Calling All Armchair GPL Lawyers

[Dann Corbit]

You can always train your own net. In that respect, it is no different than SF

[Fulvio]

MP3 players contain a complex decoding algorithm, which can be implemented in different ways in different players. I am not really into MP3, but I know it is not an exact representation of the original sound, like a .WAV file would be.

The encoding part, when the parts to be discarded are chosen, is differente but the decoding part is math.

So the reproduction of the original sound will in general not be perfect, so that there is room for improvement.

Who cares?
Hasn't it already been said a thousand times that copyright covers the song and not the sound?

[hgm]

Who cares? The people that want to listen to the song, of course. Producing High Fidelity sounds is a multi-billion-dollar industry. Typical MP3 does not have the same quality as a Compact Disk. There surely would be market for an MP3 player that would make low-quality MP3 sound as good as CD.

I don't think the decoding part is math. The encoding throws information away. There is no way the decoding can uniquely recover that information.

[mar]

Who cares? The people that want to listen to the song, of course. Producing High Fidelity sounds is a multi-billion-dollar industry. Typical MP3 does not have the same quality as a Compact Disk. There surely would be market for an MP3 player that would make low-quality MP3 sound as good as CD.

I don't think the decoding part is math. The encoding throws information away. There is no way the decoding can uniquely recover that information.

that depends on the bitrate. I claim that at high bitrates you won't be able to tell the difference.
also why bother with mp3 when we have patent-free Vorbis that's vastly superior especially at low bitrates.

plus there's lossless FLAC if you want 100% fidelity. it's funny though that many people listen to 192kHz 24-bit FLAC that's actually been transcoded from low-bitrate mp3 and call themselves audiophiles

it's quite funny how one can make the parallels with computers chess here

also: inverse MDCT isn't math?

[mvanthoor]

So the FF2 thread devolved into 10 different interpretations of how one could provide a binary of a GPL engine with an NNUE network, and still violate (in their view) the right's of the GPL'ed programs authors. In this thread, I ask each individual what they feel is an appropriate way to offer a private Network which is intended to run inside a slightly modified GPL engine. IE, what could Fat Fritz 2 have done such that there would be no doubt in your mind that all license obligations were followed, even if the entire concept was lame.

My view:
1) Fat Fritz does not need to post the code for their modifications online
2) Fat Fritz needs to provide the Network as a second file, read in using the usual UCI commands
3) Fat Fritz needs to provide the source to any customer (someone whom they have shared the binary with) who made a request
4) Fat Fritz needs to include a copy of the GPLv3 alongside their (download?) to inform customer's of their rights.

As far as I know...

- If you use and change GPL'd code, you are obligated to publish those changes
- It is not forbidden to sell GPL'd code.

Therefore, if Fat Fritz is a somewhat modified Stockfish, the code modifications should be published. Chessbase has the right to train their own network, and then sell that network, running it inside of the modified Stockfish engine. Chessbase has been selling their GUI's with Stockfish (and before, Crafty) as a standard-installed engine for some time. Fat Fritz was an Lc0 network, running inside of a somewhat modified Lc0 executable. (It actually states that in the engine name.)

I see no problem with this, but I'd never buy a chess product for the engine; I'd buy it for either the GUI, or the Chessbase program itself. I stopped being interested in Chessbase engines when Frans Morsch quit / left after Fritz 13. After that, "Fritz" just became a trademark for me... and selling GUI's running GPL'd engines with their own custom network, makes that abundantly clear.

[hgm]

also: inverse MDCT isn't math?

Well, basically dividing by zero is not math. If information is thrown away by compression (which it must be), there is no mathematical way to recover it without additional assumptions on the original. But it cannot be excluded that the a NN looking at the spectrogram can recognize things like "hey, this looks like a trumpet, let me add the overtones of a standard trumpet to make it sound better". Of course this might produce strange effect when you actually listen to the sound of en elephant sneezing, but when playing orchestral music I might sound great.

[Ckappe]

I am not a lawyer but I have a clear view of how I think this should be done.

Sell and distribute your network stand-alone as a network and clearly state the prereqs to run it. And for g**'s sake be honest and informative about exactly what you provide and what you have copied. Also, ensure you can prove the network is truly unique and original work by you (copyright and reaches a Threshold of originality) - and can prove this is not just some derivative work based on re-inforced learning from other author's networks).

Any changes to GPL based engines needed to run your network should be packaged and provided separately. These binaries and sources should not be tied to a specific network or even distributed together!!... If these distributions only apply to a proprietary network, you should apply GPL to the complete distribution.

[mar]

Well, basically dividing by zero is not math. If information is thrown away by compression (which it must be), there is no mathematical way to recover it without additional assumptions on the original. But it cannot be excluded that the a NN looking at the spectrogram can recognize things like "hey, this looks like a trumpet, let me add the overtones of a standard trumpet to make it sound better". Of course this might produce strange effect when you actually listen to the sound of en elephant sneezing, but when playing orchestral music I might sound great.

well, your analogy of comparing SF to a mp3 player simply seems off.

there's only one way to decode encoded audio data. so if you take say random values (and divide them by zero), then happy listening but that certainly doesn't qualify as mp3 player to me.

some relatively weak engines gained 500-1k elo by using the SF net IIRC, but still lacking some 700 elo behind StockFish. that's hardly what'd qualify an roundoff error, assuming I even have enough fantasy to imagine how a strong chess playing program relates to audio codec.

[hgm]

there's only one way to decode encoded audio data. so if you take say random values (and divide them by zero), then happy listening but that certainly doesn't qualify as mp3 player to me.

Well, call it what you like. The point is that it plays MP3 files, and makes them sound better than when what you would call an MP3 player.

Any discussion about this completely misses the point I intended to make: that it would be no legal problem to distribute such a GPLed 'MP3 non-player' together with a bunch of low-quality MP3 tracks that are under a completely different licence. (Namely that of the songs they represent.)

[Fulvio]

I don't think the decoding part is math.

Your opinion doesn't match reality. There is a standard, google it.

Also your entire point doesn't make any sense: it is a separate binary file, so it's not part of the software.
I suppose that all the icons or the graphic art of videogames are not copyrighted in your view...