Forecasts of the US frontier models' progress in the original AI 2027 are proving true, but their prediction that Chinese models will be right behind (~1-3 months) is proving false (at least for now).
This means the race between the US and China is not that tight, and that the US can be more comfortable slowing down. e.g., a short slowdown (e.g. 1 month) on the part of the US does not risk losing the lead to China.
Hence, the initial phases of a slowdown do not necessarily require international coordination. We just need Anthropic and OpenAI (and Google) to coordinate.
EDIT: Seems I was wrong about the China gap prediction from AI 2027 (see StanislavKrym's reply below). However, I believe the main point is that US frontier labs can do a calculated pause without significantly risking handing the lead to China.
The AI-2027 scenario doesn't actually predict that. The section on mid-2026 had the authors write: "A few standouts like DeepCent do very impressive work with limited compute, but the compute deficit limits what they can achieve without government support, and they are about six (italics mine -- S.K.) months behind the best OpenBrain models" At the time of writing, the ECI of Kimi K3.1 Pro, released on October 23, is 162, which was first exceeded by GPT-5.5 Pro on April 23. However, the scenario did have China become a couple of months behind since the theft of Agent-2 and subsequent progress being slower due to compute issues.
Thanks for pointing this out. I mistook the later parts of the timeline for the earlier parts. I have added edited the post and pointed this out.
That said, I believe the main point still holds: IRL, US frontier labs have more liberty to pause/pace than is assumed when it comes to the risks of handing China the lead.
This framing strikes me as having an underlying attitude that most likely will not be shared by people working from a national security mindset.
It seems to me that competent defense must entertain worst case scenarios, and so is effectively dealing not with a simple estimate of the progress differential, but rather with a lower bound on that differential. So even if it seems like 6 months behind is our best estimate, a security mindset would say that we should assume that the gap is closer than it appears, to account for the considerable uncertainty in the estimation (not least because of the adversarial situation).
I think that arguing for a unilateral pause will have to reckon with the more stringent requirements that this implies, and supporters should expect to confront the lower bound rather than the estimate. In addition, the risk of gap-closing actions like the weight theft in AI 2027 is not something that can be ignored either, and will also be weighed as a factor in the security analysis of pausing, since pausing increases the exposure time to such actions.
The race, IMO, is not purely about capability (although for AI companies that is indeed a, if not the, primary goal), but also about maintaining the (much larger and much less sustainable unless rapid gains are provided soon) investments from the US market into the race. In spite of Chinese models falling behind on ECI, they are growing rapidly in popularity, and when factoring in cost, they are only about 7 ECI points behind equal-cost US models. Anthropic's IPO has already fallen from 2 trillion dollars expected before the IPO, down to ~1.4 trillion now on the eve of it; with strong uncertainties about the profitability and safety of the stocks with very weak fundamentals. It is driven by market forces, in effect, to release a model that's only 2 ECI points behind its frontier model 1.5 months ago (Model 4); and yet the impact of Fable 5.2 has been somewhat muted in rallying Anthropic to a higher valuation.
All to say that unless the race is directly funded by the government, then the race is tight because US companies must maintain a comfortable lead to maintain market confidence; and the race is funded by the market, which might think that a 6-7 month lead is an incredibly small moat (and not entirely wrong: if AI capability growth is an S-curve at some point, Anthropic's moat is gone in a tiny amount of time), and which sees AGI companies as having very weak fundamentals. Therefore, a large hard-capability lead is necessary to challenge these market doubts; as large as possible would be good.
Of course, for national security or economic stability reasons, the US government might intervene to take over the companies and save the industry; but it'd be concerning, to say the least, for a government already running somewhat hot with growing public debt, the Iran war, and growing public dissatisfaction over affordability, to spend trillions covering what the market should be doing. (And of course most companies don't want to be taken over by the government, so.)
Thanks for sharing this perspective!
AI 2027 predicts that China will be right behind only after they successfully manage to steal US weights in the scenario.
An idea: AGI lab employees go on strike for a period to demand progress for a pause.
And not just AGI lab employees; ideally all technical AGI researchers/engineers, including those in institutions (government, NGOs, etc.) or independent, would strike at the same time. These people don’t build AGI directly, but they operate within the ecosystem and machinery that lets these companies go full speed.
The period can be 2 weeks to a month, basically a timeframe that makes some meaningful progress possible.
The window to act is closing rapidly, and the next 6 to 12 months might be our last chance.
The prevailing view among technical AI safety people (and many of those I talk to) is that alignment is not on track to ensure safe progress in capabilities, and that we have nothing in sight of scaling to superintelligence. This means alignment research likely won’t make it in time to save us if we continue on our current trajectory.
The current political landscape suggests governments aren’t interested in any kind of slowdown, and AGI lab leaders are unlikely to overcome the intense commercial incentives to slow down.
But we technical AI researchers, especially those working within AGI labs, have more agency and say in this than we think. We are the people on the frontier working on it, so we understand the risks best, and we could mandate a pause by boycotting technical work for a short while.
The strike would be most effective if it includes people across the board, especially those working on capabilities, but even if just those working on safety and alignment strike, it would have considerable impact since alignment is a bottleneck for capabilities.
The strike would need to be coordinated and get buy-in from all major companies, including those in the US and China (though I am not sure how difficult it would be to get Chinese employees on board). If CEOs and countries can not coordinate a pause, a few thousand technical researchers could coordinate a strike and mandate it. The aim of the strike would be to make meaningful progress toward an international framework for a pause with clear objectives. If governments and CEOs fail to make progress, they would risk further downstream strikes.
During this strike, engineers and researchers would focus all their effort on non-technical work, such as policy, lobbying, and advocacy, to make maximum progress toward achieving these objectives. This would provide a considerable push to a pause.
This will definitely involve personal risk to people, but I believe the fact that this is 1) collective, i.e. inter-lab and hopefully international, 2) CEOs already recognise the need for slowing down, and 3) is an action done in good faith will significantly lower the personal risk to any given employee.
To be clear: I don’t claim this would be easy to pull off with no risk. But I believe employee strikes are an important lever and, from my point of view, might be worth exercising if other efforts fail and while strikes are still effective, as we are not yet in a fully automated R&D regime (but soon will be, and might lose this lever forever).
I’m sharing this as an idea rather than a confident proposal, and I’d like to hear where it breaks. What am I missing?