TL;DR: Basic economics to AI research suggests that making frontier R&D more expensive will reduce its pace, shifting our median timelines outward and lowering FOOM risk. This is probably true even if only one major geopolitical player does it.
It is a truth universally acknowledged in Econ 101 classes that an action made more expensive by some means (costlier inputs, government regulations, etc.) will be done by people less often. Yet, for some reason, people often forget they can apply this to the action of producing novel scientific or technological research - something I claim is probably still the dominant FOOM risk factor of today.
(I don't plan to back up that claim rigorously here today. For now I will simply (a) point out that figuring out good boundaries to any boundary problem is often harder than solving the problem itself, and (b) clarify that I do consider training new more advanced frontier models to be a pretty good starting point for what counts as 'novel scientific or technological research' in my definition, even though you could well argue that by this point it's just an implementation detail. If you think, however, that e.g. we're a few API wrappers or Selenium sessions away from turning GPT-o3-pro into an actual FOOM risk, well, it's certainly possible, but I find myself doubtful of that.)
Research has to be done by someone[1]. There's no a priori reason to suspect that that someone is granted protection from that Econ 101 principle just because we rightfully consider research to be a vaunted profession. History overwhelmingly supports the hypothesis that when research is very expensive, relative to other things one could do with their time, it simply doesn't get done. At the extreme end, Homo sapiens with cognitive capabilities similar to where we are today first hit the evolutionary scene about 300,000 years ago, but the majority of scientific research actively discussed today is from the last ~0.1% of that time we have spent post-Industrial Revolution.
All of this adds up to a very boring hypothesis: If the act of doing AI research become more expensive for people, then less of it will be done, and our chances of entering FOOM will be lower, or equivalently, our median AI timelines should shift outward. Boring is good! We like boring hypotheses around here.
I often hear criticism and skepticism of the "AI policy" and "Pause AI" folks, the most convincing of which I find is But what about {X hostile superpower}? We can concede that international cooperation is hard, without missing the forest for the trees here. You still win if all you manage to do is push AI timelines out by 5 years by getting one of the world's most technologically advanced superpowers to ease off the gas.
Less convincing, but still worthwhile to notice, is the asymmetry: It is very easy to argue that this action delays AI timelines, but very hard to argue that this action somehow accelerates AI timelines. You can maybe do it - game theory isn't as clean as market economics - but it requires a lot of unusual assumptions about how other nation-states will act in response. Not only would they have to respond by stepping on the gas, they would need to step on it so hard that they more than make up for the loss on a global scale.
(Students of economic history may also find considering the counterfactual of where the AI timelines would be if the European Union did not in effect "freeze" most of their own frontier model research over the last few decades enlightening. I don't think the EU did this in a particularly enlightened or targeted way -- more on that later -- but it often seems this point gets forgotten. Our timelines are probably at least [redacted, do your own thinking here] years more forgiving than they would be had they not semi-consciously bit this bullet.)
Strictly speaking we can use 'someone' in the rational agent sense here, not in the homo sapiens sense. ChatGPT's Deep Research mode would be much, much less commonly used if the software that ran it had to pay additional tithes for every website they crawled.