Somewhere in San Francisco, an engineer who helped build the most capable AI on earth can no longer use it.
She did nothing wrong, but unfortunately because she’s not American, on Friday the US government decided that was enough.
If you're reading this from Delhi, or Lagos, or anywhere outside America, the same line was drawn around you.
For the first time in this whole story, the question is about nationality over capability.
Anthropic, an American company, had just released its two most capable models, Fable and Mythos. Honestly, I had freed up my Saturday just to play around with Fable.
On Friday, at 5:21 in the evening, the US government sent Anthropic a letter under the export-control law that stops a company from shipping fighter jets or advanced chips abroad, and it ordered Anthropic to block both models from every foreign national on the planet, whether they live in Mumbai or in Manhattan, including the company's own staff who hold foreign passports.
The stated reason that officials believed someone had found a way to jailbreak Fable, and the flaw they pointed to, is one other public models, including OpenAI's, give up just as easily with no trick at all.
The company is obeying the order while saying, plainly, that it disagrees, and that the whole thing is a misunderstanding it's trying to undo.
By the weekend Sridhar Vembu, who built Zoho into one of India's largest software companies, was writing that globalization is dead and India had better find its own way.
For two years, every conversation about AI and your future has circled one question.
Are you good enough?
Can you prompt the thing well, do you have taste, do you know what the model sounds like, is your judgment sharp enough to catch the model when it's wrong.
But this incident was one no one expected.
The first lock is skill, and you can pick it by learning.
The second, it seems, is permission, and it just grew a nationality.
For about thirty years, the software you used never once cared where you were born.
A teenager in Delhi and a teenager in California opened the same Google and pushed code to the same GitHub.
Your passport stayed in the drawer, and we grew so used to it that we stopped seeing how new it was.
It was never total.
But that list was short and always the same few names.
This week it widened to take in everyone else at once, on one of the best general-purpose tool we have.
You can earn your way past a price, but the passport you were born with has never been something effort could touch.
I want to be fair to the other side, because there is one.
Governments have always controlled their most dangerous exports by citizenship.
A model like Mythos can find and exploit weaknesses in other people's software at a scale that worries serious security researchers, which puts it closer to a controlled weapon than to a word processor. A government has a real duty to sit with that.
Anthropic itself says the state should be able to halt a release that's truly unsafe.
All of this has precedent, but never at this speed or this scale: a whole planet lost the tool in a matter of hours, on something already sitting in hundreds of millions of hands, over a reason thin enough that the company behind it calls a misunderstanding.
And the precedent stays even if access returns next week.
Once that move has been made, everyone else is building on a maybe.
Now move this off the engineer in San Francisco and onto your own work.
The founders I work with woke up to a dependency that sits on no risk register anywhere.
Their product was built on top of a capability that can be switched off by a directive issued in another country, for reasons that have nothing to do with their business and irrespective of how good their product is.
You can plan for a supplier that raises its prices or misses a deadline.
There's no clause in any contract for the supplier being ordered to stop serving you.
It's already reaching the products downstream.
An assistant app called Poke spent the weekend telling its users their replies had gotten worse, and that it was rerouting them to other models that still worked.
Swapping one model for another is seldom clean and unfortunately, there is no alternative to fable right now.
The outputs come back shaped differently, and you lose weeks finding everything else the switch broke.
And if you build outside the US, this cracks a reflex most of us stopped noticing, the one where you reach for the strongest American model and move on.
Vembu's prescription, once he'd declared globalization over, was something I tend to agree with.
Stop paying people who don't want to sell to you, and build the rest yourself on the smaller Indian and Chinese open models.
We've also only just started spending real money on our own models.
Building our own used to be a point of national pride but this week told us that it’s to ensure progress isn’t gated by regions.
If the best AI is only ever going to be for some of us, I want to know what you'd actually do.
Keep building on it and hope the line never moves again.
Build around it, loosely enough to swap it out in a day.
Or make the slow, expensive bet Vembu is describing, and build your own.
Reply and tell me which one, and why.
Nandini
PS: If you were wondering about the subject line, it’s about his quote about how “All animals are equal but some animals are more equal than others.”