I don’t necessarily fault OpenAI’s decision to initially train their models without entering into licensing agreements - they probably wouldn’t exist and the generative AI revolution may never have happened if they put the horse before the cart. I do think they should quickly course correct at this point and accept the fact that they clearly owe something to the creators of content they are consuming. If they don’t, they are setting themselves up for a bigger loss down the road and leaving the door open for a more established competitor (Google) to do it the right way.
It is clear OpenAI or Google did not use only Common Crawl. With so many press conferences why did no research journalist ask yet from OpenAI or Google to confirm or deny if they use or used LibGen?
Did OpenAI really bought an ebook of every publication from Cambridge Press, Oxford Press, Manning, APress, and so on? Did any of investors due diligence, include researching the legality of the content used for training?
If you give a bunch of books to a kid all by the same author and then pay that kid to write a book in a similar style and then I go on to sell that book...have I somehow infringed copyright?
The kids book at best is likely to be a very convincing facsimile of the original authors work...but not the authors work.
It seems to me that the only solution for artists is to charge for access to their work in a secure environment then lobotomise people on the way out.
The endgame seems to be "you can view and enjoy our work, but if you want to learn or be inspired by it, thats not on"
a) In many closely comparable scenarios, yes, it’s copyright infringement. When Francis Ford Coppola made The Godfather film, he couldn’t just be “inspired” by Puzo’s book. If the story or characters or dialog are similar enough, he has to pay Puzo, even if the work he created was quite different and not a literal “copy”.
b) Training an LLM isn’t like giving someone a book. Among other things, it involves making a derivative copy into GPU memory. This copy is not a transitory copy in service of a fair use, nor likely a fair use in itself, nor licensed by the rights-holder.
Training is almost certainly fair use, so it's exactly a transitory copy in service of fair use. Training, other than the brief "transitory copy" you mention is not copying, it's making a minuscule algorithmic adjustment based on fleeting exposure to the data.
Congress took the circuit holding in MAI Systems seriously enough to carve out a new fair use exception for copying software—entirely within the memory system of a licensed user—in service of debugging it.
If it took an act of Congress to make “unlicensed” debugging a fair use copy…