Expert Questions Feasibility Of US Ban On Open Source AI

The only version of an OS ban that would have the intended effect is for the US to convince China to stop releasing OS models. The Chinese might do that for self-interested reasons, but they definitely won’t do that to help out the NSA or the US frontier AI companies they finally stand a chance to surpass.

@yonashav It’s less complicated: open weights (not open source! No build files!) just want to be praised for a reckless, callous, antisocial embrace of, “the rockets go up, but where they go down, that’s not my department, said Werner von Braun”

@janekm I’m skeptical the admin would go along, if there is no actual natsec benefit. It is directly to the disadvantage of actors like Amazon and Microsoft to stop serving OS alternatives, and they’re at least as politically powerful.

@yonashav But it’s probably not unrealistic that some US labs would ask for regulation of “unsafe foreign AI models”, if for no other reason than to protect their business 🤔 if you pay the right lobbyists logic doesn’t really come into it.

@arthurctellis But this still doesn’t apply to Chinese OS models, which lead the capability frontier, right?

@yonashav Oh right, that's a good wrinkle @AndrewCurran_

Banning open source only makes sense if OAI & A\ are behind, ie OWs have surpassed Fable. Suppose it was Fable equiv, banning it would mean Fable also was banned. So it must be what followed, then Fable must be final public model

@yonashav No reason that a U.S.-origin norm can't be coopted by the PRC. The best way to precipitate Chinese analogous action in this period of pseudo-detente is through some combination of independent initiation and diplomatic engagement.

@David_Kasten But banning its use doesn’t actually hurt OW developers, who aren’t making money on that use anyway?

@yonashav It also seems plausible the the USG (and, then, PRC) could have more leverage to regulate/ban OS models in a words with a US-China agreement to slow down/pause (but not necessarily stop) AI development, esp given e.g., bio/cyber-sec risks.

Many people underappreciate the degree to which open source models actually hosted and being used in the United States impact the Constitutional analysis of any restrictions. By contrast, there are many ways to argue that various theories under First, Fourth, and Fifth Amendments don't apply to closed source hosting.

And, practically speaking, the small number of closed source providers may simply agree not to sue, which wipes out a whole set of other potential objections because they would be the only possible plaintiffs with standing.

@yonashav You have a lot more faith than me in the ability of government to pursue rational policies rather than mostly what’s being pushed by lobbyists, I hope you’re right 😁

@yonashav You're assuming the US govt actually thinks through these decisions. Based on how they handled Fable, that is not going to be the case.

2025 - the year of vibe coding 2026 - the year of vibe legislating

@yonashav I feel as if banning OS would not hold up in the courts anyways, would the weights and biases not be considered speech? Using Bernstein v United States as precedent.

@yonashav Good post. Seems like banning OS models will have a far-reaching and chilling effect on the AI industry in general, without addressing the root cause.

@yonashav But yeah, this is at best a low-medium EV proposition.

@yonashav @AndrewCurran_ The counterargument might look like the 2 labs can monitor/control access. But current ceremony (and impossibility of being jailbreak proof, in fact it's now automatable with other LLMs) clearly shows that's insufficient in gov's eyes.

@yonashav "I think the costs are too high to too many other actors"

You assume anyone making policy even knows or cares anyone else has a stake in AI. They probably think Anthropic invented it.

@yonashav @janekm I think you could say something similar about tariffs, but they happened anyway, and the logic here would be similar (protectionism).

Also, I don't think Amazon (big Anthropic investor) and Microsoft (big OpenAI investor) would be happy if those companies died.

@yonashav To be clear, I don’t think this is in the Overton window (nor likely under the current admin), but it seems plausible post-2028 if the admin is sufficiently AGI-pilled (or there’s significant political backlash against AI development/proliferation)