The Government's position is that Anthropic/OpenAI are forbidden from allowing non-US citizens use the models. This is impossible in reality because how can Anthropic know that the person sitting behind a Claude code session or API key is at any given moment a US citizen? You can check their ID on signup but how can they know that they didn't give their credentials to someone else to use? They can't.
Given the impossibility of compliance, what Anthropic and OpenAI are doing is working with the government to release it to certain organizations with the government's blessing.
If this were about missiles and not AI models, nobody would question this turn of events. If the government said that nobody can export this missile or allow non-citizens access to the missile, and then they started giving permission for certain organizations to handle the missile, that would be normal, not picking winners.
The only reason people are questioning it in this case is because they believe that these models are not dangerous enough to deserve these kinds of export controls. Personally I'd agree that in my 3 days of using Fable I didn't observe any superpowers. Unfortunately however, Anthropic undermined that argument by claiming that Mythos is highly dangerous, which set them up for any jailbreak of Fable to be considered a national security risk. Who is a court going to believe? Someone who used a model for 3 days? Or the government and the people who made that model?
Comparing AI models with missiles is a far fetch as long as citizenship is the single qualifier used to decide who’s allowed to be a customer. This is not a security related policy, it’s about strategically controlled economic power.
This is just not how it works even if we really, really want it to work that way. US government can do it, and has done it before. At some point strong encryption was considered “munitions” and export controlled. If SSL can be “munition”, LLMs can be slapped on a label just like that. SSL/TLS stopped being qualified as such eventually so some sanity was restored. But as the legal and regulatory framework, it’s certainly there and has and is being used in that capacity.
> This is not a security related policy, it’s about strategically controlled economic power.
It is both. The US and China are locked in an AI arms race with economics and security intertwined, given the perceived power of the trajectory of frontier AI models.
This has a long history and the US government are not completely stupid. They understand that some cornerstone technologies are high leverage even if they are not per se directly militarily relevant.
Some classic examples of this, on which the US places severe export controls, was advanced materials science and inertial navigation technology. Neither of these are weapons but advanced technology in these domains greatly enables the development of advanced weapons. Any work in these areas is automatically subject to the full export control regime. In extreme cases they may be nationalized and classified.
AI tech is becoming just another tech domain subject to the same level of scrutiny. I’m not making a moral judgement. This was always going to be the reality and a lot of people could see it coming.
Another one is thermal vision cameras. Anything above 9fps (I guess an arbitrary cutoff to do with night-vision goggles) is export controlled. This is despite the technology being very valuable for many industrial and hobby applications.
If that was a guesstimate on frame rate it's likely off.
Regardless of US controls, I can from Australia buy a high quality 50 fps thermal scope with built in 1000m (or ~ 1000 yard) laser range finder from Hangzhou, China.
ie. Why lock the stable at 9 fps when out on the open plain 50 fps is already running free.
The US government obviously has no control over what China sends to Australia, so it this case the cat is out of the bag. However I can assure you the limit is still real for US companies: https://oem.flir.com/support/support-center/knowledge-base/a...
Cheers for the follow-up .. and that's certainly interesting.
Chinese 50 fps thermal cameras and scopes have been about for a while now, they've been supplying the industrial specialist market even longer than the general consumer goods market.
I note the US export control is limited by licence .. which likely means the US high frame rate thermal gear is equally as available in AU as (say) TI DSP chips were when they first iterated that family (ie. restricted by licence but easy to purchase in allied countries if you just sign a form stating you're not supplying them to countries in the Chinese sphere of influence, etc.)
Now that China is also making them I wouldn't be surprised if they drop this. But FLIR has been doing milspec thermal cameras for over 20 years and I think the regulation certainly slowed down proliferation of this equipment in the early days.
I don't think this is what happened. Anthropic could have absolutely complied with the government order, even in the most minimal fashion possible. Instead they chose to block everybody, and then give a link blaming the government.
I think it's most likely that they felt that this would drive wide antagonism towards the government which would help put them in a more favorable situation for future negotiations to establish a more 'commercially favorable regulation regime.' In other words, build me a moat, now! The government responded by super-screwing them, but is doing so in a way that can help keep the corporate class relatively content while this plays out.
"Our models are so good they are banned by the government"
I don't see any way how releasing an incrementally better model could get even close to the positive PR. For every Mythos token not sold because of this they'll sell dozens of Fable, Opus and so on. Perhaps the ban did not origin at Anthropic, as a PR stunt, but genuinely at some government desk. Entirely possible. But Anthropic would be foolish to not pick up that ball and drag it out a little.
Yeah but that's not what happened at all. The government said limit access to US citizens. Anthropic could have chosen compliance or even malicious compliance (are you a US citizen - please tick yes or no). And such a response from Anthropic would have required the government to try to crack down even harder which would have made them look even worse, and Anthropic even better. Instead Anthropic chose to over-comply and looks just as bad the government here, especially with somewhat transparent ulterior motives driving their behavior.
In addition to what you're saying, Trump admin is also hand-picking which customers gain access to GPT 5.6 during this period. This is widely reported on.
That's why I raised skepticism that Anthropic has not been subject to that.
Given the impossibility of compliance, what Anthropic and OpenAI are doing is working with the government to release it to certain organizations with the government's blessing.
If this were about missiles and not AI models, nobody would question this turn of events. If the government said that nobody can export this missile or allow non-citizens access to the missile, and then they started giving permission for certain organizations to handle the missile, that would be normal, not picking winners.
The only reason people are questioning it in this case is because they believe that these models are not dangerous enough to deserve these kinds of export controls. Personally I'd agree that in my 3 days of using Fable I didn't observe any superpowers. Unfortunately however, Anthropic undermined that argument by claiming that Mythos is highly dangerous, which set them up for any jailbreak of Fable to be considered a national security risk. Who is a court going to believe? Someone who used a model for 3 days? Or the government and the people who made that model?