Good point. Given the textual transcrition of games into move representation in PGN, mostly SAN moves, without even seeing the implied FEN, most of the time, it would build a move only world of chess.
Like killer move list in engine, or our ability to use at same position, but different thinking stubs, moves from one conceived embryo of variation to another we just winged, missing that newfound move.. But in that case of human thinking, the position were we shuffle moves, are within "nearby" positions.. likely the moves are compatible...
>
Good point. Given the textual transcrition of games into move representation in PGN, mostly SAN moves, without even seeing the implied FEN, most of the time, it would build a move only world of chess.
Like killer move list in engine, or our ability to use at same position, but different thinking stubs, moves from one conceived embryo of variation to another we just winged, missing that newfound move.. But in that case of human thinking, the position were we shuffle moves, are within "nearby" positions.. likely the moves are compatible...
I’ve tried some chess with chatgpt and it went like this, I was white:
e4 d5 exd5 Nf6 (a bit wacky) c4 c6 dxc6 Bg4?? I said cxb7 and it said « that’s not a valid move because your king is in check »
It seems it can’t properly play
I tried asking it to give me an article on the lucena position with a FEN of an example and the position had no kings, a pawn of each color and two rooks
Was just completely faking it until making it
It’s a cool software but it can’t comprehend chess yet
e4 d5 exd5 Nf6 (a bit wacky) c4 c6 dxc6 Bg4?? I said cxb7 and it said « that’s not a valid move because your king is in check »
It seems it can’t properly play
I tried asking it to give me an article on the lucena position with a FEN of an example and the position had no kings, a pawn of each color and two rooks
Was just completely faking it until making it
It’s a cool software but it can’t comprehend chess yet
chat gpt is awesome my computer science teacher at school told my class about it
It may be that language is an empty shell by itself.