Potentially dumb question from a guy who isn't a lawyer:

Does Copilot infringe Google's patent(s) on the Transformer architecture? If so, then Google could potentially sue them for royalties, at least.

Further, couldn't this Copilot thing backfire for Github because customer trust is more valuable than AI training data right now? If folks don't feel they can trust Github, seems like they could move their work to other version control systems like Gitlab or Bitbucket...

Doesn't really matter, because if Google sued Microsoft, Microsoft would immediately hit back with a countersuit, since they would have little trouble finding something in their 90,000+ patent warchest that Google is infringing on. Software patents have become a matter of mutually-assured destruction for the big players. The only winning move is not to play.

> For example: Let’s train a network on Beatles music to generate new Beatles songs. I’m pretty sure music lawyers will find a way to prove that the trained network is violating the label’s copyright, as they always manage to do that.

The people making the machine that learned (and recites) beatles songs aren't infringing though (most likely). It's those that use the machine to create and distribute the new works that are.

Same here. No one will be able to say that Copilot itself is a "derived work" or somehow uses the code in a way similar to a computer program (Although such claims have already been made - I highly doubt that's the case). But those that produce a whole file full of GPL code verbatim (Which will be rare, but WILL happen), are at risk of violating the license terms if they distribute it under the wrong license.

There is an absolutely enormous archive of fan-taped Grateful Dead shows out there, someone with much more time and money than me needs to train a network on that!

username checks out lol

Wouldn’t a more accurate metaphor be “let’s train a network on all music, to generate new music”, which includes Beatles, and may generate songs that contain the same chords as the Beatles used?

Yes, but may also use the same chord progressions, lyrics, or melodies. Could even say it contains snippits of the actual recordings, depending on how you look at it

Sure but then it’ll definitely be harder to prove it’s actual copyright infringement, especially when only a very small part of the song may have some snippets of the Beatles. Could it then, perhaps, be considered fair use?

Yes you'd have exactly the same kind of lawsuits and arguments that already exist today around fair use. It doesn't matter if a tool creates the new work or a person creates it (without tools?!!?) because ultimately it is a person who claims the new work as their own and distributes it, and that is the person who will get sued by the record companies if their work is too derivative. Establishing the line for "too derivative" in any particular case is a very lucrative field already I'm sure.

Or contain new chords that it synthesized from its knowledge of the ones the Beatles used.

In ancient Rome they didn't have a police force. What they had was essentially muscle for hire, mercenary bands paid by rich and powerful folks to do their bidding. As a regular person, the only thing that could have protected you from one of these groups was another such group.

Same today with the licensing system.

They had cohortes vigilum and cohortes urbanae in Ancient Rome. Why don’t they count as police?

Or Why not find some leaked windows/office source code and try to train a model to reconstruct microsoft software, then open source it? This surely must be legal, they're doing it themselves after all :D

(Maybe bring oracle into this :D)

It would be legal! But it wouldn't "reconstruct" Microsoft software. The way Copilot works is just that, a copilot. It's not the pilot. It's your own fault for what you do with it, it's just giving you some help along the way.

So long as the "copilot" is a black box that no one can inspect, how is it substantially different than me creating a website with a link to download a licensing-stripped version of Microsoft office, but it only gives you a verbatim copy 1/10 times you try it?

There's an entire academic paper detailing exactly how it works. https://arxiv.org/abs/2107.03374

That already exists though? SongSmith and other similar tools are used by musicians a lot.

At what point is it not a derivative work?

afaik, chord progressions aren't copyrightable, and even some lyrical things aren't. Melodies are the main thing, I believe. (I could be wrong, this is just what I have been told in the past)

Note: that already exists, it's called Jukebox! https://openai.com/blog/jukebox/