The Permanent Periphery

transcript

I'm going to give a talk on something I've called the Permanent Periphery. It's both a strong claim I want to make and an investigation into a version of this claim that's a lot stronger than the basic version.

The basic version is: there are structural forces that, as more and more transformative and advanced AI systems come, asymmetrically disadvantage a lot of the currently advanced economies in the world and asymmetrically advantage the countries that actually build strong AI. I think that's fairly uncontroversial.

For the purpose of this workshop, the claim I want to investigate a little bit more is: to what extent does that lead to an actual equilibrium in which these countries are permanently, relatively disempowered — in a position where they're much less powerful, much less wealthy, much less influential on the trajectory of the world? I'm not quite sure about the second claim. I'm very sure about the first claim.

Just to motivate the very general idea, maybe start with an analogy. There's this well-known, well-trodden meme of the permanent underclass — the idea that you have a fairly calcified structure of society, and those that are disadvantaged in the current setup stay disadvantaged basically in perpetuity, because these forces lock in at the point where we get to advanced AI, and there's very little flexibility after that.

I think there are a lot of reasons to believe that this might or might not happen. But there is an analogous concept here that applies to the geopolitical international level. As opposed to your career capital being locked in at point X and then no longer being recoverable when strong AI comes around, you could think that the strategic position of countries and their economic niches and economic participation is locked in at that time — in a way that doesn't lead to an individual loss of wealth, sovereignty, and decision power, but that does lead to a collective loss of the extent to which your nation is sovereign, the ability to affect the world at a greater scale.

That said, I'm not so sure the permanent underclass is a thing. And I think understanding why the permanent underclass might not be a thing is actually helpful in figuring out why the geopolitical international version might actually be a thing.

Broadly speaking, the domestic political economy of AI happens in a system that does offer a few backstops and limits to how strictly you can get into the very permanent underclass type of system.

The first is: we just have a lot of people who still have a political voice. They take issue with the idea that they might soon be members of a permanent underclass, and I think they will exert some of that political will and power to prevent it. It's TBD to what extent that's going to be feasible. And I'm aware there are economic structural forces that pull the other way. But that does seem to be quite a plausible backstop in a functioning democracy. You could at the limit just stop these AI systems from coming into existence. You could do redistribution and pre-distribution policies. You can come up with a bunch of ways to stop the permanent underclass that are vaguely responsive to political sentiment.

The second is: we would just expect AI that is powerful enough and economically transformative enough to give rise to a permanent underclass to also have pretty outsized positive economic effects that can in theory be redistributed. It makes a permanent underclass maybe not less permanent, but at least less "underclassy" in the sense that you just get more resources to give to a lot of people. It's unclear whether they can turn that into meaningful leverage, and there might still be a disempowerment element, but it at least seems like you can get them to a pretty high level.

And then the last thing is a little more speculative, but I think you can still come up with a lot of ways in which having access to AI just empowers individuals to do new things. There are more augmentation-forward paradigms for advanced AI where people can do a lot of things because they have access to AI. Maybe that empowers them in ways that allow them to escape the permanent underclass as well.

The problem then is I don't think any of this applies to international AI policy. None of these backstops work in the same way when you think about what countries can do.

There's a lot of talk about minimizing the risks while capturing or maximizing the benefits in AI policy. I think the default outcome for a lot of countries in the world is basically exactly the opposite. They capture all the risks and they minimize all the benefits.

The reason starts with the fact that the benefits come from having access to frontier AI. Frontier AI does a lot of cool things for you — it gives you leverage, economic progress, all these positive effects on your country and your citizens. I'm not quite sure that all the countries in the world will get access to frontier AI for the next few years.

But not having frontier AI access doesn't insulate you from the bad things. All the disruption you might be worried about could still happen. Your adversaries might still have AI — they might have stolen advanced model weights, or have slightly weaker models that can still be used for dangerous activities. Your economic rivals still have advanced AI. The American economy does have frontier AI, so they're producing goods at a fraction of the price, with hyper-efficient laborers supported by much better frontier systems. Everyone you're competing with on the economic or security front probably still has AI. So you're still exposed to all the risks.

And the worst thing of all: you don't actually have anything you can do about this. If you're in the US, you have backstops from political leverage. But I'm not sure that the US government cares very much what the citizens of most middle powers think should happen in AI policy. In fact, history suggests that whenever citizens of other countries have strong ideas of what the US should do on tech policy, the US generally pushes back and does less of it. So there's a very small amount of political leverage you have.

Where you have growth effects in the US and maybe in China, you can tax some of it, redistribute some of it. But you don't necessarily have a lot of fiscal participation in frontier AI growth. If a lot of the value just accrues to the people that develop frontier chips and frontier AI, and they're all in the US, then there's no obvious scenario where you get a lot more tax revenue just because AI grows the economy a lot.

So in all the ways that the permanent underclass seems avoidable on the national level, the permanent periphery seems unavoidable on the international level because of the distinction between the international and the domestic political economy.

Where do we end up once that happens? Is there a way this incidentally bad geopolitical situation calcifies into an actual permanent distinction?

If you think through the realistic political responses to this divergence opening up, I think you can come up with a fairly likely scenario where this actually does stratify and then calcify. What does a country do when it realizes these guys have all the frontier AI, our domestic market can't compete in an international marketplace where all our competitors and adversaries have this frontier AI? A very basic instinct is to shut off your economy and try to insulate your economy and your way of life. Maybe the Europeans shut off into the European market. Maybe India shuts off into India's internal market.

The question is: if you do that, do you ever find a way back into participating in the broad-based AI gains? I'm not quite sure. There's a realistic scenario where the models commodify, things get cheaper, cheap products get built, and at some point they trickle down through the entirety of the world. But there's also a scenario where you lock down and separate yourself from the AI-driven economy outside, and you basically never have a way back in. Your economy is working at a fraction of the efficiency. It isn't outputting goods that you could realistically trade with the very AI-driven economies at anything close to a realistic trade balance. So you have an incentive to shut off further with protectionist measures and restrictive trade policy.

You probably don't end up in poverty and destitution and collapse, but you are markedly poorer than the AGI-rich economies. And you can basically never afford to open your market again, because then the competition comes crashing down — these much more efficiently produced goods start competing with your domestic economy again. And you really have a bigger problem for every year that you've shut off before. So you shut off even further, the gap grows even more, and it becomes even more risky to expose your domestic economy. At some point you just might decide: we're basically cut off from the more AI-driven economy forever.

Life then is probably not bad. You can probably keep up because there are some trickle-down effects from AGI-driven wealth and growth. You have fairly functioning economies there right now. You have decent education, decent natural resources in a lot of these countries. They can probably become self-sufficient, maybe trade with each other, and still have a decent quality of life.

But it's not as good as we can do for these countries. And on a deeper civilizational equilibria-level consideration, it doesn't seem great. If there's just a bunch of countries that are basically 2026 forever while the rest of the world is living in Paul's talk — there's a huge gap in agency between people that live in Paul's world and people that live in my world. That doesn't seem like a great outcome for the post-AGI civilization, that there's just a bunch of people who via the birth lottery end up in countries that don't have a realistic way in.

I also think it might not even be stable. You can have backsliding and destabilizing effects from having this permanent periphery. If you live in a peripheral country, you might do your best to migrate out of it — your brightest talent might try to move to the AGI countries where they can have a lot of effect. States get into real legitimacy issues if they feel they can't provide participation in this big technological trend. And if some countries in unstable regions get access to slightly more AI capabilities than others, then there are a bunch of ways that could re-trigger dormant conflicts — as always happens when there's asymmetric technical diffusion of advanced technologies throughout conflict-prone regions.

So I think we should do something about this. It's just extremely unclear what we should actually do, because every measure you could come up with basically collapses down to: there is no galaxy-brain narrow AI policy intervention that fixes any of this. The way you stay competitive in a global economy is you have a decent amount of economic growth, you lean into global trade, you can trade with the AGI-producing countries, and you have a stable strategic position. That's really hard to do, and really beyond the scope of any narrow AI policy intervention.

I think there are still some things. The first is just: buy compute. If the AI models run on compute that you own, you have marginally more access. You have leverage securing that access because the labs can't really choose not to give you the models — if they don't give them to you, they can't give them to anyone, because you'll just turn the data center off. That gives you some leverage to get access to frontier AI. You can probably also tax around the data centers a little bit so you get some revenue.

Then there's an analogy to the labor policy question of what will be scarce in the post-AGI world. You can think about that on a national level: what kinds of bottlenecks to what kinds of supply chains will become really important if AI becomes the most important product in the world? Some upstream stuff like semiconductor manufacturing equipment, some downstream stuff that can turn data centers full of genius into actual real-world goods — industry capacity, biotech, R&D, manufacturing, pharma. There's a bunch of things you could strategically think about that will be valuable in the future. If you occupy them strongly enough and make them essential enough to the value chain, you probably do get some marginal leverage and economic participation.

But really, I think it still collapses down to: go through, not out. Do some version of effective 1995-style neoliberalism — opening trade barriers, getting to a point of decent economic growth and decent alliance management, and then try to take it from there. Which is an unsatisfying takeaway. But that's the only takeaway I have.

Thanks.