AGI-based Central Planning

transcript

For this one, the thing I'm worried about less — or focusing less on — is economic growth, but what's going to happen to power and concentration of wealth and then ultimately political power. And so this is a scenario where some of the barriers that prevented that from happening start fading away because of the power of technology.

Just to be clear, this is not a recommendation or a preferred path. It's not an inevitable path. And one of the things I want to do is leave some time at the end for all of you to say something about what are some ways that we can prevent this from happening. But I'm going to lay out some of the economics of why it is increasingly possible and even likely if we don't change the path.

One principle I'm going to focus on particularly is this idea that in general, you want to co-locate the knowledge of how to make a decision with the decision rights of who gets to make that decision. It doesn't always happen, but a lot of organizational design is around trying to bring those two together. And you can either move the knowledge or you can move the decision rights. And a particular aspect of that in turn is centralized versus decentralized decision-making — whether decisions are made in the center or on the periphery.

So there are some trade-offs. The paper has lots of economic theory about this, but it basically boils down to the idea that centralization is better when decisions are highly interdependent. You could have an optimal toothpaste tube and an optimal cap for it, but you'd like the two to fit together. So maybe coordinating that would be a wise thing. Or if you're driving a truck, you want to know maybe what some of the other places are that you'll be delivering stuff. All these interdependencies lead to some benefits to coordinating.

Hayek argued — and this is the title of this talk, "AI's Use of Knowledge" is a play on his most famous paper called "The Use of Knowledge in Society" — he argued that the really important information for making decisions was inherently local. That truck driver, when they're driving around, they realize that, hey, I didn't need to complete all my delivery at the last shop. They didn't take everything. So I have some stuff left over. Maybe I'll bring some of it to this other guy that wasn't on my route, and then I don't have to come tomorrow. There's all sorts of knowledge of particular time and space, as he called it. And that kind of knowledge is inherently decentralized, distributed out through our organization. He argued it just couldn't be aggregated. It couldn't be centralized. And so you needed to push all the important decisions out to the periphery where that information was. A lot of it was dispersed, local, and tacit.

We heard yesterday about some of the ways that it could also improve markets and make agents more effective. So both these things are happening at the same time. Let's go into a little more detail about how AI can codify local knowledge and increase information processing.

You could have an allocation where the entrepreneur owns both the information and owns the truck — that's like a local owner of a company. Or you could have the headquarters own the truck. But by assumption, they can't own the information — that is, as Hayek says, stuck to be local. So basically Hayek's conclusion is that you're better off if the entrepreneur has the knowledge — they should also own the other assets and we should distribute things out to them.

But transformative AI makes it possible to move that information. And so now at least you have the option of having headquarters own both — you could move the truck and the information to headquarters. And now it depends on which types of decisions are more important — the local decisions or the central decisions. At least you have the option that you didn't have before.

Another aspect that's changing is the information processing limits. Our brains can't be overly large. We don't have that much brain capacity. So right now, if headquarters is doing some information processing and these other folks are also doing some information processing, if you have bounded information processing capacity, you can't have one entity make all of the decisions. But again, TAI makes it possible to successively move more and more information to the center and you don't overload headquarters. The folks in Bentonville, Arkansas can have all sorts of data about all the local shops — how much toothpaste there is in each shop, what the preferences are for peppermint versus spearmint. And it doesn't overwhelm them. They have a giant computer that computes all of this and considers all the demand curves and supply options. So that's now possible in a way it wasn't in Hayek's time.

Now, there are some countervailing forces. One is that I'm assuming that you can do all this information processing and you have all this knowledge. A lot of people push back on that. You guys aren't going to push back on that because it sort of goes against the assumption of the workshop. So we'll just say that, yeah, you can have really powerful AI that does these things.

Another one is that even if you did have AI making all the decisions, you'd still need some level of decentralization. There's the speed of light, so if you're making decisions on the Moon or in a faraway galaxy, you can't wait for people on Earth or headquarters to be making all those decisions. The Mars rover had to have some level of autonomy. I think that's fair to say, but it's still far more potential for centralization than we have today, even if it's not complete.

And then the third one is where I think I want you guys to focus the most — are there some other laws or legislative requirements that would prevent the AI-powered firms or the AI-powered government from making all of the decisions? Could we legislate some limits on that level of centralization?

The early empirical evidence — indeed, there is a lot more centralization. Bentonville, Arkansas is where Walmart aggregates all this information from all over the country, now all over the world, about what's happening in the local markets and makes all sorts of decisions. And apparently they're making them pretty well because they're winning more and more market share. If you look at the data — I'll focus on the lower left — that's what's happening in retailing. The market share of the biggest firms has been growing quite steadily from about 20 percent to about 50 percent, even faster in sales than in employment. In other industries, you also have more and more concentration.

Although we're not yet at AGI, depending on how you measure it, there have been technologies that have been leading to more and more centralization of the economy over time. We also see a lot of private equity firms doing roll-ups of small firms. They're rolling up funeral homes and accounting firms and lots of others into more aggregate. They believe that having more centralized decision-making makes them more efficient and they seem to be making money on doing that.

One political implication — I'm not a political scientist, I think there's a few of them here in the audience — but I think there's a lot of evidence that having more concentrated economic power tends to lead to more concentrated political power. If a small number of people or entities have the vast majority of the economic power, that can distort the politics in such a way it'd be hard to maintain decentralized political power. Maybe not impossible, but something that we need to think about.

Another implication is that this can become a recursive or reflexive effect where you get more and more feedback as you have more powerful AI systems. There's less incentive to invest in human capital. Humans are not making as many decisions. It's not as valuable. And that leads to humans being more and more disempowered and even further centralization over time. So you end up getting more powerful forces heading in that direction.

So just to summarize, we have this tension between centralization and decentralization. It's something that's been in economics for centuries, but especially the past century. Hayek predicted that the decentralized system would win. There was this so-called socialist calculation debate, and most people think that he basically won that debate — the Soviet Union fell and highly centralized economies did not succeed.

But I also want to stress that this is not inevitable. It just demands a new orientation. We can't just rely on decentralization being by default the only option for how you organize the economy. Now this centralized option could be very efficient and we could let it rip and see how things roll out. Or we can think about ways that we want to try to push back against it.

I'll just leave you with whether there are some ways that we can improve the likelihood of maintaining or increasing decentralized economic power and therefore decentralized political power. Thanks very much.