✨ Nora Tindall  🪐

True - but if it includes AGPL code, which it does, even the server-side code is covered.

And my position is that if GPL code is part of the training data, and all they’ve done is multiplied that by some sigma values, the generated code is a derived work based on that GPL code and thus must be distributed under the same terms.

you may have forgotten the triviality bit: copyright won't apply to a trivial bit of code -- and these all seem trivial, and you can imagine that's how the model is built: include only snippets with X000 or more appearances in the corpus cc @hipsterelectron who's also thinking

about this. i prefer the copyrighted model approach: the oracle v google case went around the structure sequence and organization of the N thousands APIs: even if individually e.g. java.math.max() isn't copyrightable, the whole body is maybe that's a better copyright argument

to say that the use of the code isn't transformative, but that the whole model has been trained on a corpus whose sets of licenses have to be respected anyway, I'd rather hear an actual lawyer talk about this one ;)

i have determined that we are absolutely fucked but feel free to wait for a lawyer to confirmhttps://twitter.com/hipsterelectron/status/1410391147868946433?s=21 …

you're allowing the “this was mediated by tech” logic confuse you: if a snippet is big enough and you have copyright on a similar enough bit of code -- the mechanism by which someone copied it doesn't matter copyright protects the expression of an idea regardless of copy method

if a large snippet is constructed by you+copilot in multiple steps and ends up looking just like someone else's, you done goofed ;) i would also like to attach to the model itself, though -- a ML model is a derivative work on its own, and i which this wasn't still being debated.