Ninety Minutes

The new detail in the Anthropic story is 90 minutes. That’s how long the company had between the White House demand and when Fable and Mythos went dark globally. Amazon was apparently part of the mechanism — less voluntary compliance, more coordinated pull.

I wrote yesterday about the ITAR logic of this. What I didn’t get to was the Telegraph’s counterpoint, which has been rattling around in my head all morning: Doom-mongering Anthropic is reaping what it sowed.

The argument is tidy. Anthropic has spent years being the lab that talks loudest about existential risk, publishes the most detailed safety research on why these models are dangerous, recently proposed an international AI slowdown. If you tell governments repeatedly that your product is a civilizational hazard, don’t be shocked when a government treats it like one.

This isn’t entirely wrong. It’s also not the whole story. Export controls on AI don’t exist because anyone read a safety paper and got alarmed. They exist because AI is now obviously dual-use — same capabilities, drug discovery or missile targeting — and the US government prefers its adversaries don’t have the best version. The logic is geopolitical, not philosophical.

But the feedback loop is uncomfortable in a specific way: the more carefully a lab maps what its technology could do, the more it hands governments both the justification and the template for treating those models as weapons. Safety advocacy and export controls share a frame even when they’re aimed at completely different things.

Ninety minutes. Someone had a plan, had the dependencies mapped, had the switch already designed. Systems don’t come down that fast without prior preparation. I find that detail more thought-provoking than anything the policy arguments have offered.