America First
There's only one country in the world that can produce a frontier model. Could the US pace AI progress unilaterally?
There is an ongoing conversation, highlighted by the Pacing the Frontier letter, about potentially slowing the speed of AI progress to allow alignment research (and maybe the economy broadly) to catch up. One challenge with this idea is that it could require coordinating policy globally—or at least with China, the only other country with a serious AI sector.
Global coordination is famously hard and slow, and people aware of this might be inclined not to try at all. But the need for coordination is premised on the idea that, without it, China might overtake the US in capabilities. Other arguments—sometimes by the same people—suggest that the progress of Chinese models has been driven by cribbing from public US models.
Superficially these views seem to be in tension—if Chinese progress is indexed to US progress, then pacing the capabilities of US models alone might be enough to slow progress globally. No treaty required. Is that true?
First, the vibes.
Wild Pause
There are several potent ideas swirling in the aether at the moment. Moonshot released the weights for Kimi K3; various open letters have been drafted and signed; sandboxes have been breached and prototype models have been turned off, buried in concrete, hosed down and encrypted. Margins have been called. The market speaks. Competition with China, regulatory capture, coordinated pacing, pleading slogans written in chalk on OpenAI’s sidewalk that contain, perhaps, more than the optimum number of syllables.
I challenge anyone to write coherently about AI right now without doing the textual equivalent of pinning 30 documents to a cork board and connecting them together with red string. It’s enough to make you feel… surrounded.
This frenetic speed combined with the sense that the Hugging Face incident was a warning shot has elevated to new ubiquity the idea of coordinated pacing of AI progress. (Theo won’t let me say the word “slowdown”. They’re pacing it, like a marathon. He’s right!) The Pacing the Frontier letter is one incarnation of this, and I wrote yesterday about how we might interpret the intention of the people behind it. But, as with many things, the sentiment shift was felt first on x dot come, the everything app.
My Way or the Highway
I find myself very persuaded by the arguments made by many of the signatories. Even if you don’t agree with the idea of an immediate global slowdown in AI progress—I put myself in this camp—there is a strong case for building the tools before you need them. Not only would this leave us more prepared, it would also give us more time to refine an approach that avoids the pitfalls that might attach to more rushed regulatory approaches, like the concentration of market power in a small number of firms.
But there are reasons to think that even a more deliberate approach to pacing the frontier would be undermined by the challenges of global cooperation. I wrote about some of them yesterday, using the WTO as an example of just how dysfunctional multilateralism can be. A bilateral treaty directly with China would dispense with some of these challenges, but it would also bring new challenges of its own. Chinese leaders, for example, have publicly taken a preemptive stand against “new historical injustices” in AI.
But, if the leaders of US AI companies are to be believed, the AI sectors in the US and China are not simply mirror images of one another. There is one (simplified) view of AI progress that says that US labs do research that advances the frontier of AI capabilities, and Chinese labs distill US model output and use this to release fast-follow open-source, open-weight models. If that is the case, then it strikes me that the idea of a “capability race” is somewhat misleading. Chinese labs might be racing to keep up, but, if they are dependent on frontier models for progress, that doesn’t imply that they will be able to take over if the progress of public US models slows down.
You First
First, what evidence do we have that the progress of Chinese AI capabilities is linked to US progress? Mostly, I have a few interesting tweets.
It’s not just Kratsios and Anthropic, of course. Kimi K3 is definitely borrowing something from US models.
That’s not to say that training or RL on US model outputs explains all of Kimi K3’s capabilities. Chinese labs have pioneered several techniques that have become central to model architecture in US labs as well. They are clearly extremely talented teams, operating with fewer resources and less compute and producing models that lag the frontier by only months. But how exactly could you quantify the relative contribution of these factors? In the immediate aftermath of the Moonshot release, independent AI researcher and writer Gavin Leech shared his rough estimate:
If something like this is true—and I don’t know how confident Gavin himself would be in these estimates—then it seems very plausible that tools and mechanisms for coordinating the pacing of US AI capabilities would also likely slow progress in open models from the rest of the world.
Add to this the idea that slowing the pace at which new capabilities are made public does not necessarily imply a radically slower pace of internal progress, and you can imagine a world in which coordination between US labs is enough to slow frontier capabilities while also preserving the lead held by US companies today.
This is, of course, all speculation. And not speculation that is particularly empirically grounded. But perhaps it could be. Here, then, is another thing to add to the list of “AI metrics it would be good to have”: what proportion of any given release by a non-US lab is downstream of the public capabilities of US models?
And if that metric starts to look relatively strong, perhaps we should consider whether US unilateral action might be nearly as effective—but much faster—than global action.
