I think ml models should be public unless the data itself isn't.
You can't translate a commerical book without paying the copyright holder. You own the copyright for the translation but everything else still remains that of the original author. Why wouldn't this apply in case of a machine learning model trained on my images, messages, intents, likeness, etc?
Your data? It’s their data now. (And Microsoft’s.)
One aspect of this is that historically statistical models calculated from large volumes of text (which is a notion that predates computers, e.g. frequency dictionaries and the whole [sub]field of quantitative corpus linguistics) have been considered facts about that corpus of text and thus not copyrightable at all or (depending on jurisdiction) entitled to different set of protections/limitations assigned to compilations of facts, which give some rights to the people who compiled the facts but no rights to the source of these facts (since facts as such can't be copyrighted).
This also applies to many forms of analysis of audiovisual data, where the copyrights of the source works do not transfer to the results of the statistical or qualitative analysis and can't limit their creation, distribution or sale.
The appropriate analogy to a commercial book or movie is not a translation, but some analysis of it - e.g. a thorough literary review and critique of some book or movie is a separate work with its own copyright, and the original author has no claim on it despite the fact that is (obviously) based on the contents of the work and describes it in great detail. Including verbatim fragments of the work is limited (fair use allows some inclusions but not all), but all the other details are not.
Also, the whole notion of copyrightability of ML model weight files is interesting and IMHO not settled. You could argue that there is some creative expression in forming the model (which would support it being copyrightable) or you could argue that it's a mechanistic result of the application of some algorithm and settings (which have the creative part, and are copyrightable on their own), and so the output can't be copyrightable, no matter how much work (human or machine), time and cost it took - at least in USA copyright law doctrine (e.g. Feist Publications v. Rural Telephone Service) is that mere "sweat of the brow" (no matter how much) does not entitle a work to copyright protection; it requires application of human creativity to create an original work, and automated processes can't satisfy that requirement.
So from the perspective of GPT-3 authors it might be prudent to distribute the model only to people who agree to specific, explicit contractual/licensing restrictions on redistribution and usage of the model instead of simply relying on the default application of copyright laws.
And crucially, if some output is not copyrightable, it can't be considered a derived work to which the authors of a source work could have some claim.
1. Artistic/creative expression in my comments.
2. Direct commerical interest. If you are not benefiting from my comment directly, it can be considered fair use and hard to prove otherwise.
In gpt3's case, it outputs artistic expression in the form of poetry/stories and direct commerical interest from data used by common crawl and their api service.
You can argue about derivatives and so on. It's not clear cut in any way but your analogy isn't a good one. We can also give some leeway to humans compared to machines as latter is more deterministic. We know that gpt 3 was trained on which data. We can't say that for sure about you.
Okay, so you don't own part of my brain, but StackOverflow would? Because I definitely commercially use information I've read there.
And I mean, I know English because of all the English material I've read. If I write poetry, I'm working off training data I've read online. It would be borderline impossible to be an artist under these rules.
> We know that gpt 3 was trained on which data.
We don't know what data GPT uses for a given answer. We don't know what data I use for a given answer. But we know what data I read, and we know what data GPT read.
Far as I can tell, the brain analogy is 1:1 here.
They have a well established license you need to follow to use the answers posted on their site and what freedom it provides you.
Using stackoverflow answer to add little bit of code in your codebase is different than reselling stackoverflow answers behind a pay wall.
> But we know what data I read, and we know what data GPT read
No unless you are under surveillance 24/7.
And as I already said, machines are different than people. Copyright laws acknowledge that so human situation wouldn't be comparable. You aren't mass producing creativity and selling it.
Working for a company using information I learnt on SO is exactly equivalent to reselling SO understanding behind a pay wall.
> No unless you are under surveillance 24/7.
So if I'm under surveillance, that means I'm affected by copyright?
> And as I already said, machines are different than people.
I disagree.
> You aren't mass producing creativity and selling it.
As a software developer: yes I am!
https://www.copyright.gov/comp3/
Also see: https://stackoverflow.com/help/licensing
> As a software developer: yes I am!
No. Another reason why the copyright protection not applying to purely automatic processes make sense. A machine could theoretically generate every potential expression understandable by the compiler or interpreter. You can't. You aren't capable of the same mass production. You will die at some point and your work will be released for public but same won't happen to machines.
> So if I'm under surveillance, that means I'm affected by copyright?
No. It just makes it easier to prove original sources.
The reason why I proposed my initial argument is simply for accountability and right to information.
1. Unlike stackoverflow which every participant agrees on a license when they signup, common crawl is different. I believe everyone has right to the weights derived from their data unless they explicitly sell them or give permission.
2. When gpt3 is used for interviews and other filtration: it is important for anyone to be able discern it.
Sure I can, it'll just take a while.
> You will die at some point and your work will be released for public but same won't happen to machines.
I'm not arguing machine produced work should be copyright protected. This is about quite the opposite, about how copyright protection relates to the process of learning as exercised by machines vs. humans.
> 1. Unlike stackoverflow which every participant agrees on a license when they signup, common crawl is different. I believe everyone has right to the weights derived from their data unless they explicitly sell them or give permission.
The big disagreement here is that I think posting anything on the internet gives an implicit license for "brainlike consumption" - reading, understanding and learning. GPT isn't doing anything while training that I'm not doing while reading. It would be impossible for a human to read the text at all without engaging in these processes, and so it must likewise be permissible for GPT.
Consider the futuristic scenario of a brain implant extending your memory for Alzheimer's patients. GPT is not so dissimilar from such an implant operating in standalone mode. I don't see how screwing a brain onto it should change the ethical algebra to make it more permissible.
That's another disagreement. I consider the scale and classification (machine?) to be an important distinction.
I would like to think more about this. I probably won't change my initial stance on requiring ml models to be public if they are used commerically even if point 1 shouldn't apply. However, whether they get IP protection of any sort and who owns it depends on sorting out other points I have considered after this.