Post

I'm reading Bartz v Anthropic, of course, from a very practical, interested perspective. ("What can libraries let me do, and how exactly can we stay out of jail?") The first thing I notice: I didn't realize Anthropic had rolled their own GBooks. Doesn't sound prohibitively $$$. 1/3

Central practical detail. It matters a lot, in this ruling, that Anthropic did not serve their model in a way that produced infringing copies. Having the potential to do it is not illegal. The gray area here for libraries will be "What if a model is mounted on Hugging Face?" But Alsup's ruling +

seems to imply that the liability in that case would rest with Hugging Face or whatever service runs the model / uses it to infringe. Having created a model that *could* infringe is not actionable here. Finally, not a practical point, but this analogy will piss ppl off and I'm glad.

thx for summary! im still wondering if it implies what ur saying. the judge is giving a pass because the filters around the model prevent content leakage. couldn’t one also interpret this as open model developers may be expected to provide such filters alongside released weights?

Yeah, I agree the question of liability in that case isn’t directly addressed. I started out confident but eventually worked around to reconsidering: bsky.app/profile/tedu...

And I guess AI2 might be an exception to the last sentence there. It’s not a wholly theoretical question for you.

Also I think it’s important to see that this kind of argument — repeated prompting with half of the text can eventually extract most of the other half — is having zero effect in the courts. This suit was just decided for Meta. www.404media.co/meta-ai-mode...

Neither of the rulings we’ve gotten this week is at all interested in deciding whether there is a copy of the book latent “in” the model in some metaphysical sense. Both imply that an infringing copy has only been produced if a reader in practice uses it as a substitute for the book.

It looks like memorization is very model-dependent and book-dependent. In general, it does not happen much (<1%), but it's hard to explain why some LLama models memorize some books quite a lot. (Also even 0.1% of a few million is a lot to some people)