David Manheim says European AI labs avoid distilling US models due to IP lawsuit risks

@DadaJudith have Europeans considered this

@tamaybes High effort, low nobility. Not the European way.

@tamaybes Now that a few more famous people have commented, I think the answer is that anyone in Europe who would do this, moves to America. Its culturally not European to do things.

@davidmanheim It likely breaches ToS, but it's not clear that the legal risk is greater than what highly-risk seeking startups routinely take on.

@tamaybes ...but how many of those high-risk startups are in Europe, where courts, and public, take a far dimmer view of such risk-taking?

@tamaybes Scrupulous. It’s illegal to do so. It’s also more expensive to acquire training data than for Americans (can’t just buy books)

@tamaybes Well, it's very adversarial, no? Europe is much wary of taking hostile action against the US, and the downsides could be really large.

@ZiChengCaoHuang I'm pretty sure they'd prefer being on the frontier.

@tamaybes What makes you think they don't? At a minimum I'm guessing they're using private models for synthetic data generation and probably doing OPD in post-training. Building a distillation pipeline is still really hard.

@tamaybes Because Europe can't even figure out Turbo Prolog yet.

@AdriGarriga @grok is distilling models from US labs illegal?

@tamaybes I mean, it's one of those things, that even if you're gonna do it, you're not gonna say you're gonna do it...

@marcusabramovi1 Yes, I think that’s right.

@tamaybes because the chinese aren't worried about being sued to oblivion.

@tamaybes Europe is a highly regulated environment and it's against the ToS.

Also the EU wouldn't find the API fees

@tamaybes Even ignoring the ethics, which seems relevant, it's probably because Europe has functioning courts that will enforce laws when the AI labs sue those companies.

@tamaybes same reason reta copies come from china

europeans are just honest and dont see the problem yet.

@tamaybes EU AI Act requires disclosure where you got the data for training of large model, article 53.

Next step, you need to respect opt out, and I'm 99% sure that Anthropic/OpenAI efforts to stop distillation count as opt-out.

@Simon__Grimm EU regulators going easy on startups distilling from US labs seems much less hostile than e.g. the EU commission imposing billions of $ in fines per year on US tech companies.

@tamaybes Because you dont have US models to distill. What US open source models the US have?

Mistral was literally copying Deepseek, so they released the model closed.