Gian-Carlo Pascutto wrote: ↑Fri Aug 03, 2018 7:07 pm
Now, as for the weights, if they are not copyrightable, then anyone who gets the engine can just give them away, same as with the source code. So it doesn't change much.
Yea, but you can't enforce anyone that uses Lc0 binaries to produce or run new weights to publish them. As a matter of fact anyone can take Lc0 binary, package it with his weights, and sell it as a product without any responsibility towards Lc0 copyright holders, beside giving a link to Lc0 website where source code is.
That said, making the engine closed source is also no guarantee. Rybka -> Ippolit/Robbolito -> ... anyone? Don't think Vasik managed to assert copyright there either, financial gain or not.
Even if Vas had enough prospect of financial gains, in case of winning, to sue, and perpetrators were known, he would never ever win that in court.
There was no court case ever that someone that published to public domain totally different code (from original source code) obtained by allegedly reverse-engineering of a binary was found guilty. Nor there ever be.
Karol Majewski wrote: ↑Fri Aug 03, 2018 6:18 pm
May the 1st 2019 on Chessbase.com:
DeusX 1, the first ever neural network chess engine, is now available on Chessbase. Written by Albert Silver is now the strongest chess engine on the planet (~3700 Elo). Click here to get 20% discount for DeusX 1 and get it for only 2400$.
Only in this case the GPL would come into play. If he wanted to distribute it. Even for free. As long as it's private, Albert can do what he want with it. Only the TCEC rules apply, not the GPL. As far as participation is concerned. At least I think so.
Debugging is twice as hard as writing the code in the first
place. Therefore, if you write the code as cleverly as possible, you
are, by definition, not smart enough to debug it.
-- Brian W. Kernighan
Karol Majewski wrote: ↑Fri Aug 03, 2018 6:18 pm
May the 1st 2019 on Chessbase.com:
DeusX 1, the first ever neural network chess engine, is now available on Chessbase. Written by Albert Silver is now the strongest chess engine on the planet (~3700 Elo). Click here to get 20% discount for DeusX 1 and get it for only 2400$.
Only in this case the GPL would come into play. If he wanted to distribute it. Even for free. As long as it's private, Albert can do what he want with it. Only the TCEC rules apply, not the GPL. As far as participation is concerned. At least I think so.
GPL never comes into play. Simply speaking file weights.txt is not copyrightable and putting copyright claim to it is void. If you think differently please consult a copyright lawyer.
Karol Majewski wrote: ↑Fri Aug 03, 2018 6:18 pm
May the 1st 2019 on Chessbase.com:
DeusX 1, the first ever neural network chess engine, is now available on Chessbase. Written by Albert Silver is now the strongest chess engine on the planet (~3700 Elo). Click here to get 20% discount for DeusX 1 and get it for only 2400$.
Only in this case the GPL would come into play. If he wanted to distribute it. Even for free. As long as it's private, Albert can do what he want with it. Only the TCEC rules apply, not the GPL. As far as participation is concerned. At least I think so.
GPL never comes into play. Simply speaking file weights.txt is not copyrightable and putting copyright claim to it is void. If you think differently please consult a copyright lawyer.
Even without any GPL, the weights are Albert's creation in a sense. If not everybody would have gotten the same weights, they are unique and define his 'program'. Albert could sell them probably, separately, even if he can not sell the rest that is GPL protected. Maybe copyright would not apply because they are not words. But I'm sure they will find a way to legally protect a recipe for the structure of a neural net somehow. I'm no lawyer of course. I don't think there are many here.
Debugging is twice as hard as writing the code in the first
place. Therefore, if you write the code as cleverly as possible, you
are, by definition, not smart enough to debug it.
-- Brian W. Kernighan
Eelco de Groot wrote: ↑Fri Aug 03, 2018 7:52 pm
Even without any GPL, the weights are Albert's creation in a sense. If not everybody would have gotten the same weights, they are unique and define his 'program'. Albert could sell them probably, separately, even if he can not sell the rest that is GPL protected. Maybe copyright would not apply because they are not words. But I'm sure they will find a way to legally protect a recipe for the structure of a neural net somehow. I'm no lawyer of course. I don't think there are many here.
They might find it in US one day (in rest of the world probably never). However, today you can copyright a program that produces the weights or you can patent an algorithm/method that produces the weights, but NN weights themselves can not be protected by copyright nor by any IP law.
The only way Albert can protect them (if he wants to sell them as a bundle with Lc0 binary) is encryption .
Rebel wrote: ↑Fri Aug 03, 2018 6:28 pmIt's naive to think you are safe in court. Judges often reason different than us. A judge might very well reason that the weight file is so crucial to the project that it is in the spirit of GPL (sharing) it should be shared.
Sorry but you have no clue about what you are talking about.
Said by you who claimed (still claims?) the zero approach doesn't work
90% of coding is debugging, the other 10% is writing bugs.
Milos wrote: ↑Fri Aug 03, 2018 6:09 pm
Numbers are not copyrightable, end of story, no ifs or buts. As for the other stuff, copyright is only valid if it is enforceable. And it is enforceable only if you can prove that perpetrator made financial gain by violating it. Other than that copyright is nothing but a toilet paper and talking about spirit and morality just makes ppl laugh and you naive as a newborn.
"Numbers" may be copyrightable if they represent data. You can't say executable code or MP3-encoded music are just "numbers" and are therefore not copyrightable - good luck with that defense in court. If NN weights were simply "numbers", there would be no way to protect NN implementations as intellectual property, and hence no way to make them commercially viable. That's why I have zero doubt courts will treat NN weights as IP.
As for your claim that copyright is "enforceable only if you can prove that perpetrator made financial gain by violating it", that's not true. (And would make no sense either, if one spends 15 seconds or more thinking about it.) Normally I would ignore this sort of statements, but since you tell other people they don't understand copyright law: no, it's you who doesn't understand it.
Branko Radovanovic wrote: ↑Fri Aug 03, 2018 10:48 pm
"Numbers" may be copyrightable if they represent data. You can't say executable code or MP3-encoded music are just "numbers" and are therefore not copyrightable - good luck with that defense in court. If NN weights were simply "numbers", there would be no way to protect NN implementations as intellectual property, and hence no way to make them commercially viable. That's why I have zero doubt courts will treat NN weights as IP.
Seems you also don't understand much about this particular topic. MP3 music is music, not code. If you code random noise in MP3 ofc that it doesn't mean anything. Content is important. NN weights are just that, weights, some numbers that are attached to synapses of particular NN architecture. There is no further content, they have no other meaning beside being weights. NN architecture is not patentable for various reasons, implementation could be, if it is in the form of training algorithms or methods, I wouldn't go into details coz it doesn't make much sense anyway. First you need to understand the basics.
As for your claim that copyright is "enforceable only if you can prove that perpetrator made financial gain by violating it", that's not true. (And would make no sense either, if one spends 15 seconds or more thinking about it.) Normally I would ignore this sort of statements, but since you tell other people they don't understand copyright law: no, it's you who doesn't understand it.
Yea sure, just because you believe so. Please state me one court case where GPL copyright was enforced in case of violation where individual perpetrator made no financial gain. Just one is enough.
Both Lc0 on 2x1080 Ti already losing points in this division, with opposing engines having 400-500 CCRL Elo points less than top dogs. Without mentioning that there seem to be some issues with running on one core for a couple of engines, instead of 43 cores. It seems from this small sample that Lc0 in these conditions are not yet prepared to have a fight with top dogs. Networks are improving, last time I checked, the newest testnets are by now some 60-70 Elo points stronger than the nets in the 10160-10180 ID range against AB engines. Albert Silver might come with a better net too. But I doubt both Lc0 in TCEC conditions will stand a chance against top 3 anytime soon.
.
to think you are safe in court. Judges often reason different than us. A judge might very well reason that the weight file is so crucial to the project that it is in the spirit of GPL (sharing) it should be shared.
