Why the Fable cutoff makes open weights inevitable

The US worried Anthropic’s Fable model was too good to let out, and blocked it on national security grounds. The irony is that in doing so, they’ve woken every other country up to a much bigger security problem than the one they were trying to solve.

Here’s what the rest of the world just clocked: a frontier model can be switched off from abroad, at any time, on someone else’s whim. Not because of anything you did, but because a government on the other side of the planet decided the politics had changed. That’s a rough place to be when you’re depending on these things for diplomacy, for trade, for whole swathes of industry. You’re building on foundations another state can pull out from under you, and you’ve no say in it.

The risk was never really the model. It’s the dependency.

There’s a fix, and I reckon the world ends up demanding it whether the labs like it or not: open weights. Models whose weights are released, that anyone can download and run.

The fairness case more or less writes itself. These models are trained on humanity’s collective knowledge — a great deal of it copyrighted, hoovered up without anyone being asked. If a frontier model is essentially a compression of what we already know, the argument for one company keeping it locked away gets harder to make the longer you look at it. World knowledge, you could reasonably argue, ought to belong to the world.

And once the weights are out, something interesting happens to hosting. It stops being a service you rent and becomes something a country can own outright. Run the models at home, on home soil, as national infrastructure: sovereign, impossible to switch off from abroad, and a revenue stream in its own right rather than a subscription cheque posted overseas every month. The models are enormous and still need serious specialist hardware to run — but that’s exactly the sort of thing states already know how to build and operate.

None of this means the labs get nothing. They spent the money building these models, and fair enough, they should be paid for it. But once a model is genuinely strategic infrastructure, licensing it stops resembling a vendor contract and starts looking like a government-to-government arrangement, the way strategic technology already gets licensed between states. A country pays to use the weights; the weights themselves live on its own hardware. If relations sour, that becomes a political matter to be settled between governments — while the businesses and industries built on top of the model carry on regardless. The lab loses revenue. It doesn’t get to flip a switch.

That’s the whole point. The leverage moves from the provider to the country running the thing.

Frontier models keep creeping towards must-have, just as it’s dawning on everyone that you can’t fully trust them either — or rather, you can’t trust the arrangements they currently come wrapped in. Those two facts can’t sit together forever. Open weights are the only way out of the bind, and the Fable cutoff has made that a great deal more obvious than it was a month ago.