EleutherAI Director Disputes AI Distillation IP Theft Claims
Stella Biderman explains distinctions between distilling public models and unauthorized data use in AI training.
Stella Biderman, Executive Director of EleutherAI, states there is nothing illegal about distilling publicly available models because outputs are not IP. She contrasts this with cases where companies like Anthropic illegally obtained copies of books for training. In replies with Leon Derczynski, she notes distillation of private models via unauthorized access differs entirely. The exchange addresses how sources mix terms around terms-of-service violations, espionage claims, and general machine learning techniques, without resolving broader IP disputes.
Combined views
4.5K
10 posts, first seen 2h ago
Reactions from ranked influencers
10 posts
There is nothing illegal about distilling a publicly available model, because model outputs are not IP. Sarah knows this, as the CO has made it extremely clear, but doesn’t want to admit it because it’s not in Anthropic’s interests.
Thanks @mkratsios47 for speaking out on this important issue. Illicit, adversarial distillation is IP theft and industrial espionage that supports adversary military and intelligence capabilities. It is a national challenge that creates serious national security risks for the United States and democratic allies. We'll continue to crack down on it and work with the Administration and Congress to maintain American AI leadership. 🇺🇸
If Kimi was illegally gaining access to Anthropic’s models that would be totally different, just like it’s 100% legal to train a model on data without permission from the copyright owners but the way Anthropic illegally obtained copies of books to train models wasn’t okay.
@BlancheMinerva There's a very careful line being walked between presenting a narrative that "distillation only means distillation against TOS" to one audience, and saying "all distillation is espionage" in others. I don't know if I would attempt to walk that "line".
@LeonDerczynski Can you unpack this a little? I’m not sure what you mean (and not even confidant who you’re being critical of)
@BlancheMinerva oh goodness, i'm not sure if I can in some contexts, "distillation" is being used to refer to "distillation of private models to copy them" in other contexts, "distillation" is being used to refer to the general ML distillation technique
@BlancheMinerva these are not the same - one is a strict subset of the other the first is used when building agreement about distillation being bad the second is used when talking about "something that must be done"