Stockfish: Our lawsuit against ChessBase

[syzygy]

No, neural nets are not just weights, there is not a "the neural net" or a "the NNUE" that you could take for granted. The designs, the tooling, the driving code and the precise set of numbers that made it to work are all creative work therefore subject to copyrights.

Once it is about "that made it to work", copyright disappears from the picture.

Of course the code implementing a particular network structure will normally be copyrighted, but that is separate from the network (the NNUE file) itself.

And it is implementation specific and not a mathematical theorem or ground truth, there are endless instances of them offering specific characteristics and/or performance.

Still, this is engineering and not copyright.

In this forum, there are people who would like to argue that search function is also "solved", there exists only one right way to do it and therefore it is free for all to copy paste.

What makes eval.c copyrighted is not the specific search algorithm including reductions, extensions and search parameters but the choices that are not implied by the functionality of the code.

Claiming NNUE nets are not copyrightable IMO is similar to claiming Blues music are not copyrightable, ymmv.

Only if you listen to Blues music by reading out the 0s and 1s of the mp3 file.

[syzygy]

Another interesting fact is that in some countries one cannot "give up" copyrights, therefore licenses like CC0 are created to grant as much freedom as it is legally possible. So in principle people would have to hire a lawyer and file a suit to argue and prove that they don't have copyrights over something for whatever reasons, lawyers do indeed know how to create work for themselves.

Not sure what is your point. Are you trying to imply that it is immoral to suggest that something uncopyrightable is not copyrighted?

[syzygy]

Well, IMO something is missing, artists can not protect their work from being sucked into NN molochs, and people investing work/time/energy to create neural networks can not claim Copyright on them? Obviously both cases were not considered in the good old days...

They can if copyrighted pieces of their work are sucked into the NN and potentially output.

I do not know in any detail how AI image generators work, but the claim seems to be that they do not actually copy pieces of the training images.

An NNUE certainly does not copy any actual moves or positions.

[syzygy]

(...) and people investing work/time/energy to create neural networks can not claim Copyright on them?

Should Mersenne primes be copyrighted just because a lot of time and energy goes into finding them?
Should bitcoins be copyrighted?

SCOTUS rejected the "sweat of the brow" doctrine in Feist v Rural. In Europe the law is not different.
Copyright does not exist to promote all investments of work/time/energy.

[dkappe]

Likely there will be legislation covering copyright on neural network weights, distilling of commercial nn’s into new nn’s, and copyright of generative ai’s, their output, and their training data. The current copyright laws are no longer fit for purpose.

[smatovic]

(...) and people investing work/time/energy to create neural networks can not claim Copyright on them?

Should Mersenne primes be copyrighted just because a lot of time and energy goes into finding them?
Should bitcoins be copyrighted?

SCOTUS rejected the "sweat of the brow" doctrine in Feist v Rural. In Europe the law is not different.
Copyright does not exist to promote all investments of work/time/energy.

How would you suggest to protect (Copryight or patent or ...) the original work of A0 or the initial NNUE SF or things like Google BERT or OpenAI GPT-3? Was it an creative act or just engineering or just collecting data? Maybe this is my own bias in perception, writing code and finding solutions to problems is an creative act for me, like painting a picture. As far as I got it, the algorithms themselves in Knuth's TAOCP volumes can not be protected (patent or Copyright), but the specific implementation in his MIX language can be copyrighted, this is good for me as programmer, would be hell if I had to pay for every algorithm I use a fee.

[towforce]

Should Mersenne primes be copyrighted just because a lot of time and energy goes into finding them?

No - mathematical constants cannot be copyrighted.

Should bitcoins be copyrighted?

Who would copyright them? What would be a breach of copyright look like in this case?

An NN can not do anything by itself: it's part a program - and programs/software can be copyrighted.

[syzygy]

(...) and people investing work/time/energy to create neural networks can not claim Copyright on them?

Should Mersenne primes be copyrighted just because a lot of time and energy goes into finding them?
Should bitcoins be copyrighted?

SCOTUS rejected the "sweat of the brow" doctrine in Feist v Rural. In Europe the law is not different.
Copyright does not exist to promote all investments of work/time/energy.

How would you suggest to protect (Copryight or patent or ...) the original work of A0 or the initial NNUE SF or things like Google BERT or OpenAI GPT-3?

The code is protected by copyright. (But an independent re-implementation of exactly the same functionality will not infringe that copyright.)

The network weights can be freely copied, but you will need code that understands it to do anything with it.

Was it an creative act or just engineering or just collecting data? Maybe this is my own bias in perception, writing code and finding solutions to problems is an creative act for me, like painting a picture.

Are you concerned about the programmers or about the artists? (Previously you were talking about the artists, now the programmers.)

Let's consider a program that uses an NN to generate an image in response to a textual prompt.

The following persons might try to claim a copyright:
- the programmer;
- the person training the NN;
- the artists on whose works the NN is trained;
- the operator who enters the textual prompt.

And the claimed copyright might cover the following things:
- the source code;
- the NN weights;
- the generated images.

Only the programmer will have a copyright on the source code he writes (and this extends to the compiled object code).

The NN weights are not copyrighted. Just numbers not encoding any expression.

The generated images presumably do not contain any expression coming from the programmer.
As far as I understand, typical AI image generators produce images that do not contain any expression contained in the training material.
So the programmer, the person training the net, and the artists have no copyright on the generated images.

The operator MIGHT be able to claim a copyright on the generated image, namely if he is able to control the expression in the image (not its abstract content) through the textual prompt. In this case, the operator is an author using the AI as a tool. But the AI image generators I have been reading about do not seem to work as a tool but independently come up with expression of the more abstract concept/idea represented by the textual input, in which case there is no copyright of the operator. (*)

E.g. a photographer will typically have a copyright on a photo he takes because he uses his camera as a tool and is able to control at least some of the expression in the photo. However, if he has the original idea to give his camera to a monkey and let the monkey take a picture, then this original idea does not control the expression in the picture, and he won't have a copyright.


(*)
This is different in the UK. According to s. 9(3) of the UK Copyright Act, "In the case of a literary, dramatic, musical or artistic work which is computer-generated, the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken". According to s. 178, a "computer-generated” work is "generated by computer in circumstances such that there is no human author of the work". It is not entirely clear to me who is the person "by whom the arrangements necessary for the creation of the work are undertaken", but I would guess it is the operator. So if an AI-generated image shows at least a tiny bit of originality, then the operator would seem to have a copyright on it, in the UK. (However, this provision might be in violation of the Berne Convention. I don't know if it has been tested in court.)

[smatovic]

[...]
Are you concerned about the programmers or about the artists? (Previously you were talking about the artists, now the programmers.)
[...]

My open source code on GitHub is fed into AI/NN systems, currently there is no way for me to prevent this, except going closed source, I do not like this as a programmer. I myself will pubiish the neural networks I create/train under an open source license, but I realize that others wish to protect their work.

--
Srdja

[syzygy]

[...]
Are you concerned about the programmers or about the artists? (Previously you were talking about the artists, now the programmers.)
[...]

My open source code on GitHub is fed into AI/NN systems, currently there is no way for me to prevent this, except going closed source, I do not like this as a programmer.

Again, if your code is copied, then your copyright is infringed.

Reading your code and learning from it is not a copyright infringement. If you don't want that people or computers you don't like read and learn from your code, then indeed publishing it on Github is not the best idea. I wanted to say open source, but that is different. Even a closed source license such as Crafty's does not prohibit reading and learning from Crafty's code (and we all know that this was exactly Hyatt's intention).

I myself will pubiish the neural networks I create/train under an open source license, but I realize that others wish to protect their work.

So this is a completely separate issue. It might be helpful to at least use different paragraphs to discuss completely separate topics