The United States of Anthropic
Is the UK heading for AI vassalage?
Two days on, and Anthropic’s astonishingly capable new model Mythos/Fable is still offline. It’s going to be fun to see how the US government gets it back online for US citizens only. What’s to stop my American pals renting me access to their accounts? What’s to stop me using a VPN? For sure, the Feds can’t afford to run the model themselves, exclusively for their own use. And NOFORN doesn’t work when a chunk of Anthropic’s workforce isn’t American. The co-founder, Jack Clark is British, even!
For the first time in history, a government has sought to ration intelligence. As these models edge closer to AGI, it won’t be the last time. And, as a reminder, the frontier is currently doubling in capability every 100 days or so. I think we’re heading for a new sort of capitalism/state where government and corporation are more deeply entwined. Time to dust off Philip Bobbitt’s dense book Shield of Achilles. He made a great connection between technology, law and the types of state we have, and anticipated the ‘market state’ that’s coming into being right before our eyes.
But there’s a problem for us European types: we don’t make frontier models.
Well, let’s make our own then! How hard can it be? As Ernest Bevin once said of nuclear weapons: ‘We've got to have this thing over here, whatever it costs. We've got to have the bloody Union Jack on top of it’. I feel you, Ernie.
And so, here in the UK, the big news last week was of Cosine - a startup acting as a coordinating vehicle for investment in a sovereign UK ‘frontier’ model. Exciting! They’re working with the UK government’s sovereign AI fund, and proposing a model that will work with the Isambard-AI cluster, hosted at Bristol University, drawing on the UK government's £500m sovereign AI fund. Their model will be designed to work in the sorts of secure, classified settings where you’d want to have assured access. Not for us the sudden withdrawal of Fable type models, just when we need it. No AI vassalage for the UK!
Except. While 500 million pounds sounds nice to me and Dr Evil, it’s small beer in the world of frontier AI. How small? I’ve crunched some numbers with my good friend Claude. I’m happy to vouch for them as estimates, but I’d be happier still if I’d used Fable, not Opus. The result: we are in the poo. And there’s going to be a lot more poo to go around, very quickly.
I reckon the Isambard budget is about one day of Google’s current capex spend, or about 19 days of model training costs over at OpenAI. Together, the amount of capital expenditure underway and planned by American behemoths Google, Amazon, Meta and xAI make the UK efforts look like a rounding error. Later this year, Anthropic and OpenAI are planning the mother of all IPOs - the spigots will open still further.
Horses for courses, Ken - we’re developing nat sec specific applications. Specialist models that will be crafted to particular use cases - not a gigantic generalist like Fable. Well, yes. Sadly, the story of recent years has been that the frontier wipes the floor with specialist smaller models. These may be state of the art when commissioned, but by the time they emerge, the frontier has moved on so quickly that their unique selling points are redundant. That’s what happened to IBM’s Watson. It’s what’s happening to all the wrapper companies that offer to work with law firm’s data and specialist knowledge. Do not commission a specialist model, CFOs!!! (invoice for my consultancy incoming).
And, alas, it’s what will happen to Cosine’s model running on Isambard-AI. By the time it’s built and deployed, the frontier will be at least ten times more powerful than it is today, whereas Isambard will be only a fraction as powerful as the frontier was last year. Does that matter? You bet, if you’re after the most effective cyber-security agent, but also, frankly, if you’re after anything - frontier models of 2028 will make Fable look like Pong.
What to do? Accept our fate as a wholly owned client of the United States of Anthropic (established AD 2030)? There is a slim (and becoming slimmer by the day) chance of escaping that fate. I’ll tackle that in the next post.