Calling All Armchair GPL Lawyers

[unserializable]

2) This is also pretty clear, how the parts are packaged doesn't matter.: "You must license the entire work, as a whole, under this License to anyone who comes into possession of a copy. This License will therefore apply, along with any applicable Section 7 additional terms, to the whole of the work, and all its parts, regardless of how they are packaged."

Hi Fulvio, you are overlooking aggregates that directly follow the section that you quoted -- aggregates need not consist of GPLv3 items only:

...
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
...

[Fulvio]

A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work

Hi,
obviously the nets are not "separate and independent works".
That section apply for example to the Fritz GUI: even if it is distributed alongside a GPL3 program it can have a different license.

[dkappe]

W.r.t. NN(UE) nets I wonder if these are just data or could be compared to state-machines for instance .
So maybe the distinction between NN(UE) as pure data or instruction is not so clear.
So maybe these could still be covered by GPL??
Best

This is a frequent question in other domains where pretrained models are used: can the pretrained model be copyrighted and/or licensed? I’d say the question isn’t fully settled, but the model protection train is already running, with piracy resistant model watermarking being an active area of research and commerce.

[hgm]

[Moderation] As I interpret this as a general question on NN-using engines, to which the FF2 case just serves as a real-life example, I see no need to move it to the Engine Origins section. Other moderators are welcome to do so if they object to the thread.

My own 2 cents:

I am not really acquainted with the FF2 case, but from what I read here I gather that FF2 is a Stockfish version, perhaps slightly modified, using a 'private' NNUE. Where the sources of this modified version are available on-line. I agree with unserializable that the NNUE is data, and that the GPL does not extend to it. An interpreter of a programming language is a good analogy, even if we want to maintain that a NN is a program. Stockfish is just a 'NNUE interpreter'.

It also seems to me that there is no such thing as the 'source code' of a NN, so that the GPL cannot be applied to it (e.g. if I would want to extend an existing net, e.g. embed it into a larger one).

If there is a violation here, it would be the omission of the GPL licence text with the product, but I think this is a very minor one. When the source code is not included in the distribution, but available on line, no one that receives the package can actually make derivatives, so the conditions under which they could are completely irrelevant to them.

[smatovic]

...
My own 2 cents:

I am not really acquainted with the FF2 case, but from what I read here I gather that FF2 is a Stockfish version, perhaps slightly modified, using a 'private' NNUE. Where the sources of this modified version are available on-line. I agree with unserializable that the NNUE is data, and that the GPL does not extend to it. An interpreter of a programming language is a good analogy, even if we want to maintain that a NN is a program. Stockfish is just a 'NNUE interpreter'.

I doubt that is is that easy, a audio player is also just an interpreter, and music data, but in this case the music is copyright protected and you won't ship a open-source GPL audio player with non-GPL copyright protected music in its binary, would you? If NNs are in general copyright protectable is questionable on mutliple level, but I guess in this case we have to assume that Chessbase's NNs are copyright protected unless the jurisdiction says otherwise...

It also seems to me that there is no such thing as the 'source code' of a NN, so that the GPL cannot be applied to it (e.g. if I would want to extend an existing net, e.g. embed it into a larger one).

If there is a violation here, it would be the omission of the GPL licence text with the product, but I think this is a very minor one. When the source code is not included in the distribution, but available on line, no one that receives the package can actually make derivatives, so the conditions under which they could are completely irrelevant to them.

[smatovic]

...
My own 2 cents:

I am not really acquainted with the FF2 case, but from what I read here I gather that FF2 is a Stockfish version, perhaps slightly modified, using a 'private' NNUE. Where the sources of this modified version are available on-line. I agree with unserializable that the NNUE is data, and that the GPL does not extend to it. An interpreter of a programming language is a good analogy, even if we want to maintain that a NN is a program. Stockfish is just a 'NNUE interpreter'.

I doubt that is is that easy, a audio player is also just an interpreter, and music data, but in this case the music is copyright protected and you won't ship a open-source GPL audio player with non-GPL copyright protected music in its binary, would you? If NNs are in general copyright protectable is questionable on mutliple level, but I guess in this case we have to assume that Chessbase's NNs are copyright protected unless the jurisdiction says otherwise...

It also seems to me that there is no such thing as the 'source code' of a NN, so that the GPL cannot be applied to it (e.g. if I would want to extend an existing net, e.g. embed it into a larger one).

If there is a violation here, it would be the omission of the GPL licence text with the product, but I think this is a very minor one. When the source code is not included in the distribution, but available on line, no one that receives the package can actually make derivatives, so the conditions under which they could are completely irrelevant to them.

Followup, rereading the posts I realize you mean distributing the net as data as separate file and not in the binary, so feel free to ignore my answer.

--
Srdja

[Fulvio]

I agree with unserializable that the NNUE is data

Doesn't matter: "This License will therefore apply, along with any applicable section 7 additional terms, to the whole of the work, and all its parts".
It is distributed as a part of a GPL program, therefore it is covered by GPL.
I already posted the link to how BusyBox litigated and get compensated for those type of violations.
And now the companies are providing all the code, including the .bin files (for example https://www.tp-link.com/us/support/gpl-code/ )

Chessbase could try to distribute only the "improved" NNUE under a different license, but is not clear if that would be enforceable.
However, their main selling point is ease of use. They offer a complete package with a GUI, a strong engine and a small database all pre-configured and easy to install. It is certainly an added value for which many customers willingly pay.
What I find disconcerting is the abrasive personality of the people involved in those interviews and articles; the other time it generated resentment in the lc0 community and is now on the same path with the Stockfish' one.
It also seems bad marketing, they could simply advertise it as "an improved Stockfish, highly tested and with an optimized evaluation system tuned analyzing million of games..."

[matejst]

Namely, I've been working on some strong NNUE nets for a few months with all original efforts, and many have urged me to sell a pro version of future Ethereals with the NNUE nets being paid for. I would want to do it in a way that most people agree is up to code, even if there is no chance of someone coming after me for the code I actually own.

Aside from these conversations, I support you in the idea of making a commercial version of Ethereal. I am sure that you will be able to improve it greatly to the point it would be close (and probably, at times, better) to SF, and, with some luck, the association with a company providing a good interface, you could earn some well deserved money.

[Dann Corbit]

The github project has both the code and the data.
I don't understand the fuss over code verses data here.

I also think that the data is probably not copyrightable.
But I am not absolutely positive about that,
It is conceivable that enough creative process went into the production of the data that it is more like a story than numbers.

[pohl4711]

4) Yes, and they also need to clearly state that it is a stockfish derivate

A.Silver told me, that the FF2 Binary is Stockfish 210111. Only changes are the modifications, which are needed to use the bigger FF2-nnue net. Since the update1 the Contempt was set to 24, again, which is default of all Stockfishes.

All what makes FF2 unique is the bigger nnue net, built out of Fat Fritz 1 evals (=Lc0 with FF1-net (FF1-net was built out of human-games and selfplay))
Nothing more, nothing less.