A few tips on AI matter. Final thoughts on the Topic

  1. Neural Network based AI is a MiniMax (wiki) search already. By it's nature of pytorch. An Evaluation function iteration in depth. With backpropagation (wiki) , it corrects itself depending on training results. The task is to train a model (a .keras file in tensorflow model, a .pth as a state, .safetensors - a kind of binary tensor model to share, bunch of other types), a kind of a binary tree, that "predicts" best chess moves by given chess position as a parameter. Kind of opening databases combined with statistics. "Likewise playing" in gigabytes. Yes, yes, it can be strong. Lc0 is strong when trained and updates of weights already downloaded on an appropriate hardware. ChatGPT says, Lc0 needs half a Gb. Gemini says, the Big Transformer 4 model likely occupies between 28 GB and 40 GB of storage space. Interestingly, how many times less then tetris being developed today. I mean, lots of space on harddrive for 10 NNs.
  2. Another approach. Very theoretically speaking, in case of unlimited resources we could build a simple NN with evaluate checkmate only. Then perform megatraining. After it would be just the right chess engine that guides us throughout the game till the checkmate. Unfortunately chess is not a TicTacToe 9-squares simple problem, well described with samples on internet. There should be a little evaluation logic built-in for better selection of what to compute and what to ignore. We can't just generate gazillions of dumb chess games and simply count winnings of them. The default point of view is that we simply perform a full-scan and maybe many times on an infinitely powerful new hardware. Almost this. GenZ.
  3. Nobody wants to provide free computational resources to calculate evaluations for all possible chess positions. Should be a selection anyway. And what to play in the endgame?! Stable diffusion? xexe. There is FishNet project of distributed computing resources, obviously people have lots to give for chess analysis project. Data is king there, including expensive hardware too.
  4. Nalimov endgame tablebases are calculated already. Syzygy and Gaviota tablebases available to lookup a 7-pieces position in terrabytes large files. My guess, it is done in assembly language made chess code for fastest move generation, have seen such. But may be C too. So, just a lookup is most needed in such cases when almost nothing left on the chess board. In fact, there is chess theory known for centuries already and loads of literature too. For chess openings, the polyglot .bin files of chess evaluations for given key of chess position available on the web. Chess openings on websites available, as oldschool chessgames.com, don't think they are wrong. Anyway, large collections of chess games available at Lichess.org It is the best site today.
  5. There is PyPI Stockfish for python, already usable tool for strong Evaluation function, for example. Have seen projects of tensorflow usage on position evaluation. This can be done variously. Maybe just an evaluation function and years of iterations can give results. Who knows. Maybe.
  6. The only chess calculation thing that can be easily splitted in parallel GPUs is a self written evaluation function of a given chess position, obtained somewhere before. GPU have no recursions, it should do fast and simple math and independently of data. Let's say, simply compare material, count centered pieces, etc. Checkmate verification would be too slow. So, maybe it is better not to try split movegen, or smth. Let GPUs do their job in pytorch / tensorflow part only, calculating matrices of nodes with weights. Simply speaking, write everything in tensors (matrices), and this will be processed on GPUs in parallel, or on TPUs. At least, will try. GPU is designed for simple math on numbers (pixels,vertices), not decision (ifs) on decision on (recursive) .... of numbers - the smallest part of calculations. Kinda explanation. Therefore CPU does much of work.
  7. Evaluation function of a new NN based AI can be correlated to statistics of openings played by chess masters and results of these games. So, no need to do Stockfish evaluation much there. Lc0 goes right or left only, counts winnings, no need for counting evaluation decimals. By the way, model outputs maybe can be transferred to a newer better models.
  8. Ignore chess redundancies related to human playing. Ignore the chess clock, 50-the moves rule (real masters do not play such a dumbchess), also ignore "ups! omg" touched pieces to move them after to somewhere. The chess960 case also. I suggest to put a queen always when promoting pawn to piece, ignore promotion of rook, bishop, knight. Move number also is very irrelevant - no information there at all to keep it all time actual. Simply do not think about it. The main task is to find the best move (or sequence, variants) for given chess position and NN can assist here for sure. Also, I think, the endgame can be a different model because of approach. Maybe split NN into three independent models: for opening (ready databases), middlegame (lots of pieces and should calculate much, the battle), endgame (do lookups, if less than 7 pieces). Absolutely. Make it today and it will be superstrong tomorrow. Just ideas, why not. Nvidia cares tomorrow, stock goes up today.
  9. Also, the old binary file level was designed for PCs running on 8Mb of RAM, even less. Today it is possible to operate in very large memory space. I mean, no need to encode decode anything. Simply why. Also the web is very fast now. Keep the data plain and simple, visible in text files. Dump (flush) large memory to disk, no need to itearate the old filesystem for small chunk to save anything at all. Python does lots of hashing inside itself. No gain counting bytes. Some games with extensive graphics require more than 100Gb for installation nowadays. Making short bit-squeeze records may be reasonable for extremely large endgame tables maybe, but AI should play a good game till 7 pieces left anyway. Can ignore it. If NN gets best results at the beginning and middlegame of chess playing then may say it is strongest ever. After the middlegame, when half of pieces are off the board, things get kind of less CPU requiring, always can compensate with an existing chess engine that plays at decent level anyway. Playing till checkmate is boring and just technically possible for real chess players. They are interested in gaining advantage that is obvious and clearly visible to everybody just taking a look at position on the chess board. In most cases this chess game is over then, chess clock being stopped, chess engine turned off, nothing interesting more. By the way, chess engines do not search for checkmate as most important thing, sometimes a checkmate case just happens as by-product. I mean, it is too much legacy everywhere on the web. Yeah, same old stuff. Currently Stockfish has NNUE (wiki) implementations as addons to the existing chess engine. Wrong way. Upside down. NNs should be primary today with an engine attached. Has anyone insufficient diskspace today in 2025? By the way, Intel stock goes down, sorry. Let's build a NN database.
  10. You have to get all ready data from Lichess.org, these all are mostly donated by fish-net users. Chess lovers spent money for lichess.org reasons, as I see. NN training will require lots of data. Otherwise will get a noob-chess NN. To be honest, ChessBase sells their databases products almost at price of a new PC. Lichess, at least, gives large files of games for free. So, that's it - the real value of TensorFlows for chess.
  11. Speaking of pythons. For Windows users there are DLL-libraries to do the same on local PC, or even process Stockfish position evaluation right in python. Go to git.
  12. Good luck.