There's no code red on UK defence AI - there should be
Mythos and my SDR agenda revisited.
Take a look at this:
That’s right - Anthropic’s new model, Mythos is off the charts, literally. The chart in question is METR’s now-famous analysis of how good models are at completing complex coding tasks. We are into the steep part of the exponential curve now. Model capabilities are doubling every 100 days. Let that sink in, as ChatGPT is prone to saying.
What does it mean? A whole lot of trouble is brewing for the UK’s ability to defend itself, that’s what.
Not many are still parroting the line about ‘stochastic parrots’, or claiming that AI can’t reason. And yet, in most senses, it’s business as usual. There’s a fair amount of chat about AI, and even about AI and national security. But there’s nowhere near enough action.
This is an emergency of the sort that in generations past would have mobilised a concerted national effort. Let’s sober ourselves up fast - the UK is in a hole and it’s getting deeper. The Defence investment plan is so overdue it’s a grim joke, especially its ambition to accelerate innovation at ‘wartime pace’. The UK has ruinously high energy prices, and contributes next to nothing to the AI-stack, except for the brain-power of its redoubtable computer scientists. Soon, these ‘biological units’ as one entrepreneur I met recently telling called ‘humans’, might not have all that much to contribute there either. Jack Clark, co-founder of said Anthropic, thinks we are months away from models that can do their own science.1
The crisis that’s coming for us isn’t a procurement issue - at least as conventionally defined. It’s not about more ‘uncrewed aircraft’ of the sort that the head of the RAF called for recently. It’s not even about defence planning more broadly - like the lumbering process we are currently mired in, where the Treasury batters the armed forces into submission and a round of capability salami slicing ensues. This is AI that will out-think and out perform humans, including the most elite domain experts, in very short order.
Two years ago, I made a submission to the defence review. You can read it here. I called for:
A Vice Chief of the Defence Staff for AI
A second, air launched leg of the nuclear triad to counter the vulnerability of our SSBNs to AI detection.
Accelerated risk in the MoD on AI - and in particular a huge boost to government venture capital funding of startups.
A new forum is needed to bring together likeminded democracies, encompassing NATO and AUKUS, but with broader partners, like Singapore and Israel.
A citizen’s convention or a Royal Commission or discuss the implications of Artificial Super/General Intelligence, of which Mythos is the early warning.
I think these are all defensible asks today. None of them has happened. Let’s consider each:
On the leadership role: Recently, the MoD advertised for a Chief AI officer. It was advertised as a two-star civil service position. Not nothing, but not enough. There are about 470 general officers across defence. Let’s say, 50 or so 2-stars. And that’s before you get to the civil servants. The govt is having a hard time filling it, I’m told. The people who can make this happen know it can’t be done from a 2 star berth in MoD. I read once that Solly Zuckerman, Churchill’s advisor on strategic bombing in WW2, refused rank, so that the generals and Air Marshals wouldn’t be able to place him quite. If you’re not going to give this person 3, and ideally 4 star rank, this is the way. They must be the PMs personal appointment, with a chit in their pocket to get and to deliver exactly what they want. Once again, this isn’t business as usual.
Mythos is from an American company. So is the F-35. The MoD plans to get some nuclear capable F-35A to diversify our nuclear deterrent. I argued here for the restoration of an air launched nuclear leg, but also explained why it had to be independent of America. I’ve not changed my mind. I love America, and celebrate our longstanding alliance. But where we are going, alliance shades into vassal status.
Which brings me to the forum. I don’t recall Mark Carney mentioning AI in his middle-powers pitch, but he ought to have. The UK can’t afford a full stack for frontier AI. It probably can’t even afford a sovereign model, though it must try, I now believe. The alternative is dependence. In my SDR memo, I wrote about the ‘empire of the F-35’ and the very deep political dynamics that stem from being part of that circle of aircraft operators. That’s barely scratching the surface of where Mythos is taking us. Carney’s vision offers a hint of a way ahead, that balances our relationship with our supercharged superpower friends.
And lastly, Mythos is the first hint of radical economic and social changes that AI will bring. Many will be positive. And let’s not underestimate the intervening treacle that is culture - an AI revolution isn’t giving every employee a CoPilot licence and training course, but that’s certainly how a lot of bureaucracies see it, including some in national security.
And yet. Frontier AI is doubling in capability every 100 days. Where will we be when the Defence Investment Plan finally emerges? Where will we be when the MoD’s new 2 star appointment moves into Main Building? De Gaulle once argued that, ‘No country without an atom bomb could properly consider itself independent’. I feel the same about frontier AI.
Clark, incidentally, is a Brit.