Wow, that's crazy! It is a wild concept that certain numbers could be deemed "illegal" https://en.wikipedia.org/wiki/Illegal_number

Any piece of data can be represented as a single number. In the limit your entire computer is a single number (all those 0s and 1s can either be interpreted as separate numbers or as one gigantic number).

If any data is ever illegal then by association certain numbers are also illegal.

EDIT: The Wikipedia article points this out.

It isn't the number that is illegal, it's the number plus information on how to use it.

child porn images are numbers

You're not wrong, but this is a slippery slope to go down.

8.3.4. "How will privacy and anonymity be attacked?"

[...]

like so many other "computer hacker" items, as a tool for the "Four Horsemen": drug-dealers, money-launderers, terrorists, and pedophiles.

https://en.wikipedia.org/wiki/Four_Horsemen_of_the_Infocalyp...

Not really. We've decided as a society that information can be illegal. Whether that information is encrypted, obfuscated, or whatever, it is still illegal. Everything is a number if you abstract it enough and any number can be transformed into any other number easily. As long as it can be shown that the numbers you're publishing are intentionally being published for the purpose of providing illegal information, you're breaking the law.

The legal system is not like a computer system and tech people who think they can fool it with simple tricks like it was a computer system are in for an unpleasant surprise.

Just wait till GAN neural networks are more wide spread and refined. Is someone liable for copyright violation if their GAN neural network was trained on copyrighted images? The a refined version of the resulting GAN model can be used in reverse to generate false convincing variations of the original copyrighted media. Is anyone liable?

If it goes the same way as the rest of the internet, we’ll be in a situation where they’re illegal, but large “too big to fail” companies will be exempt from the law.

(For examples of what I mean, see any copyright dispute involving YouTube, Dropbox, etc. or the repercussions of illegal content on pornhub vs. Facebook.)

Should not be different than a human artist producing something after looking at other media. The purpose of copyright is supposed to be to allow people to profit from their works by prohibiting their distribution by others. It seems to me the question of whether one work is a copy of another would be made on the basis of the works themselves, and should be independent of who the artists were. There is also a right to parody, and no inherent reason parody can't be produced in an automated fashion.

In the eyes of the law it might not be different, but in the practical upshot it is. Running the program may well produce a copyright-violating image, and you'd never know. An artist would have a much harder time doing that without knowing he was doing it. So then somebody sues you for using their copyrighted work and shows you their original, and you're left scratching your head, pointing at the computer.

If you put copyrighted works into the NN training data then only copyrighted works will come out. The NN itself would be a derivative work and a copyright violation as well without the appropriate licenses.

If you don't put copyrighted works into the NN but then feed a copyrighted work to it as an input the result is still copyrighted.

I really don't find this that complicated. NN's having a license that is a composite of the software license and the licenses of the training data feels super natural.

See: What color are your bits? https://ansuz.sooke.bc.ca/entry/23

> The NN itself would be a derivative work and a copyright violation as well without the appropriate licenses.

This seems reasonable at first, but not the more I think about it. Copyright is not like an AGPL software license that claims/poisons everything it touches.

The NN is not an artistic work and probably should not be considered a derivative work any more than your brain or memory of a copyrighted work should be considered a derivative work. Likewise, the outputted work of a NN is not automatically a derivative work any more than the work of an artist that has been inspired by other artists' copyrighted content is. That's just not how copyright works.

> Copyright is not like an AGPL software license that claims/poisons everything it touches.

Obviously, it is, since the only mechanism copyleft licenses have to claim/poison things the licensed material “touches” is the copyright doctrine of derivative works, and the licensable exclusive rights the original work’s copyright owner has to create derivative works.

I don't think that's quite right. They can prohibit your use of the program unless you agree to and follow the rules they establish. They are able to do that prohibition using copyright, but isn't it proof that copyright doesn't necessarily do that the fact that there are other very different licenses?

Copyright, by default, is not at all permissive. Every license (including viral ones) are more permissive than the legal default standing, not less.

This is why one company I worked for refused to let us use Webpack 1.x- a dependency of a dependency was an ancient npm module from the time that adding licenses to modules wasn't very common. It wasnt until the 2.x line came out that they finally updated to a version that updated to a version that didn't have that problem.

I see.

Yet my comparison of the human brain still stands, I believe.

Just because you learned about something in art class from looking at a copyrighted work doesn’t make everything you produce after that a derivative work.

I hope it doesn't turn out like patents, where everyone is infringing and there's a huge pressure to own "defensive" bodies of works.

In this context I can see that being another pressure which assures that large models are only able to be built by FAANG style companies.

Why not just add 1 to the number (and mention it in the text)? When this fails, add 2, and so on.

Laws are not bytecode sequences, and courts are not dumb VMs, no matter how much programmers want them to be.

I would assume most judges would view this akin to ignoring the DMCA request. It would be like altering a single bit in an MP3 and saying that it's no longer the exact original, even though you put out instructions telling the downloader to change the bit back.