Surprisingly, I find myself slightly skeptical of Stockfish's claims, which is... a disquieting feeling. Could someone help me out by finding the flaw in my logic?
2. But the weights are already being released with the program they're selling. Otherwise, how would the program run? Why would they refuse, when they already distribute the weights with every copy of the program? I thought maybe Fat Fritz 2 was a cloud-hosted solution (weights aren't provided locally), but https://en.chessbase.com/post/running-leela-and-fat-fritz-on... seems to indicate it can run on a laptop with a GPU.
I suspect Stockfish may be trying to argue that whatever code was used to generate the weights is also covered under GPL, and that ChessBase is in violation because they refuse to release this code. I hope they don't try that argument, because it's certainly not true; you're free to train a chess engine however you want, insert the weights into a GPL'd program, and then are under no obligation to release the code to train the weights.
3. Setting aside the question of copyright entirely, it's unclear that Fat Fritz 2 is actually worse. Stockfish references https://www.sp-cc.de/ as an independent rating system, but if you look at the detail view (https://www.sp-cc.de/files/programs.dat) Fat Fritz 2 only seems to be compared to Stockfish 13 (210218) and not the latest versions of Stockfish.
I'm on Stockfish's side here, because I suspect ChessBase simply duplicated StockFish's training methodology and tried to pass it off as their own. (The engine's performance seems almost identical.) But I hope Stockfish has an airtight case for court. It'd be a bad idea to hinge the GPL violation on "they refused to release the neural net weight training code."
Typically you have stockfish source code, and human readable model file with weights. Build system and compiler bakes them together to produce binary. Inside the binary, the NNUE weights are no more readable than the source code, which is to say none at all (unless you have reverse engineering expertise). If providing the final binary counted, then there would be no such thing as GPL violation?
>> 2. But the weights are already being released with the program they're selling. Otherwise, how would the program run?
In the same way that, e.g. Windows can run from binaries, without distributing the source code? That is to say, the weights are in a binary format that is not amenable to re-use without decompilation and reverse-engineering.
The reason why they're going after them is because of their refusal to publish the weights of their network, yes, but that's not the basis for this lawsuit. After violating the GPL license, StockFish has terminated ChessBase's license. They are no longer allowed to run or distribute StockFish's code.
Given that the chess engine in question is just a tweaked version of Stockfish, this will have a severe impact on their ability to sell their product.
I'm not sure how this would work with derivative code, though. I wonder if they could take a fork of StockFish, which is not StockFish and therefore subject to a different license agreement, and go on their merry way. I don't think that tactic would work (the original author can still claim copyright over the code from the forked repository, I believe) but it might be a possible defence? I think? Not a great one, though.
I don't see any other way ChessBase can get out of this, unless the court decides that the termination clause is unlawful for some reason.
They mean that the weights are not open sourced. If the weights are distributed separately and not considered part of the "Corresponding Source", that may not be a GPL violation, and the blog post does not imply that it is. (They only say that they condemn the approach.)
> 2. But the weights are already being released with the program they're selling. Otherwise, how would the program run? Why would they refuse, when they already distribute the weights with every copy of the program? I thought maybe Fat Fritz 2 was a cloud-hosted solution (weights aren't provided locally), but https://en.chessbase.com/post/running-leela-and-fat-fritz-on... seems to indicate it can run on a laptop with a GPU.
The GPL can be interpreted to require that all the input necessary to reproduce and run a functioning binary must be distributed. If the weights and engine are integrated in one binary, and not "all the source code needed to generate, install, and run the object code and to modify the work, including scripts to control those activities" is provided, that could be a GPL violation. This does not apply if the weights are distributed separately under a different license.
> I'm on Stockfish's side here, because I suspect ChessBase simply duplicated StockFish's training methodology and tried to pass it off as their own. (The engine's performance seems almost identical.) But I hope Stockfish has an airtight case for court. It'd be a bad idea to hinge the GPL violation on "they refused to release the neural net weight training code."
They may have terminated the license in response to the Houdini revelations[1] which is another chess engine based on Stockfish that Chessbase has sold as original work, disguising its Stockfish origins in violation of the GPL.
As the GPLv3 says under Section 8 [2] the license can be reinstated if "this is the first time you have received notice of violation of this License (for any work) from that copyright holder" - the first violation would have been Fat Fritz 2.
1. ChessBase was in violation of the GPL, but then cured the violation by releasing code (https://stockfishchess.org/blog/2021/statement-on-fat-fritz-...). But the post claims they refused to release the weights.
2. But the weights are already being released with the program they're selling. Otherwise, how would the program run? Why would they refuse, when they already distribute the weights with every copy of the program? I thought maybe Fat Fritz 2 was a cloud-hosted solution (weights aren't provided locally), but https://en.chessbase.com/post/running-leela-and-fat-fritz-on... seems to indicate it can run on a laptop with a GPU.
I suspect Stockfish may be trying to argue that whatever code was used to generate the weights is also covered under GPL, and that ChessBase is in violation because they refuse to release this code. I hope they don't try that argument, because it's certainly not true; you're free to train a chess engine however you want, insert the weights into a GPL'd program, and then are under no obligation to release the code to train the weights.
3. Setting aside the question of copyright entirely, it's unclear that Fat Fritz 2 is actually worse. Stockfish references https://www.sp-cc.de/ as an independent rating system, but if you look at the detail view (https://www.sp-cc.de/files/programs.dat) Fat Fritz 2 only seems to be compared to Stockfish 13 (210218) and not the latest versions of Stockfish.
I'm on Stockfish's side here, because I suspect ChessBase simply duplicated StockFish's training methodology and tried to pass it off as their own. (The engine's performance seems almost identical.) But I hope Stockfish has an airtight case for court. It'd be a bad idea to hinge the GPL violation on "they refused to release the neural net weight training code."