Beg, borrow and steal

A British strategy for the post-Mythos world

A very long time ago in AI history (last week in human time), I argued that the UK is in the soup because we lack access to Mythos, with no guarantee that the American government will give it to us. ‘Tough shit’ as one very well-informed American insider told me subsequently. How important are we in the discussions that go on in the White House about all this, I asked my new friend. ‘Not at all’. Time to sober up.

But not for nothing does it say ‘strategy’ in my job title. There are a few options for the UK government, and we should explore some of them, quickly. Read on:

We need it, and it must have a union flag on it

As I see it, we have a few options, not necessarily mutually exclusive. These are:

• Embed ourselves in the American stack. Ryan Fedasiuk argued this a few months ago. There are some things the UK and EU can do that will make us useful allies and partners for America. ASML and their top-notch chip carving machinery is the cleanest example, but we also do well at other things, like wind turbines. As for the UK, well, we’ve the vaunted AI Security Institute (AISI), world leaders in assessing the performance of these models.

  
  My verdict: Maybe. I like the theory of comparative advantage as much as the next economist. But alas I don’t think our USP is very U. AISI in particular strikes me as a new version of the complacent British habit of aspiring to be Athens to America’s Rome. With our unique wisdom, we hope to overcome the harsh fact, so apparent to my American interlocutor, that we don’t make models. Still, the weak do what they must, don’t they?
  

• Build it in a coalition of middle powers. Let’s call this Project Carney, after the Canadian premier who issued the call to arms earlier this year. If you add up non-US NATO, EU countries, non-US 5-eyes, and throw in a few odds ’n’ ends, like Singapore, you ought to have the economic clout to build a proper frontier model. There’s just no way of getting there on our own. Mistral, the French effort to keep pace is at least a couple of tiers behind, and their total capital raised is more or less what a frontier company spends in a quarter. Collaboration is the only way to get there - and needs must.

  
  My verdict: Fat chance. This is the mother of all coordination problems, and the record of national security cooperation suggests it won’t come close to happening. Germany and France just folded their tent on their 6th generation fighter project because they couldn’t agree terms. National champions are still very much the order of the day. And the clock is ticking - the frontier is doubling in performance every hundred days. Can you imagine how long the negotiations for this sort of thing would take? Can you imagine the conferences in exotic locations? We’d end up with some sort of Heath Robinson contraption - yesterday’s technology, delivered tomorrow.
  

• Steal it. The one thing we are genuinely world class at. We have first rate intelligence agencies, who do a sterling job of pinching info that might be to our advantage. Also working in our favour: the frontier companies are getting better at securing their secrets, but are still some distance from Fort Knox. They’re not a soft target, maybe, but it’s not like hacking the NSA either. And look, if China can obtain OpenAI’s knowhow and magic up DeepSeek out of nowhere, it can’t be all that hard. We should be able get the weights and the source code. Once you have those, you don’t need the massive GPU arrays to train them. Just tweak and deploy - sorted! Let’s call our knock off Claude something catchy, Alan, maybe.

  
  Verdict: We don’t spy on the Americans - this goes against the very powerful grain of many decades of the closest intelligence cooperation. And I’m a massive America-phile too, so this goes against my personal grain too. But then our closest ally has threatened to invade Denmark and absorb Canada within the US. And this technology is going to transform economic and social life - the stakes couldn’t be higher. I wonder if some folks in the British establishment are feeling the tectonic plates shifting under their feet. Could it be time to channel Palmerston - no permanent friends, just permanent interests? But there’s a problem: emulation doesn’t get you to the frontier, not at the pace it’s moving right now.
  

• Hope there’s a cheaper way to get there - ingenuity to the rescue! This is another classic British tactic: muddle through and rely on our top boffins. We’re certainly betting on this approach: the UK government is putting tens of millions into labs working on alternative architectures. That’s in addition to the recent announcement of serious public-private investment in building a sovereign UK transformer model for national security applications.

  
  My verdict: This is the way. But we’re going to need to spend a lot more on it than tens of millions. There is a lively debate among specialists about whether the vast spending on transformer architectures will deliver super-intelligence. I’m in the camp that it probably can: LLMs plus coding plus tools will get us there. But even if it can, it’s not clear that it can do so before even the very deep pockets of American private capital are exhausted. And it’s not clear that going public will help either - let’s see what happens when Anthropic and OpenAI have their IPOs later this year. Maybe a British bolt from the blue could deliver the spoils. Maybe - I’m an optimist, but I’d not bet on it. And as the frontier gets closer to ASI, remember, the models will start innovating themselves.
  

• Accept vassalage. I still don’t think folks have fully grasped the implications of this technology for economic performance and hard power. Most of the discussion of AI is about using chatbots at work. But Copilot and a mandatory training course do not an AI revolution make. ASI won’t retool your existing workforce. It will radically outperform them and replace them. Or not, if you don’t have it. America halted Fable’s rollout because it was worried about cyber vulnerabilities. There are massive implications from that capability alone. As a friend told me - if American companies understand those vulnerabilities, they can price them in and mitigate the risk. Rivals cannot. But that’s only the thin end of the wedge.
  
  My verdict: All good things come to an end, and maybe that’s true for us too. The empire had a good run. There could be worse endings, after all, America has air conditioning aplenty. But those are the stakes - if America controls access to post-Mythos class models, we face the prospect of being outcompeted across great swathes of economic activity. I’m not so keen on that. You?