Cliff Sosin：持有智能的互补要素——[Invest Like the Best，第497期]

You called me, saying, “It’s just intelligence,” and that was sort of the theme of our conversation. Describe that simple idea.

There’s this idea that intelligence starts in an amoeba and gets better and better, and then it’s like a monkey, then a person, then Einstein, and then it goes zoom and becomes God. I think that’s something of an intellectual mistake. What people have wrong is that they’re missing the real limits of what intelligence can do.

Another way to put it: if you’re thinking about intelligence as an unbounded scalar, the height of a tree isn’t correct. What you should really think about intelligence as is more like a ratio, like the roundness of a ball. Intelligence is a gadget, and we use it to make predictions about the universe. We start with this thin scattering of facts about the universe, where we have relatively few of them, and the universe is complex, so we use intelligence to interpolate between them.

As your intelligence goes up, your ability to color in the spaces gets better. Some domains are really amenable to that. In mathematics, you can solve things, and in those areas, you get to perfect. But for most of the interesting stuff—the questions that make up most of what we want to know about the world—the mathematical term is that they’re complex dynamical systems.

What that basically means is that they’re kind of like weather, in the sense that you can’t just guess your way to an answer. You can’t make a closed-form answer. You can try to simulate it, but your simulation will invariably have some degree of error in its initial conditions or in each step going forward, because, at a minimum, your steps are going to be truncated. That error builds, and invariably, after the simulation runs for some period of time, you’ll be off by some amount.

So, for mixing cream into your coffee, predicting the weather, or thinking about how groups of people, who are all very complicated, are going to interact with each other, intelligence just isn’t that powerful. My sense is that you get increasing intelligence to the point where you’re making all the predictions something super smart could make. It turns out that’s a relatively small set of all the interesting things you’d want to know about the world. To get past that, you need to go out and get more information about the world, and that just takes time. You have to do experiments. You have to touch reality.

One interesting corollary is that one of the things you can turn your intelligence toward is reducing the cost of intelligence, which we’re seeing. So what you actually end up getting is everyone gets unlimited, really great, very round-ball, so to speak, intelligence for essentially free. Obviously, that’s a limit statement; it won’t actually be free. But the point is that really good-enough intelligence is rapidly falling in price. And it turns out really good-enough intelligence will solve the vast majority of questions where people need to use intelligence.

Yeah. Can you say more about why this notion of a complex, adaptive, or dynamic system means that even perfect intelligence—or whatever intelligence we have in a bottle ten years from now, or pick your horizon—will still be limited in its applications?

Yes. Nonlinear dynamic systems. The famous story about how these were discovered, and how chaos theory came into the branch of mathematics, goes back to the days when there was a guy building a weather forecasting model on big computers that took up rooms. He ran the model, and it made predictions of pressure and temperature over some forecasting period. He thought, “Cool.” Then, just to make sure it was still working, he reset it and ran it again. It started out with basically the same results, but then the results started to diverge. By the third, fourth, or fifth day of the forecast, the results were entirely different.

The guy thought some version of, “Darn, I must have a bug in my system,” and began trying to figure out why the computer wasn’t producing the same results in repeated runs. It turned out that the technician who had loaded the data into the computer for the second run had rounded off at one fewer decimal place—

Mm.

—on all the initial conditions for pressure and temperature and such. You would have thought that maybe this would have produced a small but consistent error over the whole forecast period. But what had actually happened was that the small error compounded through all the repeated steps: you take your current conditions, calculate your next set of conditions, and the errors compound. After a short distance into the forecasting period, the forecast was entirely different and the weather unrecognizable. One model said it was 70 degrees and sunny, and the other said it was 20 degrees and pouring.

People called it the butterfly effect because, at least in principle, a butterfly could flap its wings in Africa and that could lead to a hurricane in North America or whatnot. It turns out that many systems we’re interested in have this characteristic. You can know everything about pressure and temperature, but you can’t necessarily write down an equation that you can easily solve to tell you there’s going to be a hurricane. You have to model it, and the model, in turn, is going to have some level of granularity unless you actually take the step of having the whole universe.

Ultimately, in the limit, if you want to simulate the whole universe, what you actually need is the whole universe. If you don’t do that, there are many circumstances where simulation is super useful for approximations, and in those circumstances, you can use intelligence to solve them. But there are many more where it doesn’t work.

Yeah.

And so intelligence is limited by that fundamental unknowableness, I guess, about the universe.

So, back to this notion of “it’s just intelligence.”

Yeah.

What does that mean about where you think we’re going from a more practical standpoint?

We’re all going to have a team of 200-IQ people working full-time in our pocket, and that team will have a team of many more that it can deploy. They’re going to be trying to figure out all the little optimizations that we’re too lazy to do. Every company will make similar teams available to their employees to help them do their jobs.

There are just a lot of businesses where that’s not good for them. Consumers are boundedly rational at present. They only have so much cognitive space. You go to a consumer and say, “Look, we’ll take your money, we won’t steal it, we’ll pay you a decent interest rate, you don’t have to think about it, and you can go about your life.” The consumer doesn’t say, “Look, if I think about this every day, I could save $400 a year, so therefore I’m going to think about this every day.” The consumer says something like, “Eh, sounds good enough, and I’m going to move on.” But with an agent, you would think the agent will continuously optimize for them.

There are also companies where what they produce is fundamentally some sort of artifact that is just crystallized intelligence. I’m by far not the first person to realize that AI could be bad for certain software companies. But the idea is that an AI model is kind of like a general-purpose intelligence tool, and many software tools are crystallized and hardened pieces, like projections of intelligence.

As I’ve been thinking about companies and investments, I’ve found it helpful to assume everyone has really great intelligence and ask what’s left. The pushback is often some version of, “But you don’t understand how hard this is, and you don’t understand how complicated this is, and you don’t understand how many smart people have had to think for so long about this.” My hand-wavy and approximately right response is, “Okay, sure, but that’s all just intelligence.” To follow this logic all the way through to the end, if intelligence becomes cheap and ubiquitous, what you want to own are the complements to intelligence.

The complements to intelligence are data, attention, love, affection—whatever sort of emotional attachment you want. It’s physical proximity, and it’s scale. It’s a little cynical, but it’s regulatory capture.

For some businesses, their ability to deploy intelligence is finite, whether because of the economics and the cost, because they’re only a certain size company and don’t have access to 10,000 of Google’s best engineers for 10 years, or because the organization is behind. The ability for you to deploy way more intelligence in your business, for your competitors to deploy more intelligence in theirs, and for your customers to use lots of intelligence can work out to be better for you. The lack of intelligence can be bad for you if intelligence is a barrier to serving your customers. Or it can be bad for you if your customer’s lethargy is the fundamental element of your economic value.

Maybe pick Meta or something as the perfect example of a company that you think owns perfect complements to—

Sure.

—to ubiquitous, cheap intelligence, just to bring the example to life.

I’m actually going to set aside Muse and agents and that whole thing for a moment—

Okay.

—and focus on business. The core business at Meta is personalized TV.

Personalized TV.

It used to be social media, but now it’s personalized TV. It turns out that people really like watching Instagram discover that they love dark British humor. It’s great.

People talk about the 10,000-foot view. Let’s talk about Meta from space, from the highest view. They have $250 billion of revenue, and they spend $250 billion a year. Some of it is accounted for as costs, some as CapEx. It doesn’t really matter; they spend it all. The question you might ask is: What do they spend it on? A really reductionist view is that they spend it all on intelligence.

What do I mean by that? They’re buying data centers. What is a data center? It’s kind of an intelligence-producing tool. What else are they doing? They’re literally buying inference. That’s just intelligence. What else are they doing? They’re paying really smart people to write algorithms. That’s intelligence. They’re writing code, also intelligence. They’re serving videos. That’s intelligence-adjacent. Sure, it’s intelligence.

In some highly reductionist sense—and you shouldn’t take these things literally; I’m just trying to highlight the really big directions—all they buy is intelligence. If you think forward with the fact that the price of intelligence is falling by a factor of 100 or something over the next 5 years, first principles suggest their costs kind of go away. And that’s just thinking about the cost side of things.

On the revenue side, there are all of the creators. You are a creator. How much more quality creative can every creator who makes content for Meta make if they have access to intelligence that’s 1/100 the cost and 10,000 times more effective? The content is probably better. Then Meta, in turn, maybe doesn’t take its costs to zero. Maybe it spends more than zero and does a far better job of picking the videos to show people next, which should make the content better targeted and more enjoyable.

On the ad side, advertisers could make far better-designed, far better-targeted personalized ads that Meta can then target much better. Advertisers can also build much more slippery funnels or whatever to convert. The whole point of all this is that maybe, instead of taking their costs to zero with much better intelligence at much lower prices, they can sell a few more ads.

All of this is downstream of what they possess, which isn’t destroyed by ubiquitous intelligence: data and attention. They have an enormous amount of data about all these people and how they view videos and whatnot, and an enormous amount of attention. These things compound each other. I don’t think the availability of cheap and ubiquitous intelligence is going to cause people to stop being interested in watching funny 20-second videos. If anything, they can win some time at the margin from other things people do with their time.

I don’t actually own it, but a company that benefits in theory from having a lot more intelligence is Grindr. Basically, Grindr is the gay dating app, but it has been constrained over its life by the fact that it’s just not that big a company. It turns out they’ve got lots of attention and lots of data, but their ability to build all the things you could imagine building—on what ultimately amounts to a finite-size team and budget, and a finite ability to create things—has meant that, in some ways, Grindr can’t be all the things you could imagine it could be.

Again, I don’t actually own any Grindr. I study lots of companies, and it just occurred to me that this is another example of something where the mental model is: What if you had 10,000 of Google’s best engineers to go work on your problem for 5 years? What could you do? Grindr is one where maybe they could do more.

1. The Frontier Labs Face Commoditization

I want to talk about the labs. Back to the original point, I think lots of great investors would say intelligence, specifically frontier intelligence, is not a commodity. Historically, the conversation has been about the gap in intelligence between closed source and open source. Increasingly, the conversation is about the gap between what’s publicly available from the labs and what’s privately used by them inside their own businesses. They seem to have figured out how to stop or slow distillation. Lots of people would say there’s this escape-velocity thing happening, and that the frontier token is going to be way more valuable and useful, and that the labs themselves can then distill their own models and provide the best version of cheap intelligence, too.

I’m curious for you to riff on the labs as companies through the mind of an investor.

If they have models that are better, and the ability to produce adequate intelligence at way lower cost than everybody else, and nobody can copy it, then it’s going to be a great business. I guess I have a couple of thoughts.

As I’ve watched models get deployed, one way to think about it is that you start out with these very inadequate models that are basically able to do a very small subset of jobs. The models get better, and they work their way up the stack. With each successive step up, the frontier model has unlocked a larger, call it, addressable market. But behind it, what’s happening is that once the evals or whatever get saturated, you would never use a frontier model for a chatbot in a customer service application. You’d be an idiot. Those areas clearly commoditize. There’s a wave of hard tasks where you really need a frontier model, and then behind that, you don’t.

The friends I have who are big bulls on the frontier-lab companies—and obviously these companies’ financials aren’t public, and they might have an information advantage over me—would say something like, “Don’t worry, we can also, along the whole efficient frontier, produce cheaper tokens.” Maybe they can right now, maybe they can’t right now. Maybe they will in a few years, maybe they won’t. I find this very hard to think about.

Certainly, the trend has been rapid commoditization of these things, if you look at the rise of Chinese models at every step. At a minimum, the cost has been collapsing. So even if they do have a position where the token is necessary to do tasks that are well behind the frontier, it might not be a very big market because it just doesn’t cost that much in the scheme of things to deliver. And it is competitive, even if they are a cost leader in a shrinking commodity market.

The other thing is that I think programming was this huge use of models. As you push forward, the question becomes: For the next, next, next layers of intelligence, do you still have bigger and bigger places to go after? The bull story would be some version of, “Yes, people will pay anything for a model that can figure out how to cure cancer.” Maybe that’s true, but to go to our prior conversation, I think figuring out how to cure cancer requires touching reality, and I’m not sure that more IQ gets you that.

If you don’t believe there are some super-incredible use cases where you can just charge an enormous amount for more intelligence, then you start getting to the point where the next step up—you go coding and then maybe something else—but you get into smaller niches. It’s like this diamond-shaped demand frontier where initially you get wider and wider uses, and then it actually starts to get more and more rarefied. How often do you really need a 300 IQ, one might start to ask?

We’re going to find out, and I don’t want to be too dogmatic, but it strikes me as totally reasonable that what we end up with is some version where there are very few use cases for super-elite intelligence. There are lots of use cases for very good intelligence. That very good intelligence is highly competitive and commoditized. If these frontier labs can somehow become cost leaders in doing that, there’s a margin to make in being a cost leader. But then there’s also just rampant deflation. This question becomes: How big is the market if good-enough intelligence is very, very cheap? How much good-enough intelligence are people going to buy?

I’ll just keep playing the devil’s advocate.

Yeah.

2. Demand Meets Physical Limits

So the 2 things people might say in this case are Jevons paradox, that—

Sure.

As it gets cheaper, we’re just going to want to use more of it, so I’m curious what you think about that. The second point would be that, historically, new uses of intelligence have typically been discovered by highly intelligent people. A good friend of mine says, “Intelligence creates its own demand.” What about that? Maybe it’s diamond-shaped for humans if humans are driving the new questions that intelligence is set to answer. But what if the thing itself is creating new questions, and we train it to do that really effectively and just explore in an open-ended way, as all of humanity would, historically led by people with 150-plus IQs or something?

Jevons paradox is just a narrow example of price elasticity of demand. In general, all markets are a U-shape. Initially, when the product is terrible, there’s no demand. Then the product gets better, and people will pay a lot, but they won’t buy very much. As the price goes down, the size of the market rises because of Jevons paradox, essentially. But then you get to a point where further price decreases don’t actually drive enough demand, and the market shrinks. Eventually, of course, if you take the price to zero, the size of the market goes to zero.

One example I’ve thought about is the cost of fiber-optic communication, or just communication. The cost of moving information used to be super high, and people moved very little of it. Then, through a whole series of innovations, the cost of moving information went down a lot, the amount of information moved went up a ton, and the market grew a lot for a long time. Then the market stopped growing. The cost of moving information kept falling, and the market kept shrinking.

Nowadays, Cogent Communications routes, I think, 30% or 40% of all IP traffic across the internet through one of its divisions, which isn’t even its biggest division. Cogent Communications is a company with a market cap of a few billion dollars. It’s just not a very big industry. I’m not saying we’re going to get quite that small in terms of intelligence, but Jevons paradox is a statement about a particular spot on a curve, which doesn’t necessarily have to be true.

As for your other question—can more intelligence create more desire for intelligence?—I have a couple of thoughts on that. Certainly, maybe. I started out by saying you need to touch reality to answer questions. If you had a model that was going to think more about physics, it could certainly think lots of interesting thoughts about physics, but eventually it’s going to need to do experiments. That’s not to say we couldn’t also do experiments, but then we’re bound by our ability to physically do them.

I think there’s going to be this incredible period where there are so many cool discoveries, which we’re already starting to read about: “This paper was written, and this paper was written, and this paper was written. If somebody had read all 3 papers and thought about them, they would have realized this, but nobody had.” It turns out the models read all the papers, and they’ll figure that out. All kinds of great stuff will come out of that.

But at some point, we’ll finish that, and then it’s going to be about finding some way to think about things outside of what we know. We’ll have a hypothesis, but a model, even a super-smart model, won’t necessarily be better at guessing that. It will be better at guessing than the smartest person, but the smartest person’s odds are super low anyway. The model will come up with great ideas for experiments to do, and they’ll be worth doing, but that will limit our ability to expand.

The other thing I think about is my own use of AI. To go back to the commoditization point, I’ve totally commoditized my use of models, for what it’s worth. It’s because I’m lazy about how I do a lot of my building. I’ll tell a Groqbot to build things, and because I don’t want to use up all my Groqbot weed tokens, I then have it run all of them. I subscribe to all the things, and I’ll have it build using those subscriptions on a server. So I actually don’t know what model I’m using anymore. I used to know which model was better or worse. I no longer know. All of this is being optimized on the back end for me. All I know is that stuff is getting built, and it’s not using up my subscriptions, so we’re good.

There’s this batch of companies I’m curious what you think about: real-world experimentation companies that create an automatic loop. This is happening in robotics, in several places in bio labs and automated wet labs, and in automated testing labs. Any commentary on that kind of company?

Yeah, super great idea. Closing the loop in the physical world is going to be great, and it’s going to accelerate the process of learning. This is a great idea. But again, there are speed limits. The thing I keep coming back to is some version of, “This is amazing, and I’m super AI-pilled in terms of what I think these things are going to do and how useful they are.” But I’m also cognizant of the idea that things bump up against limits.

The way I like to tell this to people is, “Look, you’re really smart, so you must understand: being smart has limits. Being smart has limits.” The ability to just think your way to solutions is limited. If you’re going to build an automatic biology testing laboratory, connect it to a computer, have the computer come up with an idea, run an experiment, and iterate, that’s awesome. But if the experiment requires breeding mice, and mice need a year to grow up and die, that experiment is going to take time. It might be a great experiment, and you might learn a lot, and that’s great. It’s still going to take time.

The lab is only so big, so you might have 5,000 great ideas. You’re going to pick 20, and the model may be better. Your odds of that experiment working might even be 1%. That might turn out to be 10 times better than the average researcher, and that’s amazing. It’s still 1%.

None of this is to be pessimistic. Discovering all the things we should know because they’re already in our corpus and we forgot them is awesome. Discovering all the things and better directing our research is awesome. Automating that research to move faster is awesome. We will learn, and we’ll push the frontier, but this is not a case where the model suddenly looks around and says, “You guys missed it. Here’s some new world you can’t comprehend.” I don’t think that’s where we’re headed.

3. The AI Infrastructure Capital Cycle

It seems like the market, especially in private markets and among investors thinking about AI at the leading edge, has been very focused on investing in supply-related stuff: how we’re going to build and serve all the intelligence. There’s less focus on what I’ll call demand-related stuff, like someone watching personalized TV on their Meta app or whatever. What do you think about that first batch of companies as investments?

Super hard. I’ve joked that I’ve been willfully ignorant about them, and I’m probably worse off for it. But there’s this big build-out, and then you’re trying to forecast inference demand. I can’t forecast supply or demand within several orders of magnitude. How much more inference people will use as intelligence gets better and cheaper is a really, really hard thing to figure out. Equally hard is figuring out how much better and cheaper inference is going to get.

Even if you could somehow guess at those things, then there’s the whole value chain, and you need to be at the bottleneck in order to capture the rents. Finally, you have to consider your actual competition in that particular piece of the value chain. It’s just a super hard thing.

As I think about my use of AI as a consumer, I’m increasingly doing things like this: instead of having an AI chatbot talk to my daughter or text with her about how to do math, I have it make a 3Blue1Brown-style video explainer so she can watch a nice 5- or 10-minute video on powers, area, circumference, or something.

It’s super cool. I don’t think I would do that if it were going to cost $5. There’s real price elasticity. The point I’m trying to make is that I’ll film myself working out with my Muse app, and it’ll show me pointers on my form. I’ll do that for very little money, but I won’t do it if the price point is real—and I’m probably much less sensitive to price than most people. AI demand is going to be really sensitive to price, which intrinsically, I think, limits the total amount consumers are going to be willing to spend on this.

Maybe this is a time to talk about capital cycles or something. It sure does seem like the smartest people in the world, and many or most of the best entrepreneurs in the world, are heavily personally invested in this supply build-out. If you watched SpaceX’s Investor Day yesterday, this is basically the theme of it. Arguably, maybe the greatest entrepreneur of all time is dedicated to effectively doing this—building out the data center supply, et cetera.

This is just a historical question. If that’s all true, and there’s so much uncertainty about supply and demand, and there are other ways to build businesses and products and make money, why do you think so many of the best entrepreneurs, the smartest people, are so completely dedicating their lives to the build-out of the supply?

This is a huge infrastructure build-out of a huge and important technology, and it would be impossible for there to be a huge infrastructure build-out of a huge and important technology if there weren’t a strong price signal and a ton of talent going at it. So you may have your causality backwards. Obviously, these things work both ways. Could you imagine a telling of the story where nobody interesting who’d made no money was building out AI? That’s probably the best answer.

To be clear, this is a huge and important project. It’s only happening because super smart and talented people are focusing their lives on it. It’s great that they are. They will have and deserve to make lots of money. It’s just very, very hard from where I sit, and there are probably people who are smarter and have figured it out. From where I sit, it’s very, very hard to figure out where one would invest along this and whether or not you’ll make any money doing it. That’s not to say that there isn’t plenty of opportunity potentially there.

I’ve just decided to focus my attention more on the longer-term equilibrium results that come from what the world goes to. Instead of betting on which railroad tycoon is going to make money and how many railroads we need, I’m trying to think, “Okay, once you have railroads, I’d love to have the foresight to figure out that CPG brands come next.”

Hmm.

That’s the way I’ve been trying to think about it.

Can you riff a little bit about what you’ve learned about capital cycles or innovation cycles like this? I would love your perspective, having studied so much market history. If we’re in the middle of one of these, maybe the biggest one ever, what does history teach us about these cycles and how they work that’s interesting to you?

You could think about American society right now as a household that’s saving money for a house. We’re building a house, and we’re all feeling pinched because all these resources are dedicated to building out this big AI infrastructure. All these electricians and plumbers and everyone else are going to do these tasks, and as a consequence, there are higher costs elsewhere in the economy.

Undoubtedly, we’re going to get a lot of value for it, and we’re going to keep pushing it because no one really knows. We’re going to keep pushing it until eventually we overdo it. Along the way, people have to make tons of money, because otherwise people won’t keep pushing it. And because things have lead times and because it’s a winner’s curse, invariably, the last person to stop will regret it. It’s just so hard. It’s just so hard.

Then you combine that with the fact that there’s so much dynamism in the technology here, which I think was less true of other historic infrastructure build-outs. There’s so much change in how you do these things. Other people have made the point, which I don’t think is necessarily wrong, that we’ve talked about the models, but even at the chip level, there’s the potential for AI to commoditize even things like NVIDIA or—

Recursive self-improvement everywhere at the—

Everywhere, and a lot of it ends up being just intelligence. NVIDIA quite literally makes flat files that they send to Taiwan. Their flat files are very hard to make, but what they have is a lot of people who think really hard about it. So, in theory, you could make them. And then the question becomes: What if you step back a little from the frontier? The cost of making something comes way down.

In that world, would your model be that physical assets and scale are the things you’d want in a business you were to back on the supply side of this equation, like TSMC?

I don’t want to hold myself out as some sort of TSMC expert. Certainly, what they do has better attributes, I think, in that regard. They’ve also got this interesting problem where they did this super-hard thing that requires all this process knowledge built up over many years, and it’s super hard to replicate and much less amenable to AI, just because it’s so physical and because it’s all these little things, many of which aren’t even legible to an AI. It’s like some guy who’s in the system knows that you have to do this.

Because there’s so much demand and because they haven’t been able to meet it, they’ve basically caused everyone in the world to seek to copy their moat. That’s not a great fact. But more generally, stepping outside the AI build-out, if you think about businesses, we talked about the Meta example: businesses that have physical proximity or physical scale might benefit if they can use intelligence to serve their customers better or have lower costs as a consequence.

4. Agents Reshape Markets

I want to go back through the natural complements to intelligence. Could you explain why each one is a complement to intelligence, so people can go away and think about what companies have this feature that might be interesting? A good example is physical proximity. How would you think about DoorDash or Uber, or these businesses that have local density effects that matter a lot? Does intelligence being ubiquitous naturally magnify all of the scale winners of this type, where physical proximity is some sort of key advantage?

If you’re the low-cost producer of something, I’ll give a stylized example. If everyone has an AI agent, many decisions get offloaded to the AI agent because it helps you optimize and pick up all the pennies, dollars, and nickels that you were leaving on the ground.

I think about consumer choices on a spectrum. On one end is maximizing my bank deposit. This is something where I think I can almost entirely defer to the AI’s judgment. As long as it’s FDIC-insured, I just want the highest rate. On the other end of the spectrum might be picking my favorite sweater. That’s something where it’s all my judgment. The AI can source things for me to look at, but in the end, I need to decide. I can’t really delegate that judgment to it.

There’s a spectrum. Let’s pretend for a moment that garbage collection is something where I can defer the judgment. You might say, “No, I really care about the guy, and I trust the service, and he’s driving to my house,” or whatever. But let’s set all that aside for a second. Right now, if I’m in the waste collection business, I might be great at it, but it’s still really hard to get people to bother to switch. I doubt you know what you pay for garbage collection off the top of your head. Many people do, but I doubt you do.

So you have to do all this work to get people to pay attention and bother to switch garbage collection companies, even if you’re offering a better service at a lower price. You could imagine that their agent could talk to my agent and say, “Hey, I’ll save you 50 cents a month,” and the agent will ask, “Will the garbage go away?” I’ll say, “Yes, the garbage will go away,” and it’ll say, “No problem, then I’ll change.”

The first-order effect is that you move to the low-cost provider. But the general market effects are interesting, too. Let’s say there’s a street, and before, it would be more efficient for society if all the people on my street had the same garbage collector, because the guy could drive up the street and pick up the trash. He could offer all of us a lower price and make more money.

But it’s really hard for us to coordinate that, because we’re all lazy about switching garbage collectors. The guy with 2 people on the street can’t call everyone else up and say, “Hey, I’ve got 2 people on the street. Everyone else has just 1. I’m the low-cost producer. Let me roll this up.” A second-order effect might be that the guy who has 2 people on my street can go to everyone else on the street and say, “Hey, I’ll offer you a lower price on the garbage,” and everyone will switch. So everyone moves to the low-cost producer, who in turn gets to be the lower-cost producer. That actually creates, in some sense, a better business, where the low-cost guy wins a bunch more volume at much lower margins. As you study them, a lot of companies have mixed effects that point in different directions because of these things.

What do you think the impact of agents is going to be on your world? What will your role be? How will it be different in 5 years?

Yeah.

Will the best agent be able to build a portfolio of Carvana, Hilton Grand Vacations, and Meta in the same way that you would, because you trained it to think the way you do?

Well, hopefully better. In terms of what I do, I think that in the long term, we'll probably get competed out, right?

Basically, pricing risk is the thing you do?

Yeah. When computer-aided chess came along, there was a long span of time when freestyle chess was better: a human working with the computer program was able to outcompete either the human or the computer by itself. That's because humans thought about things differently from computers, and they were synergistic. I think in investing, for a while—

We're in that era now.

We're in that era, and that's because there are things I understand about the world in ways the model just doesn't, by virtue of being a person and having lived a life. And there's information I have about the world that just isn't in models, because as much information as there is on the internet, not everything is actually on the internet. I think that's going to be true for a while. The ability for me to use models to assist me and then make judgments will lead to better judgments when we work together.

You could imagine a circumstance in the future where that gap diminishes to the point where it doesn't matter anymore. I don't think models will be able to predict the future. As we talked about, the future is, in some ways, intrinsically unpredictable. But the model's distribution of outcomes may be the best you can do, and adding a human to the forecast would be like adding my judgment to whether or not it's going to rain in 7 days. You'd just be like, “Cliff, I don't really care what you think. I'm using the forecast. This is going to be better. However you fudge it, it's just going to make it worse.” That day may come.

It's an interesting moment. There's a lot to invest in. Think about judging management or something. Maybe the models will be able to judge management. They'll be able to interview them and figure it out. Then it gets really interesting: How do you raise capital? You'd be like, “You know, I'd like to run an amusement park,” and the model will say, “Sure. Here are 6 locations that have a 51% chance of an amusement park working, and as long as you're really committed to it, and I've assessed you, you have a 51.2% chance of succeeding because you're actually slightly below average.” You're going to be like, “Okay. Well, what can I do better than average?” And they'll say, “Well, have you thought about running a go-karting park?” Fortunately, I don't think it's going to affect my career that much, because I think it'll take a while.

Hmm. What about replacing price as the organizing principle and signal that coordinates economic activity?

Price is a miracle because it's a signal wrapped in an incentive.

What does that mean?

Price is an incentive. It causes you to want to buy more or less of something because it goes up or down, and it's also a signal because it says something about the relative value of something, the scarcity of something, and the cost to produce something. It's a miracle of a mechanism to coordinate things, and I think you'd be really hard-pressed to find a better system, even with unlimited intelligence. It'd be pretty hard.

Maybe another way of asking the question is like investing. If ubiquitous intelligence gets so much better and so much cheaper, and in some way it can replicate what you do, and you're really good at pricing risk... A friend of mine, Dan, was saying there are 3 levels to this. There's: I put in a prompt and I get something back. That's level 1. Level 2 is I set up a watchtower. My agent every night runs a set of tasks, and it happens in the background on a schedule. And level 3, which produces the most use of intelligence, is it's just always on, proactive, figuring out its own way to use its own intelligence, and that just goes on forever, and it's unlimited in how much it can do. If I think about that level 3 applying to everything, what things should be priced—

Sure.

—efficiencies, optimization—

Sure.

—all this kind of stuff, does that add something interesting to the price—

It sharpens markets. There's a lot of stickiness to prices. It's a well-known macroeconomic phenomenon: price stickiness. There's a lot of lazily priced stuff. Think about deposits and such. It's a really competitive world out there, but it's actually not as competitive as, in theory, it could be.

Mm-hmm.

So I think if everyone has agents working on their behalf and constantly optimizing, the price of garbage collection will move around a lot more. As all the input costs of garbage collection evolve, and every garbage-collection company is constantly rerunning its financials, trying to figure out its true marginal cost, and bidding for every customer it could serve for more than marginal cost, every agent will be switching to the lowest-cost provider. You won't just pay $60 a month for 12 months and then have the guy come back to you and say, “It's a 5% increase. It's going to be $63 next year,” and you don't even notice because they're just picking up your garbage. Your garbage collection will vary by day. One week it'll be whatever. There'll be all this interesting, much more accurate, much more rapidly moving pricing, I would imagine.

Have you seen anything interesting in capital markets innovation? There was this awesome The Economist story talking about the behemoth that is American capital markets, how good it is at adapting to new stuff and creating new instruments and new ways of financing things, and how that's been as much a part of the American story as technological innovation.

Sure.

There's this amazing interrelationship between Wall Street and Silicon Valley, one way of thinking about it. Anything interesting percolating for you in markets, in capital markets, in ways of financing things?

It's fun to take a moment to pause and maybe congratulate ourselves, as capital markets, for what we're doing with AI. People were worried about a decline in American dynamism. I think we just redeployed, what is it, 3% of GDP right into this highly speculative venture, basically on a dime. It's remarkable. Of course, who knows? Maybe it should be 2%, maybe it should be 6%. But it is amazing that we can redeploy resources at this scale this quickly, and there's no one else in the world that's been able to do it.

Certainly, the inability to move investment in productive resources can be a binding constraint on productivity growth, even with—and especially with—rapid technological progress. And so, at least with regard to capital—and obviously we talk about capital—the capital investments are tied to people doing different things. Americans are doing different things today. Every day, people are waking up and going to some place in, I don't know, Pennsylvania, and building a data center. That wouldn't be happening if it weren't for this whole infrastructure we have that's somehow been able to move this enormous amount of capital. It's incredible.

As for deep innovation, I think the fundamental idea that people are taking risks on speculative ventures and slicing up the pie in different ways is irreducible. I think people come up with lots of different names for what are essentially the same things over time. I don't think that's deeply different, but it is incredible.

I should also note that I'm sure we'll discover all kinds of fraud and bad behavior and stuff eventually when the tide goes out. But I think, given the magnitude of this, I suspect it's so clean. It's amazing. If you read about the railroad build-out or something, the amount of waste and fraud—and invariably there is some—but there are a lot of people trying to basically do the right thing here. The governance parts that we just take for granted are incredible.

America's capital markets, and capital markets in general, are incredibly underappreciated. The idea that any person can buy $100,000 of stock and get a portfolio of well-run, well-motivated teams building companies that are doing mostly the right things, do it basically instantly and at basically no cost, and then sell it as soon as they're done is amazing. And the fact that all those companies and other ways people invest can then allocate society's resources pretty well in the scheme of things is just incredible. The whole system is very underrated.

Why has America so dominated this, do you think?

My pet theory, and everyone has their own theories, is some sort of interplay between genetics, cultural evolution, and institutional evolution. Basically, the people who came to the United States were the subset of people most interested in taking risk, because going to the United States was very risky. I think that risk-taking has a meaningful genetic component, and so I think that tendency is just more common among Americans.

When you're in a place where risk-taking is more the norm, there's also a whole bunch of other things that come from that. It's more the norm, and then that leads to buildups of cultural and institutional things that reinforce it. It's understood that people can take risks and fail. That leads to bankruptcy laws that make sense. That leads to labor laws that support that. That leads to venture capital that supports that.

And all of those things reinforce that. And then that leads to success stories. You get this whole soup. But that’s my pet theory.

So if there’s a moon base at some point, you’re going to be bullish on that community’s ability to do really well economically.

Yeah. Actually, that’s a great way to put it. The United States has been persistently innovative in a very amazing way. So if there’s a moon base, I wouldn’t be shocked if it’s ours, even though the Chinese are brilliant and numerous, and the Europeans exist.

When you see rates go up, you see that great long-term chart of rates—let’s say the ten-year rate in the U.S.—and it just looks like this forever, basically our whole lives. And I was going, “Ah.”

Mm-hmm.

Comment on that chart.

Well, there’s so much, and I’m certainly no macroeconomics expert, nor do I spend too much time thinking about it, but I’ll make a couple of observations. One, obviously, is just in general, when you’re a society and you’re building, you’re making a giant investment in the future, it means you’re going to have to consume less today. Fundamentally, all of the resources going into this AI build-out—

Are about the future.

—are about the future. And fundamentally, the plumber building that today is not making consumption today. He could be making today’s present consumption. Instead, he’s making a data center, which is future consumption. So one piece of a very complex puzzle that I’ve thought a bit about is this idea that rates are higher because we’re inducing people to save for the future. And the way you do that is you pay them.

People would kill me if I didn’t ask about your portfolio a little bit, which is so interesting to me. It’s 80% Carvana, vacations, 5% Meta, basically your portfolio. Roughly, it’s like 5 things. And one thing is 80%.

Sure.

So if you think about your portfolio and those companies, which obviously you know extremely well, and you were advising them on how to think about using AI to transform their businesses, what’s that speech look like?

Well, it starts by making sure that everyone is vibe coding.

Yep.

Starting at the board and the CEO, everybody needs to have hackathons. And of course, I brought this to Carvana, and they said, “Yeah, we call them pizza parties.” I don’t mean the developers; I mean everybody needs to have hackathons. It’s not so much that you’re going to build hugely useful software necessarily that comes out of these hackathons. It’s that you’re going to teach everybody what can be done, and that’s going to change their expectations, I think, for what can be done.

And then the second thing I tell them is the CEO needs to basically turn to the head of IT and say something to the effect of, “So I want my Grok bot, Meta Muse agent, whatever, to be able to access every system in the company on a read-only basis. Everything I could look up if I asked somebody, I want to be able to get that in a prompt.” And the guy’s going to say, “That’s really hard. We have umpteen trillion problems with that.” And he’s going to say, “That’s my point. I want that.” And that’s a big process, and it’s system by system and all the rest, but if you can get to that, then that gets you read-only access.

To the starting line.

Yeah, to the starting line. That gets you read-only access, and then the next logical thing is the CEO needs to use this everywhere. Everyone else is going to get jealous, and you’re going to say, “Oh, you can have it too, but we need to work out your permissions.” And then that, in theory, percolates down, and now you’ve got permission for read-only access.

And then invariably, someone is going to say something like, “I’d really like to put data into the system.” And you’re going to say, “That’s fantastic. We’ve got to work out write permissions.” And then someone is going to say, “I’d love to change the system because we just did this vibe-coding thing, and I’m confident we could do it.” And you’re going to say, “That’s also fantastic. We need to work out a process by which we’re going to manage changes to the system.” What I just described, sure, it’s 4 bullet points. That’s like—

Years of work.

It’s impossibly hard. So the question is, are you in a business where you have time for that to happen before someone who starts—

Before the jump ball, someone else that starts from go beats you.

Exactly. And so it’s good to have a really good business.

It seems like one of the conclusions from that is it’s the best time ever to start companies, both because of capital market conditions that people will fund at fairly low cost of capital, highly experimental things, but also just because with these tools, if you start without the baggage, you can go so far so fast. Do you think that’s roughly right?

I mean, yeah, it’s the best time to start a company until tomorrow, so I agree. It’s amazing. I’d also just say that, from my perspective, people are such a pain in the butt. If you’re running a company, they’re such a pain in the butt, and the idea that you have all these agents and stuff means you can do so much more with so many fewer people.

Of course, one of the challenges always is if you can easily do something, then so can everybody else. So there has to be some wedge by which the thing you build produces enduring value. Does it have a path where you end up with some sort of lock-in or proprietary data? In this world where all the agents are happily changing providers every day, what gives you any enduring economics?

But then, of course, there’s the further thing, which is that obviously, if you do the thing where there seems to be an obvious path to enduring economics, then that’s a bad thing to do because everyone is trying to do that, so it’s very competitive. The really great outcomes are the ones that make no sense. You shouldn’t do it at all, but then it turns out there is a path, or the world turns.

What do you think are the most underrated sources of future competitive advantage, then? If ultimately good businesses need sustainable edge, sustainable competitive advantage to have great profits, is anything underrated? We talked about attention and data and physical proximity. These are examples of things that might be super or more valuable in the future than they are today because of cheap intelligence. Is there any type of—maybe it’s regulatory capture—are there any types of edge that you think are underrated?

I get the sense that a lot of these are appropriately weighted.

Right.

I guess there are 2 interesting questions I’ve had. One is: if you think of corporate margins in aggregate, do they go up or down?

Down, like this?

Right. Do they go up or down? And I think everything we’re talking about argues for them to go down. There are many businesses where they make margin because—

More consumer surplus.

—more consumer surplus.

Everywhere.

Everywhere. What’s interesting is if you go back and read a lot of the research reports from the ’90s when the internet was coming along—

Same thing.

—they said the same thing. So the base rate here might not be great. Maybe we’re wrong, but I think the main reason they were often wrong was aggregation economics, Ben Thompson’s Aggregation Theory. So maybe absent that, maybe it wouldn’t have been so wrong.

And I do think that a lot of aggregation theory breaks down because that basically relies on the idea that consumers only look at the first thing on Google. But it turns out if your agent looks at everything on search, then being ranked first doesn’t matter, and all the core things like quality and price and everything else—service—become more salient.

But the other thing is that maybe another downstream effect of this is you’ll be able to build more complicated businesses. Right now, a lot of industries are more horizontal in structure because the complexity of running your slice is very high, and that makes it hard to do what’s above and below you without undue complexity. With AI agents and basically cheap and ubiquitous code and intelligence, you could build the whole vertical stack. People are trying to do this, I think, in other countries, but you can have a full-stack pet care business that does everything from make pet food all the way up through pet insurance, pet hospitals, pet vets, and all these things that historically were multiple layers.

But now, because you can solve the coordination and reduce the principal-agent problems by having AI do it, maybe those sorts of businesses are possible. And then maybe those businesses are higher-margin in the end, more efficient, and more consumer surplus.

Carvana is an example of that, built before—

Hmm.

—this was easy to do, but maybe people can do it in other verticals, in which case margins go up and there's consumer surplus.

The other thing is how this affects wealth and income distribution. I can tell 2 stories. On the one hand, a lot of people have talked about the idea that it concentrates income distribution. You've got all these people who made money on various labs and data, NVIDIA, and this idea that there are increasing returns to small groups of people.

But the counterargument is that if we have cheap and ubiquitous intelligence, a couple of things happen. One is that there are all these rent takers in the economy. Think of the software company that charges bowling alleys. Those guys all get smoked, which sucks for them, but it does lower the cost of bowling. That increases living standards for everyone who bowls. And that's a flattening effect, because my spending on bowling is a small proportion of my income, but people who make less money spend more of their income on bowling.

If all these little rent pieces go away, corporate profits decline. All the money that the average American is giving to Bank of America goes away from Bank of America. The capitalists lose; the average consumer wins. So that's leveling.

The other piece is that there's a whole professional class whose role you could basically define as intelligence. They take big corpora of information and apply them to certain circumstances in a fairly straightforward way. If those all get commoditized, everyone gets the best intelligence, ubiquitously, which is good for everybody, but particularly good for people who can't afford the best right now, or anything at all. Those individuals earned these extra rents because they were smarter. They then have to drop to their next-best alternative, and in theory, that might flatten—

Hmm.

—the curve. Now, the counterargument to that might be something like, “Well, yes, but returns to initiative, ambition, and intelligence could go up because now I can do so much more. These models are like a bicycle for the mind on steroids. They give people who are smart and have initiative the ability to accomplish so much more.” Maybe that actually causes increasing returns to that subset of people, and it steepens the curve.

Truman asked for a one-handed economist. I seem to very distinctly have 2 hands on everything. I think that's—

They're big, interesting questions.

They're super interesting questions, and I find that in a lot of situations I look at, there are a lot of crosscurrents. We've started the conversation about the AI build. On the one hand, I can see all these incredible ways to use AI, and I use a gajillion tokens. I can see how I could use a gajillion more. Obviously, I am at the far right end of the distribution. If everyone uses as many tokens as me, we're so short. It's going to be this huge thing that's going on forever, and everyone who's in the build is going to make a fortune.

On the other hand, the amount I get per token is going to the moon. It turns out I'm a pretty unusual guy in a lot of senses, and people are going to adapt more slowly, and the cost per token is going to collapse. So maybe we're totally overbuilt. I can tell both stories with equal certainty.

I have a good friend who's been running this very cool experiment. If you break investing down into picking what to own and then how much to own of each thing you've picked, that's an oversimplification, but an interesting way of thinking about it. He's been running this experiment where he continues to pick the things to own, but then runs his own portfolio construction as he would have in the past in one lane. In a second lane, he's completely handing that off to an AI harness he's built and saying, “You do all the portfolio construction, but stick with the stuff that I've picked.”

What he's found so far is that it's drastically outperforming his version of portfolio construction, which itself is quite sophisticated. The simple argument he would make is that AIs are just really good at optimization. They've read every paper on portfolio construction and all that kind of stuff, and on bet sizing and the Kelly criterion. They know it all. In the future, perhaps we make that transition where the investor's job is just picking what to own, and not how much of each to own.

I ask you this question because you have such an unusual portfolio where, again, 80% of it is in one specific asset. My guess is that if you were in lane 2, AI would weight your portfolio very differently from how you've chosen to do it.

I'm underweight.

So I'm just dying to know what you think about this idea and the future.

That's clever.

Underweight Carvana. Should be 90%.

Time scales are probably interesting to think about on that, too. As I think about how AI is going to encroach into this business, I think it's going to start at shorter time scales and work its way up to longer time scales. One of the challenges is that there's no verification loop. Your friend's ability to say that is based on results over—I mean, certainly not 20-year results.

I strongly suspect that the march of AI into markets may have already started. I don't know how Renaissance Technologies did it, but I assume—

It seems like they might have discovered a transformer for everybody else. I'm not kidding.

It wouldn't shock me if people who invest over shorter time scales, starting with seconds and working their way out to minutes, hours, weeks, and months, find themselves delegating more and more to models. That makes perfect sense. You could definitely imagine bifurcating that, trying to enrich the model.

I think other people do things like this already, and this is just a more sophisticated version. You can have a fully quantitative strategy, then enrich it with what you think is qualitative, human-designed analysis, and then have a quantitative strategy to incorporate the qualitative. I think many of these multi-strats do these sorts of things, and what he's doing is just a slightly different flavor of that. There are so many ways to be successful, and models have just added more dimensions of freedom.

But what about your own portfolio? Say a word more about how you feel comfortable having a position that big, forget the company that it is—

Sure.

—when there's so much research that would suggest that's a crazy way to build a portfolio.

A lot of that research, I think, is more about looking at price changes in securities prices and less focused on thinking about it the way someone who owns a company might think about it.

I mean, we're talking about Carvana. Were it not for my sense that Carvana's infrastructure and brand and everything else are just so differentiated that there are few scenarios where it works out really badly, then I wouldn't feel so comfortable owning so much of it. It might just be as simple as that.

That portfolio approach would have worked absolutely. Certainly, if you discovered Walmart or Home Depot or Fastenal back in the day, you wouldn't have taken full advantage of it.

Hmm.

And that's my sense with Carvana. If you read history, of course, all the people who did really great tended to be very concentrated. But of course, history is littered—

Diversification buys that.

Yeah. Or with the history of people who didn't.

I also think that a lot of that literature is about what most people who are looking for average or slightly better than average results should do, and it's totally the right thing to do. But that's not what I'm looking for. I like Carvana a lot. It wouldn't upset me if the stock went up a bunch and I felt compelled to diversify. That would also be okay. It's just more of a trade-off between expected returns and knowingly taking an unusual risk.

What do you think about education? You and I have kids of roughly similar ages. Comment a little bit on how you're thinking about their educations, the institutions that have driven education in this country and have been really important in that whole path-dependent story you've described. What do you think the present and future of education looks like, or should look like?

I've been talking to the head of my kids' school about this, and more and more it seems that his thinking and my thinking are heading in this direction: the way you're educating kids is getting more bimodal in terms of what you're looking for.

Let me step way back. What is the benefit of intelligence, personal intelligence—your own intelligence—in a world where there's infinite intelligence from models? You might say, “Oh, it's useless. It doesn't matter. Someone with a 50 IQ can navigate the world just as well as someone with a 150 IQ.” I don't think that's quite right, and intuitively it doesn't feel right.

But I think the way to nail down where it is: the more intelligence you have, and the more crystallized intelligence you have, which would be education, the better you are at communicating with a model and telling a model what you want. I would have a very tough time having a deep conversation with a model to create leading quantum physics experiments because I don't know anything about quantum physics. The model could try very hard to explain things to me, but in the end it would be explaining things in crayon and Nerf, and it would be very hard to get to the leading edge and get the best out of what the model has to offer.

Whereas in an area where I do have some expertise, the model and I can communicate effectively, and I can direct the model to my will much better. To direct the model to do your will—which I think is going to be an important skill in the future—you do have to understand something about the world. Smarts help with that, and education helps with that. You can’t ask it to assume something is a stochastic process unless you know what a stochastic process is. It could teach you what one is, but that’s a lot of friction in figuring this all out. You don’t have to work out the math, but you have to know what it is.

When I’ve been thinking about education, there’s this idea that you’re teaching your kids a bunch of foundational skills that build on one another. I think, in the end, they have to be able to do the sum total of those skills without models—not because they’re going to do them without models in the future, but because if they don’t know how to do multiplication, it’s going to be really hard to understand matrix multiplication. And if they don’t know what matrix multiplication is at a high enough level, they don’t have to be able to do it really well, but they have to be able to understand what the model is doing when it’s doing it. That’s so they can direct the model.

I’m increasingly of the view that we should spend a bunch of time with kids basically separated from these models, teaching them basic skills that we think are important. What those skills are is a really hard question. But the idea is that, yes, the model can do it, but you need to know it so that you can tell the model what to do. On the flip side, some English teacher is like, “I don’t like models because the kids just cheat on their Shakespeare essays.” I’m like, “You’re not being ambitious enough. If the kids are cheating on a Shakespeare essay—”

Give them something harder to do.

“Give them something a hundred times harder and more ambitious, and have them use models to their fullest extent. Have them bring the work product back, and then make sure they didn’t just give a prompt to the model to do the homework. Make sure they own it. Ask them why it’s like this. Force them to justify it.” I think that’s another really important thing, because kids need to be good at using them, and they also, frankly, just need to be taught that they can do anything.

You and I get to talk every so often on the phone. Every time, I wish it was recorded so that other people could hear it, because I learn so much. I think this particular conversation today, which is sort of a version of a call we had several months ago, is such a valuable set of food for thought about the future of intelligence and what that means for markets and companies. It’s a little bit—not in opposition to, but—a different perspective from the most common one you’re hearing, which is unlimited build-out forever, short everything forever, no cap on the demand for intelligence. I’ve loved talking to you about it both on the phone and today, so thank you so much for your time.

Thank you.