An original composition based on a statistical analysis of the training data. Statistical data about a copyrighted work obviously isn't necessarily a derivative of that work. Otherwise Tolkien could sue me for telling you how many times The Lord of the Rings uses the word "the".
If you trained an LLM repeatedly on nothing but the text of LOTR until it could re-produce the books verbatim and then tried to sell copies of that LLM, then I agree that would be blatent copyright infringement, yes.
The industry is banking on Author's Guild v. Google to be precedent in such a way that it's functionally transformative enough to be a completely new work.
I don't know what the legal answer will be, but I believe it should be considered distribution. A model is basically a highly lossy and extremely compressed copy of its training data, available as a content-addressable database. To anthropomorphize, the model is trying to perfectly replicate its training set, its brain just isn't big enough to do so.
How about if I play your song at just the right speed with just the right EQ and I can get an exact reproduction of some of the songs you claim to have written? Because we can get large excerpts of exact copies of short and long form content as demonstrated clearly by the New York Times research on chatbots and their own content.
Vstream looks super cool. Can we also use it create subscriptions that can bind with ReactHooks on the front-end ? I think PlanetScale can easily deliver amazing or better than firebase subscriptions. All we need is React and NextJs SDKs to get started with :-)
