Hugging Face CEO Clément Delangue argues AI regulations should target closed frontier APIs instead of open-source models

@roanoke_gal They’ll still be much less dangerous than their closed-source counterparts

@ClementDelangue 🎯 and if you look at the actual misuse, it is dominated *heavily* by closed models https://florianbrand.com/posts/open-model-safety

This entire post is backwards

Giving anyone access to models with these cybersecurity abilities is reckless and you know it

Your arguments are only true because open weight models haven't gotten to fable-level cybersecurity abilities yet.

Once that's true, all the same problems of closed models are also true for open weight models, except you have zero control over who uses them. Open weight models will also become far more commercially viable and adoption will accelerate even more than it already has, making it no longer a niche hobbyist industry but likely the majority of inference served

Also "regulation is supposed to protect small businesses" - lol, you must be new here

@PiotrPadlewski I’m saying closed source APIs and interfaces will always be inherently more dangerous than open weights (mostly because there’s 0 transparency on their capabilities and risks)

@ClementDelangue Are you claiming oss will never catch up with frontier capabilities?

"They're not as massively distributed or as easy to use" are you saying hugging face and other oss providers are hard to use?

Also would love to see a bit more publications on safety and interp as you claim

Rare L take.

"They're not as massively distributed or as easy to use, especially the big ones, as APIs and assistants. We (including governments) can quickly and accurately analyze what they're capable of, and for now everyone confirms they're not as good as the APIs at doing bad things."

Ok, but what about 1-3 years from now, when they are capable of doing 'bad things', as current closed-source models are?

@ClementDelangue The war between open source and closed source all over again

@ClementDelangue A government official reading this post won't go "Oh, whew, we'll ban closed models and only allow open ones", they'll go "Closed models are dangerous but can be gatekept, restrict them to the elite. Open models are somewhat less dangerous, but can't be restricted, ban them."

@ClementDelangue Frontier models is a term the AI labs gave themselves. I can still have a frontier model and not call it a frontier model. It's a label.

@ClementDelangue More specialized models may also be a path to protect open source

@ClementDelangue How in the world can someone consider open source AI as dangerous except that it threatens his business 😅

@ClementDelangue this take only works if open models never catch up

what happens when a Llama-sized release outperforms GPT-6?

@ClementDelangue @roanoke_gal But still much more dangerous than any closed api model is right now, and without a kill switch or any way to monitor and mitigate usage from bad actors.

@ClementDelangue The push to regulate open source is right around the corner. The AI safety crowd ultimately wants to regulate open source more than proprietary. They want to ensure open source is never the frontier.

@ClementDelangue Sorry, so are Anthropic intentionally regulating themselves to sell more or does it make sense to regulate them?

@ClementDelangue It is quite rational to regulate both of them in production.

@ClementDelangue open weight being "easier to analyze" only holds until someone fine tunes it for something nasty offline where nobody's watching.

closed APIs at least leave a log

@ClementDelangue been saying this for months

the transparency argument alone should end the debate - we can actually audit open models vs black box APIs

@ClementDelangue They aren't doing it for any of those reasons but because they want to have a cyber edge on Chyna for their microwave reheated cold war.

@ClementDelangue Argument 1’s first two points are pretty strong. Closed source frontier lab’s model are built behind closed door with zero transparency of how it works, and megacorp who controls them have profit seeking incentive that’s fundamental misaligned with incentive of society — yes.