This past Friday, the US Government issued an export control directive that prohibits Anthropic from giving foreign nationals access to Claude Fable or Claude Mythos, their latest models.
The leading AI researchers today are the people who got in on the ground floor a decade ago. Most of them did so because they believed artificial superintelligence was a real possibility. This is why OpenAI did a bunch of stuff that no other tech company did, like incorporating as a nonprofit! That's just far too costly an action to take for PR reasons, otherwise other companies would do it.
Anthropic was founded by the people who felt OpenAI had lost the plot on AI safety. This culture may have been diluted by the people who joined since, and obviously power and money corrupt anyone over time, but clearly Amodei and Anthropic leadership in general still believe in AI safety. Their actions just don't make any sense otherwise. If they were profit-oriented, they would be deeply stupid to warn everyone about how terrible their technology is going to be - talking about it like it's nuclear weapons, instead of like it's electricity. Clearly they're not stupid, or they wouldn't be leading the most valuable AI company in the world. I see lots of people bending over backwards trying to decipher their 4d chess game. But sometimes the simplest answer is the truth.
So yes, they did ask to be regulated. But the regulations were applied with no warning, to one model, apparently in response to jailbreaks which have been possible for years, apparently used to make the model do things which many other models are capable of. The administration has been asleep at the wheel, someone finally got through to them, and they responded with this half-assed, knee-jerk intervention--even giving them the benefit of the doubt and assuming this isn't retribution for the DoW episode. Seriously, access is limited based on US citizenship? Meaning Anthropic employees who helped build the model can't use it. How long until they go work for Deepseek instead? Meanwhile recently having relaxed export restrictions to Nvidia. It's like nobody in the Trump administration is reading Zvi Moshowitz at all!?
Details matter though. Do you really think there was no alternative way of the USG handling this that would still be congruent with Anthropic's position?
Do you think Anthropic's statements necessarily imply that their own employees get cut off from their own models, that their company possibly gets destroyed over this?
I think Zvi has it right on this one (I get you're not a fan, but at least this has a lot of additional context):
I basically think he just accepts Anthropic's overall narrative here, and Anthropic's POV is "coincidentally" that basically any situation resembling this, but not concerning them specifically, would be bad, but somehow all of the details of this are such that they are clearly in the right
Or I guess: Basically Anthropic thinks 1) that in a judgement-call way, they're right and not wrong for thinking that this specific jailbreak is no big deal and the government shouldn't step on them, but 2) in general, private entities like Anthropic shouldn't get to make their own judgement calls and the government is better equipped to make those judgement calls.
I don't think this is consistent, or that they'd be arguing the same on behalf of anyone but themselves.
It's not nefarious in itself for anthropic to believe that they are doing the right thing. Isn't that what most people believe? Or they would do something else?
In principle the government should be the one making the rules, but they also need to be making correct ones. It should not come down to a judgement call.
I would celebrate this development if it were the result of a transparent, objective decision-making process that would be applied fairly to all companies. But from the available information, generalizing this restriction would probably mean taking down many other models.
In the absence of objective regulations, nerfing Anthropic is more likely to just allow OpenAI and Google to catch up. If someone's gonna build ASI I think Anthropic is the least bad option, although I think that their safety commitments have a shelf life just like OpenAI's did.
In general, I think governments are better placed to handle policing than private corporations, but that doesn't mean I have to agree with everything the police does or every instance of concrete policy decision a particular government makes about policing.
I do think there's probably reasonable criticisms of Anthropic to be made here (though I wouldn't call myself an expert on the situation etc). In particular, I'm not sure why they thought they could work closely with this specific government and that not going poorly, which is I think related to your points.
But there's also this meme going around right now (including a literal meme image) that says just because they talked about ai being possibly dangerous and about governments needing to get involved, therefore they deserve literally every government action directed at them, no matter how arbitrary or authoritarian... And I think that's going too far.
Anthropic believes - pretty reasonably - that it is vastly smarter than the people running the US government. Anthropic also believes it holds the moral high ground in AI safety, and that the government ought to defer to its expertise. It openly advocates for regulatory capture, by itself, as a moral good. People can call that incredibly arrogant (and I do), but it's consistent and coherent.
But I just can't wrap my head around the naïveté in them believing that smarts and morals somehow convert to real power, and that they'd actually be able to play the government - particularly _this_ government, after aggravating them several times already.
I don't use cloud LLM models and could not care less about government regulating them into oblivion. My local models, however, you would have to pry from my cold, dead hands.
The leading AI researchers today are the people who got in on the ground floor a decade ago. Most of them did so because they believed artificial superintelligence was a real possibility. This is why OpenAI did a bunch of stuff that no other tech company did, like incorporating as a nonprofit! That's just far too costly an action to take for PR reasons, otherwise other companies would do it.
Anthropic was founded by the people who felt OpenAI had lost the plot on AI safety. This culture may have been diluted by the people who joined since, and obviously power and money corrupt anyone over time, but clearly Amodei and Anthropic leadership in general still believe in AI safety. Their actions just don't make any sense otherwise. If they were profit-oriented, they would be deeply stupid to warn everyone about how terrible their technology is going to be - talking about it like it's nuclear weapons, instead of like it's electricity. Clearly they're not stupid, or they wouldn't be leading the most valuable AI company in the world. I see lots of people bending over backwards trying to decipher their 4d chess game. But sometimes the simplest answer is the truth.
If they care about AI safety, why develop AI at all? Their attitude in general is that they would like AI to slow down or stop (https://www.reuters.com/business/anthropic-says-ai-labs-need-coordinated-plan-halt-development-if-risks-rise-2026-06-04/), but if there's an AI race they still want to be in the lead, because they are the most concerned about safety.
So yes, they did ask to be regulated. But the regulations were applied with no warning, to one model, apparently in response to jailbreaks which have been possible for years, apparently used to make the model do things which many other models are capable of. The administration has been asleep at the wheel, someone finally got through to them, and they responded with this half-assed, knee-jerk intervention--even giving them the benefit of the doubt and assuming this isn't retribution for the DoW episode. Seriously, access is limited based on US citizenship? Meaning Anthropic employees who helped build the model can't use it. How long until they go work for Deepseek instead? Meanwhile recently having relaxed export restrictions to Nvidia. It's like nobody in the Trump administration is reading Zvi Moshowitz at all!?
Details matter though. Do you really think there was no alternative way of the USG handling this that would still be congruent with Anthropic's position?
Do you think Anthropic's statements necessarily imply that their own employees get cut off from their own models, that their company possibly gets destroyed over this?
I think Zvi has it right on this one (I get you're not a fan, but at least this has a lot of additional context):
https://thezvi.substack.com/p/american-government-takes-down-claude
I basically think he just accepts Anthropic's overall narrative here, and Anthropic's POV is "coincidentally" that basically any situation resembling this, but not concerning them specifically, would be bad, but somehow all of the details of this are such that they are clearly in the right
Or I guess: Basically Anthropic thinks 1) that in a judgement-call way, they're right and not wrong for thinking that this specific jailbreak is no big deal and the government shouldn't step on them, but 2) in general, private entities like Anthropic shouldn't get to make their own judgement calls and the government is better equipped to make those judgement calls.
I don't think this is consistent, or that they'd be arguing the same on behalf of anyone but themselves.
It's not nefarious in itself for anthropic to believe that they are doing the right thing. Isn't that what most people believe? Or they would do something else?
In principle the government should be the one making the rules, but they also need to be making correct ones. It should not come down to a judgement call.
I would celebrate this development if it were the result of a transparent, objective decision-making process that would be applied fairly to all companies. But from the available information, generalizing this restriction would probably mean taking down many other models.
In the absence of objective regulations, nerfing Anthropic is more likely to just allow OpenAI and Google to catch up. If someone's gonna build ASI I think Anthropic is the least bad option, although I think that their safety commitments have a shelf life just like OpenAI's did.
You should consider the current debacle evidence for whether or not Anthropic has the judgement required to build ASI.
OpenAI has shown good judgement in the pursuit of profits. Would you prefer Anthropic do more of the same?
Both options are bad, but on balance I think I have more faith in Moloch than whatever's going on at Anthropic.
In general, I think governments are better placed to handle policing than private corporations, but that doesn't mean I have to agree with everything the police does or every instance of concrete policy decision a particular government makes about policing.
I do think there's probably reasonable criticisms of Anthropic to be made here (though I wouldn't call myself an expert on the situation etc). In particular, I'm not sure why they thought they could work closely with this specific government and that not going poorly, which is I think related to your points.
But there's also this meme going around right now (including a literal meme image) that says just because they talked about ai being possibly dangerous and about governments needing to get involved, therefore they deserve literally every government action directed at them, no matter how arbitrary or authoritarian... And I think that's going too far.
Fwiw, even more details: https://thezvi.substack.com/p/the-once-and-future-fable-2
Anthropic believes - pretty reasonably - that it is vastly smarter than the people running the US government. Anthropic also believes it holds the moral high ground in AI safety, and that the government ought to defer to its expertise. It openly advocates for regulatory capture, by itself, as a moral good. People can call that incredibly arrogant (and I do), but it's consistent and coherent.
But I just can't wrap my head around the naïveté in them believing that smarts and morals somehow convert to real power, and that they'd actually be able to play the government - particularly _this_ government, after aggravating them several times already.
I think they have just lost perspective on anyone who does not work at Anthropic, having wrong-footed their customers and the government in one week.
I don't use cloud LLM models and could not care less about government regulating them into oblivion. My local models, however, you would have to pry from my cold, dead hands.
fantasic article. you are right. they tried to lobby it against others, only to have it used on themselves.