Stockfish vs ChessBase hearing a “meh” affair

[smatovic]

can neural network weights be copyrighted? Licensed?

I am sure every pro NN creater/trainer would disagree with the statement that no human hand/human act of creativity was involved to create the neural network. Only cos you can afterwards automatize your work via a script it does not mean that it was no creative human act in the first place.

I see here some analogy in digital music studio software.

--
Srdja

[dkappe]

can neural network weights be copyrighted? Licensed?

If they are generated using an automated process, I don't think that copyright can apply because that doesn't pass the required creativity threshold to be declared as works, given that it's the machine that generates the data, not a human author. That holds even if using Stockfish in self training, just like Adobe doesn't gain copyright to images made with Photoshop.

If on the other hand the training material is hand selected individually using some human judgement (e.g. for style), that can be different, but again, this would not apply to just exporting games from a huge database using some filters. Just like you cannot copyright some mere list of data, but you can copyright an individually chosen and commented list.

I try to keep up with the thinking on this subject, but the amount of publications on this topic has exploded in the last 18 months. Unfortunately there aren’t any court cases or legislation that would settle the issue.

[smatovic]

[...]
2) The FF2 NN data file is anyway free of copyright, so it can be freely distributed whether CB likes it or not.

Strong statement, AFAIK you are the expert on this, how do you back it up?

--
Srdja

[syzygy]

If you went and made a deal with chessbase to sell Ethereal V11.00 (some old engine that isn't even the best free one) with some catchy engine rename I guarantee it'd sell more than your NNUE, because those guys would publicize it and make people know about it, and convince them to buy it. They know what they're doing and that's why the Stockfish team are going to spend years fighting a lost cause.

The SF team does not (or at least should not) mind that CB or anyone else makes money from selling SF itself or some modification of SF. This is simply what the GPL allows and has always intended to allow.

What the SF teams fights is the alleged violation of the GPL by not making the FF2 net available for free. The SF team argues that that the FF2 net is an inseparable part of FF2 and that CB should therefore not only release the FF2 source code under the GPL (which they do) but also the NN data.

I have two legal views on the whole issue:

1) There is no GPL violation (now that the NN net is no longer part of the FF2 executable). The NN is not an inseparable part of FF2 but merely input data. The fact that you need an NN file to be able to run FF2 is irrelevant. id Software has released the source code of quake under the GPL without releasing the data files that you need to run the quake executable compiled from the source code. It is perfectly legal to release a modified version of SF under the GPL that is entirely disfunctional: buggy, incomplete, or just missing input data files.

2) The FF2 NN data file is anyway free of copyright, so it can be freely distributed whether CB likes it or not.

Are you sure that is what the SF team is fighting for? For the net to be public?

That is not exactly what I wrote. They fight the alleged violation of the GPL. They want to show in court that the GPL was violated (and then request the court to order certain measures based on that violation).

I assume that the alleged violation consists in the NN not being offered for free download together with the modifications to the SF source code. As far as I understand, all other things relating to FF2 have been corrected and are totally minor. Starting a lawsuit just because some formality was not originally complied with is abusive and against the spirit of the GPL.

I'm fairly certain that there issue is that they revoked the GPLV3, and yet ChessBase still distributes.

If FF2 is still being distributed in violation of the GPL, they don't need to revoke the GPL to stop that.

In addition to preventing the further distribution of FF2, the SF team might also want to have a declaration that CB cannot invoke the GPL in the future for SF-derived products. But that would still be based on the alleged FF2-based GPL violation (and probably also the Houdini-based violation, but that argument seems weak as CB had no reason to suspect there was a problem).

[syzygy]

My understanding is that Houdart is the one violating the GPLv3, not CB, but I’ll leave that to someone with actual knowledge of the law.

Technically CB would have violated the SF copyright by selling an SF-derived Houdini, as would anyone who bought a copy from CB and reselled it.

CB would not have violated the GPLv3 by selling an SF-derived Houdini, as they did not receive the program under the GPLv3 license, and they also never pretended to distribute it under the GPLv3.

So it seems SF-derived Houdini cannot count towards a "repeated" GPL violation.

[syzygy]

2) The FF2 NN data file is anyway free of copyright, so it can be freely distributed whether CB likes it or not.

And here the only interesting part of the case, which wasn’t actually addressed in the hearing: can neural network weights be copyrighted? Licensed? This has implications far beyond chess.

If neither side raised this issue, I suppose the court will just assume that the FF2 NN is copyrighted.

To be clear, the FF2 on github is distributed with a NN, so it is functional.

So that makes it easier to decide 1) in favour of CB.

[Rebel]

If you went and made a deal with chessbase to sell Ethereal V11.00 (some old engine that isn't even the best free one) with some catchy engine rename I guarantee it'd sell more than your NNUE, because those guys would publicize it and make people know about it, and convince them to buy it. They know what they're doing and that's why the Stockfish team are going to spend years fighting a lost cause.

The SF team does not (or at least should not) mind that CB or anyone else makes money from selling SF itself or some modification of SF. This is simply what the GPL allows and has always intended to allow.

What the SF teams fights is the alleged violation of the GPL by not making the FF2 net available for free. The SF team argues that that the FF2 net is an inseparable part of FF2 and that CB should therefore not only release the FF2 source code under the GPL (which they do) but also the NN data.

I have two legal views on the whole issue:

1) There is no GPL violation (now that the NN net is no longer part of the FF2 executable). The NN is not an inseparable part of FF2 but merely input data. The fact that you need an NN file to be able to run FF2 is irrelevant. id Software has released the source code of quake under the GPL without releasing the data files that you need to run the quake executable compiled from the source code. It is perfectly legal to release a modified version of SF under the GPL that is entirely disfunctional: buggy, incomplete, or just missing input data files.

2) The FF2 NN data file is anyway free of copyright, so it can be freely distributed whether CB likes it or not.

Are you sure that is what the SF team is fighting for? For the net to be public?

I'm fairly certain that there issue is that they revoked the GPLV3, and yet ChessBase still distributes.

We don't know what's in the subpoena.

[syzygy]

[...]
2) The FF2 NN data file is anyway free of copyright, so it can be freely distributed whether CB likes it or not.

Strong statement, AFAIK you are the expert on this, how do you back it up?

I have two mostly independent arguments.

The first argument is that copyright protects expression, and the FF2 NN does not contain expression. Expression relates to something a human can perceive. JPG data is copyrighted because you can look at the image. Text is copyrighted because you can read it. Source code is copyrighted because (the law says so) and you can read the code. (Object code is copyrighted because the law says it inherits the copyright on the source code.) An NN is just numbers. You can read the numbers or look at them, but it won't tell you anything sensible. If you load them into FF2 they will influence the move FF2 calculates for a given position, but I don't believe this amounts to expression.

The second argument is that the expression has to be the result of a modicum of creativity (= non-functional choices), and there is no such creativity in the FF2 NN. The creation of the FF2 NN involved many choices relating to learning parameters and the selection or generation of training data, but all of these choices were made with the aim of creating a strong chess playing entity and were therefore functional.

The two arguments are not fully independent: the expression has to express the creativity. But in my view an FF2 NN fails both tests: no expression, no creativity, let alone an expression of creativity.

If one considers, which I do not, that the FF2 NN is a "computer program" in the sense of the EU directive 2009/24/EC (regulating copyright on computer programs), then the first argument is overcome to some extent, but the second argument not.

[syzygy]

can neural network weights be copyrighted? Licensed?

I am sure every pro NN creater/trainer would disagree with the statement that no human hand/human act of creativity was involved to create the neural network. Only cos you can afterwards automatize your work via a script it does not mean that it was no creative human act in the first place.

I see here some analogy in digital music studio software.

If you create music, then you make lots of creative choices.

But if you create sound that somehow cures cancer or makes a cow play better chess, then the choices you make to achieve that purpose are functional and not "creative". You are engineering, not "creating" in the sense of copyright.

There is no copyright on endgame tablebases

[Sopel]

Again, 1. whether the NN file is copyrightable or not is irrelevant because in FF2 it was part of the binary therefore required as a corresponding source. 2. Curing previous violations doesn't prevent license termination under section 8. 3. The way houdini was distributed to chessbase doesn't matter, the judge said it's chessbase's diligence to verify legality of their distribution.