为什么科学家无法重造一台 Polaroid 相机 [César Hidalgo]

I'm César Hidalgo. I'm the director of the Center for Collective Learning, and I recently completed this book, *The Infinite Alphabet and the Loss of Knowledge*. The book has 2 ambitions. The scientific ambition is to establish the scientific study of knowledge by showing that, actually, it can be organized around 3 laws: a law governing how knowledge grows in time, a law governing how knowledge diffuses across space and activity, and a law showing us how we can estimate its value.

But it also has a policy ambition, which is that when we try to develop the knowledge sectors of our economy, we have to make sure that we incorporate these laws into our policy strategy and design. If we don't do it, we're going to have failed development efforts. So the book also tells a number of stories of failed development attempts that defy the laws of knowledge, and that I equate to trying to build a rocket without respecting the law of gravity or understanding chemistry or aerodynamics.

We're going to try to understand how knowledge grows and flows, and what the speeds and rates are at which that happens. What are the functional forms that govern the growth of knowledge over time, and how does the growth of knowledge over time change when you move from the scale of individuals and teams to that of industries? We're also going to look at how knowledge crosses mountains and oceans, how it moves between activities, and how those movements also satisfy certain laws and principles.

We're then going to try to figure out how we can count knowledge in a world in which the idea of 1 plus 1 knowledge equals 2 knowledges doesn't make a lot of sense, because knowledge is non-fungible. It's made of a lot of unique components, and we need to find ways to score them that can help us understand the potential of economies.

Many economists try to develop economies by creating cities of knowledge or science parks and so forth. Usually, when they do that, they do it in ways that we could say are boneheaded because they defy these principles, and those projects involve vast amounts of money and end up in failure. So the idea is that, well, if we understand that knowledge follows principles like other quantities that we have learned to understand in the past, like temperature or other things, then we can think of policy in light of these principles and try to create policies that do not contradict them, so that we can develop knowledge in ways that are compatible with its nature.

Intelligence is about the efficient acquisition of coarse-grained knowledge, and you develop this idea that knowledge is incredibly important and we've become obsessed with it. So we've been thinking, well, what does it mean to understand something? In a way, we've developed this incredibly abstract view of knowledge, almost like it's a probabilistic graphical model, or it's a symbolic expression, or maybe the weights in a neural network are knowledge. But this idea that knowledge as a quantity is really important is something that has been impressed on us.

Yeah. I think that's something that you have in common with other disciplines. In my case, I'm coming more from the perspective of economics and social psychology, in which we look at knowledge as this sort of quantity that is essential to explain economic growth and the wealth of nations. This is something that has led to a couple of Nobel Prizes. In 2018, Paul Romer got the Nobel Prize for endogenous growth theory. This year, Philippe Aghion and Peter Howitt also got the Nobel Prize.

The idea of Romer in particular is the one that is interesting and I think might be different from the way in which maybe a computer scientist thinks about knowledge, but it's the following: when you're trying to explain economic growth, you're trying to explain output. Imagine you have 10 carpenters that have access to hammers, nails, and boards of wood, and they have to produce birdhouses. These 10 carpenters produce 10 birdhouses per hour.

Now, let's say that you want to produce 20 birdhouses an hour. Well, you might need to double the number of carpenters because if they're doing the same thing and you want to do more of them, you're going to have to have more carpenters, more nails, more hammers, and so forth. So that tells you that labor and capital are rival inputs. If you want to increase output, you don't increase output in per-capita terms; the birdhouses per carpenter remain the same.

Imagine now one of the carpenters figures out how to build a nail gun that embodies knowledge, or figures out a technique to organize the workshop differently that saves them some time, and they're now producing 12 birdhouses an hour instead of 10. Knowledge has this property of being non-rival, that it can be shared without being depleted. I can teach you a song, but I still know the song. If I give you a hammer, I cannot use the hammer while you're using it, because it's rival.

What economists figured out in the 1980s and 1990s is that if you wanted to explain economic growth, which happens in per-capita terms, the only way that you could do that is by assuming that growth was a consequence of a non-rival quantity, something that could be copied without being depleted. That was ideas or knowledge, and that became a big revolution in the 1990s. In the 1990s, everybody was talking about the knowledge economy and the idea that knowledge is the secret to the wealth of nations.

But what my book tries to do is bring that to the next level, because in that interpretation from Romer and other people in the 1990s, knowledge is still some sort of quantity that you can accumulate in a barrel. It's undifferentiated. So my book focuses a lot on the fact that knowledge has another property, which is that it is non-fungible, not only non-rival.

That non-fungibility is what makes it interesting to study, because it has all of this categorical differentiation that requires you to use math and a set of representations that are more similar to the ones used in machine learning, which also deals with non-fungible things like language. Words are non-fungible.

This non-fungibility thing is fascinating. All of us have this intuition, right? You hire someone, and I'm a big believer that knowledge is situated. There is this fanciful idea about knowledge that it's a completely abstract thing: you read a book and you acquire the knowledge, and it can just be copied any number of times. In fact, that's the argument that AI existential-risk people make. You can just copy the language model, and now you've got 1,000 Einsteins instead of 1.

What we find in practice is that it's quite difficult to exchange knowledge. Why is that? There is a tacit and implicit idea there that knowledge is something that something can have, while my view is that knowledge is a much more collective phenomenon.

Okay. It's not something that you can put in something like a book. In my opinion, the book doesn't have knowledge. The book is an archival record of some ideas that I was able to put together in a nice structure. But you cannot have a conversation with the book in the way that you can have a conversation with me, in which I can tell you the story of Jacquard or the story of Sam Slater or the story of how Sony got started based on what we're talking about and have that dynamic response.

Knowledge can only go to work when it's embodied. You cannot throw a bunch of engineering manuals and cement into a gorge and expect to get a bridge, because the books don't have knowledge; teams have knowledge, organizations have knowledge, and all of that.

The diffusion of knowledge is something that is hard, and that's something that we have established really well. What is interesting to me also about this field of study is that people say, well, there are no real laws in economics. But when it comes to economic geography, there are a lot of things that are very well established and that are law-like.

For example, the fact that knowledge diffuses more effectively at shorter distances, and that short-distance diffusion is explained by social networks, has been established not by 1 or 2 papers but by dozens of studies, if not maybe hundreds of studies, that have verified those effects. The idea that knowledge moves more easily among related activities, and that this can come from the complementarity of the inputs that are required to develop each one of those activities, is something that we also call the principle of relatedness, and that has been documented by hundreds of studies.

So we do have law-like behavior for the growth, diffusion, and value of knowledge that we're starting to understand. Some of that law-like behavior goes into your question, which is why it is difficult to diffuse knowledge.

The words in a paper or in a book are completely different from the actual physical, embodied process. But is there a middle way? Are you saying only the physical, embodied process is a form of knowledge and understanding, or do you think there exists any type of model or representation which we could say is a form of understanding?

The problem of talking about knowledge—and that's something that I addressed in the introduction of the book—is that it's a word that we use to mean vastly different things. We kind of understand that from context, but it's good to classify those different ideas, and there are people that have done that.

One of the ways that you can understand different types of knowledge is by thinking of a detective novel. A detective novel or a detective TV show usually starts at a murder scene and ends in an arrest. It's beautiful because the writer just needs to fill the space in between.

And it starts usually at that murder scene, when the detectives come in and start collecting what we call factual knowledge. There is a bullet hole in the wall. There was a call at 7:00 p.m. last evening. Those facts don't tell you about the motive of the murder or who was involved, any of that stuff.

Factual knowledge is knowledge that is very easy to diffuse. I can tell you that Santiago is the capital of Chile, and you can remember that. You can transmit that information, or that little piece of factual knowledge, very easily to someone else, and so forth.

Then you have conceptual knowledge, which is what usually the hero of the detective novel would do: putting everything together in a story where all of the facts are little anchors that can be used to validate that story. That's the story. The bullet hole was there because, when the murderer tried to shoot, this person moved to the side, and the phone call was placed because someone was trying to warn him. They figure out the entire story.

Now, to validate that story, what they need to do is sometimes collect additional evidence that is not factual knowledge but is collected through procedural knowledge. They may have a little bit of blood, and they need to send that to a DNA lab for sequencing. The DNA lab has procedural knowledge because it understands how to perform that procedure to sequence DNA, and that produces another fact that gets put into the concept.

So when we talk about knowledge, we talk about all of these different things. Now, there is another distinction that is extremely important, and I use a story at the beginning of the book to illustrate it. Especially among academics or people who are highly educated, we talk about knowledge as these sorts of truths that have been validated by the scientific method and so forth. The book is not about that knowledge.

In economics, knowledge is not just about validated truths that have come from universities, scientists, or researchers. There is knowledge in a lot of different things that is much more pedestrian and common. A car mechanic has knowledge. A baker who has been producing different types of pastries and breads would have knowledge. Everyone has knowledge.

Knowledge is highly specific. It is not necessarily made up of things that are 100% guaranteed to be true because of the scientific method, but it includes all of this experience and received wisdom that people have. That allows the world to work, because the world works not because everybody is operating according to a scientific theory, but because car mechanics know what they're doing, gardeners know what they're doing, and the guy who comes and cleans the pool knows what he's doing and has his own experience.

Maybe it even includes knowing how to deal with pesky dogs if you're a pool-cleaning guy. They have knowledge about how to deal with that, which comes from experience, and it's not what you would find in a book. So it's about that more democratic definition of knowledge.

Yeah. I do agree that we should have a notion of knowledge that doesn't completely depend on humans in the most abstract sense. We might say that it's a form of modeling. There are certain types of systems that we might say are alive, and part of the process of staying alive and minimizing this free energy, you might say, is being able to model the world.

The only reason I'm bringing this up is that you said there are facts—which is what the state of the world is—there's procedural knowledge—which is what we can do—and there's conceptual knowledge, which is how to think. You were talking about this detective film: he was doing this conceptual type of reasoning, where he was imagining possible worlds. He was saying, “What if this person killed the person? What if this person did something else?”

That kind of imagination is simulating without direct physical experience. This person had knowledge, and what they did was like a jigsaw puzzle. They were just trying different configurations of counterfactual futures. They found one that was plausible, and then they generated a hypothesis. So that is a form—not necessarily of a physical, embodied process—but also a form of mentalized, internal thinking.

Yeah, and I think that's correct. The reason I think that is correct is because it involves knowledge that is simple enough to fit within an individual. The thing about knowledge is that it can be such that you need multiple individuals to hold it.

There could be a detective who can put all of the pieces together, generate multiple representations and alternative stories, and use facts and evidence to decide among those alternative stories. But when we talk about economic growth and development, we're talking about knowledge that tends to be procedural and that tends to produce products or services that can improve people's standards of living.

For instance, manufacturing an aircraft is an operation that simply cannot be done by a single individual. No individual has all of the knowledge needed to manufacture a large jet passenger aircraft. That knowledge tends to be distributed and embodied, in this case, in networks that involve humans and machines.

They include printed material from manuals, an LLM that is now helping retrieve some information from those manuals, and the experience of people who have worked on that same model in the past, and so forth. The book focuses a lot on knowledge at that collective level. I run a center called the Center for Collective Learning for a reason, because I think about learning and knowledge at that scale.

I think that's a very different story from figuring out the right theory in the detective novel. It's a story that I think is not so logical; it's much more experiential, and we do have models for that.

The model that I love in that space is one by Linda Argote. She's a professor at CMU, and she says that an organization is a network that connects 3 types of nodes: people, things, and, let's say, ideas, concepts, and procedures—something more intangible.

At any point in time, an organization is a network in which some people are working with other people, some people are using tools to produce goods, and so forth. An organization learns not only through the learning of people; it also learns as that network reconfigures.

That's kind of interesting because it's a parallel to the deep-learning idea that you're adjusting weights. In an organization, we're also adjusting weights. We discover that maybe Tim is working with Robert, and they don't get along; they compete, whatever. So maybe he'll work better with Charles.

The moment that management—or maybe organically—has Tim start working with Charles, the organization learns something. Maybe Tim was working in marketing, but he hates marketing. Maybe Tim wanted to work in engineering, and if we assign Tim to this different activity or this different tool, then there is learning. There is organizational learning that happens only by reconfiguring the same parts in a system.

That's a model of learning that goes beyond the individual and that has an analog to the types of learning that we're trying to reproduce, I think, in silico right now. But still, I would say the in-silico models are still individual-learning systems. They're not so much collective-learning systems that involve all of these other social relationships and complexities.

I absolutely love that. There was a wonderful example in your book. You're talking about Barnes & Noble, and they're over there in Seattle. They went up against Jeff Bezos and said, “Well, Jeff, we've just launched a website, and we think we can do what you do better than you do.”

Jeff said, “I don't think so. You might have a website, but you're a completely different type of business from us. We are geared up. We have the logistics. We can send individual things anywhere in America.” They were set up for wholesale and retail. The thing is, they could do the same thing, but they were wired differently.

So that brings in the idea of architectural innovation and architectural knowledge. It's a very interesting concept that was introduced by Rebecca Henderson from HBS. The idea is that when you innovate, often you have what would be called incremental innovation, in which you are changing a component.

One of the classic examples in this literature is the manufacturing of aircraft. If you had propeller aircraft with combustion engines, changing one engine for a more powerful engine or a newer engine model was something that you could do relatively easily, because you didn't have to redesign the entire airframe. You just brought in the new engine, replaced the old one, and you were done.

When jet engines were invented, you needed to redesign the entire airframe to be able to produce an aircraft. So the companies that were operating with combustion engines went bust—most of them. There was a new wave of companies like Boeing, which were newcomers at that time and specialized in jet engines because they were designing the entire airframe around the new engine.

In the case of Barnes & Noble and Amazon, those are very good examples of architectural innovation, because you might think that Barnes & Noble is able to ship millions of books to all of these stores.

It has thousands, if not maybe tens of thousands, of employees who are experts on the business of books and dealing with clients. The idea of shipping the book directly to a consumer might look like a small incremental innovation, but in reality, it was an architectural innovation. When I do talks and present this with slides, I show a picture of a Barnes & Noble, and next to that, I show a picture of an Amazon fulfillment center, which looks kind of like this part of the airport that is sorting all of the different luggage. That shows that that little idea of just shipping directly to the consumer required a completely different organizational design, and the distance between the Barnes & Noble organization in this network that we were describing before, in that model, and the Amazon model was enormous in reality, just because of that change.

Exactly. This is the reason why, in my opinion, LLMs are not intelligent: They don't have this coarse-grained dynamic adaptation of their architecture. But we're getting ahead of ourselves a little bit. At the beginning of the book, you spoke about this concept of a person-byte, which is roughly how much one person can know, and we're a collective intelligence. You spoke about this kind of power-law learning curve, and we'll get to that as well. There was one fascinating example you gave: You were talking about this shipbuilding company, and over the course of—I think, was it the Second World War or the First World War?—they became much more efficient at building ships. Was that because of experience, or was it because of process?

The first law of knowledge, the law of time, is divided into several subprinciples, and the first one is about the growth of knowledge in individuals and teams. That's a story that starts with Louis Leon Thurstone. He was the first one to, in my opinion, map a really good learning curve in 1916. Funnily enough, he started as an engineer and then actually produced a camera that got him an interview with Thomas Alva Edison. He decided not to work with Edison and go teach at the University of Minnesota instead. He becomes frustrated that he's really good at math and engineering, and it's hard to teach it to students. So he becomes interested in learning. He goes to Chicago, enrolls in a PhD program in education, and after a year switches to the program in psychology. There, he gets access to a data set that was being collected at Duff's College of Business in Pittsburgh, in which you had records of how well people learned how to type.

Imagine you have a mechanography class. People are learning how to type. These are 18- and 19-year-olds typing on a typewriter for the first time. You see every week how many words they're able to type per minute—every 4 minutes, actually—and then you see how many pages they've written throughout the semester. When you put those 2 things together, you get a very neat learning curve that follows this sort of power law, like a square-root type of shape, in which learning is really fast at the beginning and then it peters out.

Then, in 1936, that's about 20 years later, Theodore Wright, an aircraft engineer in the United States—he was actually important enough to be in charge of aircraft manufacturing for all of the United States at the end of the Second World War—publishes a paper in which he looks at the cost of producing an aircraft. He's very smart: He looks at the cost of the last aircraft produced in a batch, because aircraft are produced in batches, and he finds also that the number of man-hours as a function of the number of aircraft in the batch decreases as a power law. So it's the same result that Thurstone got. In one case, you can look at capacity; in another case, you can look at cost.

Then, in 1965, Leonard Rapping, an economist, grabs data from the Liberty ships. The United States was producing, during the Second World War, an insane amount of Liberty ships in multiple shipyards, so he could use the fact that shipyards started at different times to have a more causal story. Economists love having that extra little hint. He was able to show that this learning that was observed—the fact that the man-hours needed to complete a ship were decreasing over time—was not a consequence of changes in technology, an increase in capital expenditure, or an increase in labor—that basically, more people were working on the ships—but was a function of experience: how many ships your shipyard had already built. So that provides evidence of learning.

Now, what happens is that the phenomenon is true only at the level of individuals, teams, firms, and so forth. Once you transition to the industry level, you get to Moore's law, which is very different. It's qualitatively different; it's exponential. Part of the book focuses on explaining the connection between the 2.

We know from experience, right, that when we have experience, we get better at things. And we have this weird—I don't know whether it's an illusion—that all we need to do is write down our understanding into a wiki document. It's the same thing with what I do on MLST. I've tried to write down how I edit the videos and how I do the sound design and the video and so on. It just became more and more and more content. I could probably write an encyclopedia about it at this point, and I realized at some point that a lot of it is tacit and it's very, very difficult to transfer to any new staff who come on. It is simply just a function of experience, and this is really depressing to anyone who wants to start an enterprise or a business, because the biggest problem is this knowledge-transfer bottleneck.

Are you saying that that cannot be overcome in any other way than just having lots of experience and lots of people working?

No, I think experiential learning is important to transmit that tacit knowledge. And I think you have stories of people who have that intuition and have been successful because they have developed careers following that intuition. One example that comes to mind is—did you watch Arnold's documentary?

No, no, no, no.

Arnold Schwarzenegger had a fantastic documentary on Netflix—a 3-part documentary that basically goes through his entire life. The 1st part is about him as a bodybuilder, the 2nd part is about him as a movie actor, and the 3rd part is about him as a politician. There's a constant in the documentary: He says, “Look, I wanted to become the best bodybuilder in the world. If you want to become the best bodybuilder in the world, you have to be with the best. So I figured out that the best were in California, so I moved to California and I became the best.”

“Then I wanted to become the best-paid actor in the history of Hollywood. So you have to work with the best. I had money that I had saved from my bodybuilding activities. I had real estate that could keep me alive, so I could be picky about the roles, and I wanted to work with the best.” Eventually, 10 or 15 years later, he becomes the best-paid actor in Hollywood. He accomplishes everything that he has set his mind to. But he's very conscious that the only way to do that is not by figuring it out on his own in a quiet room at the back of his house. It's by trying to make sure that he's with the best.

When it comes to politics, he was part of the Kennedy family. He marries Maria Shriver early on, and he learns from the Kennedys for more than a decade before he decides to run for governor of California. So again, you learn from the best.

The example I have in the book is that of Samuel Slater, who is a local lad, and he's truly a hero. This is a guy who is born in the Midlands at the time that the Midlands were the place that had, for the first time, figured out how to do water-powered cotton spinning. That was a devilishly difficult technology. The first patents for water-powered cotton spinning are from the 1730s. They tried to build a mill in Birmingham; it doesn't work. Another in Northampton doesn't work either. About 50 years have to pass after that for people like Arkwright and Strutt to develop water-powered spinning, first in Cromford, which is a very small town but had water power, and then eventually they created mills in Derby, Belper, and so forth.

Now Samuel Slater is born at the time that these mills are first being erected, and he joins one of Strutt's mills at the age of 14. He's very smart, becomes an overseer, and at the age of 21 says, “Okay, I know this business, and I know that I'm not going to make it in this business because this technology is just spreading like wildfire. Everybody figured out how to build these mills, but in the US they have not figured that out.” So he escapes Belper in the middle of the night without telling a soul. He goes into London. In London, he boards a ship pretending to be a farmer, and he lands in New York 66 days later.

He goes into a mill that was in Manhattan. He immediately sees that the machines were no good. They had no water power, so he quits after 4 days. Then he learns from a sloop captain that there was a man in Pawtucket who was trying to develop water-powered cotton-spinning technology, but they were not able to produce yarn of good enough fineness and strength. The thing about cotton yarn, like the one that we have in our jeans, is that to resist the tension of the loom, it has to be very well spun. If you manually spin cotton, you cannot produce jeans or fabric of that type because it would just snap under the tension of the loom.

No. So he moves to Pawtucket and eventually develops the first water-powered cotton spinning technology in the United States within a period of about a year. It starts the American Industrial Revolution, and mills start to spread there just like before.

So it’s a very good example of the embodiment of knowledge. The people in Pawtucket had tried to develop water-powered cotton spinning based on hearsay. There is this story from the book where I got that story, that there were some Scotsmen who had seen one of Arkwright’s mills and had told these other guys how they worked. But based on that hearsay, they were not able to develop it.

You had to have someone who had that experiential knowledge, someone who had worked with the best—with Arkwright and Strutt—to come all the way to the United States in an act of treason. It was a punishable act of treason to bring that technology to America, but he eventually built that capability.

You’re strongly asserting that there’s a huge physical component—that there’s an embodiment to the propagation of knowledge, the way it flows and decays. Everyone can access GitHub, and people on the other side of the world can start playing with software. Do you think that, at some point, at least, the propagation of knowledge becomes more virtual?

I think there are 2 things. One thing is to be precise about what we mean by physical. Everything has to be physical because even GitHub has to store its data in some sort of hard drive, magnetic field, or whatever technology. It’s not storing it in nothingness, so knowledge and information always have this form of physical embodiment.

I think we tend to think about it as nonphysical because it is a thing that is not a thing, which is the same as temperature. In the book, I have a chapter in which I tell the history of temperature. Temperature is kind of funny because today you wake up, look at your phone, see the temperature, and decide how you’re going to dress. Nobody has any doubt that temperature is something that can be measured.

But it took about 2,000 years for us, as a species, to figure out what temperature was and the fact that it could be measured. There were 2 fundamental difficulties that made it difficult for us to understand temperature.

The first one is that people initially thought that hot and cold were 2 separate things, so that temperature was a mixture of the 2, like when you make green out of blue and yellow. It took a while for people to understand that cold was the absence of heat, and not that cold and heat were 2 different quantities that were tempered together. They were mixed. Temperature actually means mixture, not what we now mean by temperature.

The other thing that was very difficult to understand is that people thought temperature was a thing, some sort of fluid that grabbed onto things. Let’s say you had a steel rod that was hot. They thought that the steel rod had this sort of invisible fluid, heat, and they had good reasons to believe that it was an invisible fluid because it could flow.

Let’s say you could connect that rod to something that was cold. That cold thing was going to warm up because the fluid was going to flow in that direction, and so forth. They thought that it had a physicality as a thing.

A brilliant Englishman, Joule, basically figured out that this was not the case—that temperature was not a thing. The way that they did it was through an observation involving how cannons used to be built. I don’t know if you know how cannons used to be built, but if you just grab a piece of sheet metal, make it into a cylinder, and try to make a cannon out of that, the exact moment that you shoot the cannon, it’s going to open up like a flower in a cartoon, like a Looney Tunes situation.

So what they would do was make these solid cylinders of metal and bore a hole in them to create the cannons. Boring those holes released an enormous amount of heat. Joule thought, “How come all of that heat is there? It’s like an infinite amount of heat. If I continue to bore a hole in a piece of metal for an infinite amount of time, I’m going to get an infinite amount of heat. It cannot be a thing, then.”

That leads him to realize that temperature is actually something that has to live in things, but it’s not a thing itself. It’s related to the kinetic energy of the particles in the thing, but it’s not a thing itself. It doesn’t have its own particle. There isn’t a temperature particle. Temperature is a property that matter has and that holds on to things.

Knowledge is similar in that it holds on to you and me, and to the collective, to exist, but it doesn’t have a physicality in itself. It always exists in some sort of physical medium or substrate. In that sense, it’s always going to be physical. No matter how virtual it gets, it has maybe a different type of physicality.

Even electromagnetic waves that are transmitting data from your Wi-Fi router to your laptop are technically a physical embodiment.

There’s an interesting perspective here. David Krakauer says that temperature is an intensive property of matter, and he says intelligence is an extensive property. I think when he says intelligence, he’s actually talking about knowledge.

He says that when we look at these complex adaptive systems, there are 2 types of systems in the world. There are the sort of Roger Penrose, symmetry-dominated systems, and then there are the systems that break symmetries, which are these complex systems—things like life and evolution.

The amount of information they’ve accumulated in their lifetimes is a good proxy for the amount of intelligence they’ve had. You’re basically saying that that accumulation of information, roughly as a physical property of matter, is how we should think of knowledge.

I’ve thought a lot about whether knowledge is intensive or extensive, or whether complexity, as we measure it in the technical literature, is intensive or extensive. The key insight is the following: Let’s say you’re wondering whether putting together 2 countries would lead to having more knowledge than having those 2 countries separately. You’re going to proxy knowledge by the specialization that these countries have in the activities that they perform.

When you put those 2 countries together—let’s say they’re 2 developing countries in Africa that are specialized in a few activities—they’re going to be specialized mostly in activities that have low knowledge intensities, and a few that have high knowledge intensities. When you put them together, you’re going to realize that the high-knowledge-intensity activities, which are the ones that are not in common, may get subtracted because now they’re not specialized. They’re not producing enough to justify the now-combined area, and therefore the complexity of those economies, when they’re put together, doesn’t go up.

An intensive quantity is one that, when you put 2 units together, averages out. An extensive quantity is one where, when you put 2 units together, you add them up. Knowledge has a little bit of that intensivity that depends on complementarities.

If you put 2 people together who know the same thing, you don’t get twice the knowledge. In that case, it’s clearly not extensive because it would be redundant. If you put 2 people together, one who knows the same thing and one who is really dumb, and that one is okay, you don’t get a super-smart person as a result. You get a half-dumb person, maybe. You average them out.

For that extensivity to kick in, you need to have a good level of performance, but also complementarities. You need to be able to put things together that, when they’re put together, are more than the sum of the parts. That’s not always guaranteed.

You have a lot of examples in which putting things together is going to give you an intensive result or property. You need that diversity and those complementarities to have that extensivity kick in.

For example, if I take a Chinese TV show and try to launch it in the States, it won’t work. It’ll fail because the cultural fit doesn’t make sense; there’s no shared phylogenetic history.

This idea that we can just take random bits of knowledge and stick them together, even if they have different histories, doesn’t make sense.

No. Yeah, you reminded me of something. I’ve been in the UK promoting the book for about 3 days now. The book is about the laws that govern the growth, diffusion, and value of knowledge, but there’s 1 chapter about the laws that govern forgetting.

Yes.

It’s the last chapter of the time section of the book.

Okay.

Yes, and that’s the one that has resonated the most with people here. It connects to what you’re saying because, after telling some stories from that chapter, I had a colleague from here in the UK send me an email. He said, “This reminds me of the Ise Shrine story.”

I thought, “What’s the Ise Shrine story?” So I looked it up and read about it. There is a shrine in Japan that they rebuild every 20 years. By rebuilding the shrine every 20 years, they basically train the new generation of people who are going to know how to build a shrine, and that generation is going to have to train the next generation of people in 20 more years.

So if you think about it, you can think of a temple as a structure, and you might want to preserve that structure. Here in Europe, we have this historical heritage and so forth. But in this particular example, they're deciding not to preserve the structure. It's not that this beam here is 300 years old or 500 years old. What they're preserving is the knowledge of how to rebuild the structure by keeping the muscle fit, by continuing to do the activity over and over again.

Our current state depends on everything that went before. That's not really true, because so much knowledge decays and gets lost. There was this interesting thought experiment: What if we could have a parallel universe? We could do a simulation and just skip alchemy. Would we still be in the same place now? In a sense, do you think that garbage collection, this deliberate kind of pruning of the knowledge tree, is actually a feature of evolution?

It's hard to tell. What we do know is that knowledge decays, and we maybe don't tend to focus on that as much because it's a more depressing story. But we know that knowledge, for instance, in the case of the Liberty ships, was estimated to decay about 3% to 6% per month. You might think 3% sounds like a little bit, but it's 50% a year. That means that if an organization were to stop working one day and wanted to resume after a year, they would have lost 50% of the knowledge. So knowledge decays really fast.

The reason why we don't observe that decay as frequently is because, of course, knowledge is being accumulated in a way that offsets those decays. But you have very good examples of knowledge decay, and one of them is the history of Polaroid.

Polaroid was an amazing company created by a quintessential Harvard dropout and entrepreneur, Edwin Land. It's called Polaroid because what they did was polarized film. They didn't start with instant photography. They started doing polarized film. Land figured out a way to create polarized filters that were large enough to have an industrial use and application by stretching this goo with this material in a magnetic field so that you could create the polarized filters.

After the Second World War, they needed to find a new business model because polarized filters were usually sold to the military. So they came up with this idea of doing instant photography, which was devilishly difficult and involved tons of patents, because the chemicals that you use to develop a photograph are extremely reactive. If you put them on an envelope that people are going to be shaking around and transporting around the country, most likely they're going to react at some point. They might mix, and your envelope becomes worthless. They developed extremely high-quality instant photography, so photographers like Ansel Adams and so forth would use Polaroid cameras.

In the 1990s, Polaroid failed to transition into digital. In the late 2000s, there was 1 plant left in the Netherlands producing Polaroid film, and there was a man in Vienna with an online shop that sold vintage film and so forth. When he found out that this plant was going to close, he flew to the Netherlands and got a group of people together to acquire the plant, because Polaroid was simply going to close it. They said, “Why don't we, instead of closing it, sell it to ourselves?” They sold him the plant, and they tried to restart production of the technology.

They had access to the factory and to all of the original equipment because they hadn't dismantled the factory. I interviewed him when I visited Vienna a couple of years ago. I asked him, “The people that you hired to work on the plant, did you get your pick, or did you have to deal with whoever wanted to work there?” He said, “No, no, no. We had the A-team. It was the star team. We couldn't hire everyone back, but for each machine, we hired the top player that we had.”

Still, after they ran out of the stock of film that had been produced previously and had to start selling their own film, it was black-and-white film. It took 30 to 40 minutes to develop and often had aberrations. The knowledge was lost, and it took this impossible project several years to start producing film of any quality. Maybe only about a decade later did they start producing film of a quality that could be considered comparable to the one that Polaroid was producing in the ’70s.

So knowledge can really disappear rather fast if we stop exercising it. If you don't use it, you lose it when it comes to knowledge.

No, and it's so fascinating, that example. You use the term “embers of knowledge.” In that particular case, it was almost lost. The embers were there, the flame had gone out, and it was possible for them to get the folks back in from the factory. What they discovered was supply lines they couldn't access—some of the chemicals and some of the other things. So what they had to do painstakingly was almost reinvent some of the knowledge.

You use this wonderful term “absorptive capacity.” There was the story of the US after the Second World War, when they were in Japan, and there was this inventor in Japan. He studied one of these tape players in great detail, and he could go away and create a prototype using a frying pan, shellac, and hemp-reinforced paper. That's Ibuka, who was one of the technical co-founders of Sony, which basically developed magnetic tape using a frying pan, a badger-hair brush, reinforced paper, and shellac, after observing it only a couple of times.

Exactly. And because this scientist and his collaborators were researchers and had lots of adjacent experience, it was possible for them to reignite this knowledge and almost reimagine it and resituate it in a different environment.

But just in terms of this knowledge-decay thing, isn't it amazing that we think that we are at the pinnacle? We think that we've accumulated so much and we haven't lost anything. There are so many examples. Concorde is a great one. What if we wanted to build Concorde again? Would we just be able to, 30 years later, pull up the blueprints and stitch this aircraft together, and it would be just like Concorde was 30 years ago? Probably not.

Another example is in software engineering. There are famous examples of, I think, IBM publishing their source code online, and everyone in the company said, “No, no, you can't publish the source code online. This is all of our knowledge. If someone gets hold of this source code, they're just going to be able to recreate everything that we've done, and they'll know everything.” That's not true, because the actual knowledge is just in the minds, the interactions, the ecosystem, the organism of the company.

You also said in the book, which I thought was fascinating, that the purpose of an organization is to retain as much knowledge as possible.

Yeah. To retain and preserve knowledge. There are beautiful examples about the role that large teams play in the development and diffusion of knowledge.

Before, I told you the story about Sam Slater. This kid, at age 21, starts the American Industrial Revolution. But there's another story later in the book: the story of the city of Donetsk.

Most people have learned about the city of Donetsk today in the context of the conflict between Russia and Ukraine. But Donetsk is a city that was actually created by a Welshman by the name of John Hughes. John Hughes was a Welshman and a successful entrepreneur here in the UK. He was in ironworks and in the business of ironcladding ships.

You have to remember that in the 19th century, wooden ships were being ironclad so that they would be more resistant to damage when you had sea battles. Later in life—this is not a story of a 21-year-old like Sam Slater—when he was in his 50s, he went to ironclad a ship for the Russian Empire. Then he developed a relationship that led him to apply for a concession to develop the coal and iron resources that were available in this area of the Russian Empire, where the city of Donetsk would eventually be.

He won that concession, came back to the UK, and loaded 7 ships with more than 100 men and all of the equipment that they would need to set up that operation. They took those ships and sailed all the way to the Sea of Azov. They unloaded and then dragged this thing through the mud into what later became Donetsk. They set up a camp and started building the ironworks from scratch in that location.

In about 3 years, they were producing pig iron. Then they started developing what eventually became one of the main iron- and steel-producing regions of the Soviet Union. They built the schools, they built the hospitals, and eventually all of that. The city was originally called Yuzovka, or Hughesovka, because it was John Hughes.

There were some people who, when the conflict started, said, “We should secede back to Britain because we are British.” This was a city that was created by the British. But the whole point is that Hughes understood that if he went only with the knowledge that he had to that area and tried to set up ironworks on his own, he was not going to succeed, because the knowledge had to be embodied in a much larger crew—a crew that required 7 ships to be taken from England to Ukraine.

It’s almost as if there’s a sufficient embodied carrying capacity for knowledge. It reminded me—I don’t know if you’ve seen the Apple TV series Foundation. It’s the Isaac Asimov Foundation, with the genetic dynasty and all of that. Again, this is science fiction, but they still had this possibly wrong-headed idea that all they needed to do was take a library of knowledge and a small group of people, put them on a planet in the outer galaxy, and then they would be able to restart civilization.

You would probably argue that there’s actually a threshold point where it’s just not possible to do that.

If that were true, shipwrecks would be much more successful than they are.

So, coming back to this, you were describing this kind of asymptotic curve of learning. It’s a function of experience, but then, in the next chapter, you went on to talk about disruptive technology. You gave a great example with steel. To manufacture steel, there are roughly 4 or 5 grades of steel, and what happened was that as companies started to develop higher grades of steel, entrants would come in and develop lower grades of steel that were kind of crap, but they had more upward trajectory. That would create this disruptive cycle, which wasn’t this limiting curve but was actually a bit more like Moore’s law. It was more like an exponential curve. Explain that.

That’s the idea of disruptive innovation by Clayton Christensen, and it’s the connection, in some way, between Moore’s law and the laws of S-curves and Wright’s law. If I can draw in the air here, we had curves first that are a little bit like a square root: they grow fast in the beginning, and then they slow down and peter out. Then you have Moore’s curve, which is an exponential that grows over several orders of magnitude, and you have to reconcile the 2.

The way that you reconcile the 2 is that you realize that when you’re operating at the industry level or at a geographic level, like at a country scale, and over long periods of time, you don’t have the development of a technology that involves a single learning curve, but a collection of generations of technology that have multiple learning curves. Moore’s law is like that envelope that captures that collection of other individual learning curves.

For example, when you look at the manufacturing of LCD panels, they go through generations, and each generation is able to produce panels that are bigger and that have fewer defects, and so forth. You have lots of examples of that. What’s interesting about that is that when you move from one learning curve to the next, the next learning curve, even though it can soar higher, starts at a lower point because you’re at the beginning of it.

When new technologies are introduced, they’re worse than the incumbent technologies. When digital photography was introduced, it was much worse than chemical photography. People in the 1980s who were involved in chemical photography, like the people at Polaroid, laughed at digital photography and said, “These guys are never going to be able to get the type of colors and resolution that we’re able to get with our chemistry.”

The technology gets laughed at because it was worse. Instant photography was also laughed at at some point because it was worse. But eventually, that curve keeps on growing, and it gets to a plateau that goes even higher. Then a new technology comes along and moves into a new curve that is higher.

Every time you have that intersection, you have a window of opportunity, because in that window of opportunity, you have a technology that is worse, that the incumbents don’t take seriously, and that is going to be able to surpass them. The moment that those 2 curves cross is when the incumbents get desperate and think that they’re going to be able to get there.

Usually, those changes also involve architectural innovation, which is what we discussed earlier. Moving from chemical photography to digital photography would have required completely redoing and restructuring the entire operation of Polaroid, because they weren’t set up to do that. They were a chemical company that started with polarizing film. They were heavily into chemistry, not electronics.

You also gave the example—I think it might have been the Sony Walkman, or a Sony technology—where they came in and originally were much worse than the incumbent, and then it had more upward trajectory.

The transistor radio did not have as good a sound quality in the beginning as tube radios. It was a cheap alternative. It was a radio for a security guard who wouldn’t be able to have, of course, a tube radio in their booth, so they would have their little pocket radio that was low quality. Then transistors eventually became good enough, and nowadays we can get amazing sound quality from transistors. Yeah.

You spoke about transistors because this is like the Moore’s law thing. I think it was originally 12 months, and now it’s been set to about 18 months or something, but it’s still holding, which is amazing. In a way, I’ve got a few questions here. First of all, it sounds like this appeal to infinity, that we can have these exponential curves and they just keep growing and growing.

It’s interesting that Moore’s law has kept on, and you said that the reason for that is they have all of this disruptive technology. They’ve got larger and larger teams, better and better processes, and they can just keep innovating. How long will that go on? What is the reason why we can have these disruptive cycles?

What we were saying before about Barnes & Noble is that it’s really difficult for an organization to do this architectural rewiring. The way that disruptive innovation happens is that usually you have an independent thread. Somewhere else in the epistemic ecosystem, you have different people with different ideas and different objectives, and they come up with a completely different architecture and innovate.

But now we’re in this domain where we have these large technology companies, and they might acquire any new startups. That means the new startups will be contaminated with all of the cultural knowledge of the incumbent. Don’t you need to have independence, diversity preservation, and competition to keep us on this exponential curve?

At least in the history of the transistor, I do think we have some good examples that the teams required to develop those innovations have been growing over time. This is something that Nick Bloom at Stanford—he’s a professor of economics there—has emphasized: the fact that there’s an increasing cost of innovation. To innovate again, we have larger teams and larger budgets. It’s getting costly and costly and costly.

Some people have arguments against his evidence and so forth, but I think it’s the right question to ask because there definitely looks like there’s something in that direction. You can see it in the history of the transistor. The first transistors were developed by a team of 3 people—originally, actually, a team of 2. It was Brattain and Bardeen, and then Shockley got jealous and, over Christmas, developed a transistor design that replaced the first transistor, which had been developed by his lab assistants, Brattain and Bardeen. That became the point-contact transistor in 1948.

Then you had several different transistor designs. For example, Shockley Semiconductor Laboratory was not successful at producing transistors. But when Moore moved to Fairchild, they produced the mesa transistor, I think in 1958, and then a planar design in 1959. They produced an integrated circuit in 1959.

There was also another team at Texas Instruments that produced an integrated circuit: Jack Kilby. Jack Kilby had just joined the company, and because he had just joined the company, the company was kind of empty. During the summer, he had nothing to do, and he created the integrated circuit as an experiment on his own. You see, it was a team of 1.

Nowadays, to produce the next generation of NVIDIA CPUs or Intel CPUs, or whoever, you probably have enormous teams involved in the design and manufacturing. I think that might be, at some point, the limiting factor. At some point, maybe, to double again, the teams are going to get larger than what we’re able to coordinate.

If that coordination capacity doesn’t scale to the amount of knowledge that we would need to generate another doubling, we might see this curve petering out. Is that close to happening? I don’t know. It’s hard to bet against a law that has been stable for so long.

I would love it if you read Kenneth Stanley’s book, *Why Greatness Cannot Be Planned?*. He’s a good friend of mine, and I’m going to tell him to read your book because I think there’s a lot of overlap. His basic idea was that consensus, committee meetings, and objectives are actually quite toxic for progress, because creativity is about following your own gradient of interest and basically preserving diversity and having new ideas about things.

So, yes, we have a new generation of these large language models. They’re getting 10 times bigger every few years.

The bullish people say, “Oh, we’re on an exponential curve, and it’s just going to keep going up.” But I think that, because there is so much groupthink and it’s fundamentally the same technology and the same people, with no fresh new ideas, in a sense it’s converged and it’s not disruptive anymore.

I remember using GPT-3, not even ChatGPT or GPT-2, and this has definitely been going on for a while. The improvement has been sustained not for 5 years, but for longer.

Yeah, these technologies started at a rather simple level of proficiency, and that has continued before people knew about them. You know, instant photography—those are the Polaroids before Sony makes it big with the transistor radio, and so forth.

So, yes, I agree that there might be incumbents that are really big. Those are like the Barnes & Noble bookstores and the Walmarts of the tech industry, and there might be disruption that comes from teams that have thought of a different way of creating artificial intelligence that might overtake those incumbents. That might be disruptive because maybe it requires doing things in a different way. But we’re not going to know until those curves cross. Usually, it’s very hard to see those disruptors early on simply because they’re not yet getting the attention.

Let’s talk about the flows of knowledge.

Okay.

So, this was Chapter 6 of your book. You had many wonderful examples of, let’s say, migrant flows, for example. One example that really sticks out to me was this trade embargo with Vietnam.

Vietnamese boat people, sorry.

Tell me about that.

Yeah, that’s a good story. When the United States left Saigon in 1975, it was as glamorous as when they left Kabul a few years ago. It was rather quick; they expected to have more time to evacuate. What happened is that the Vietnamese Army came into Saigon rather quickly, and lots of people went into boats and went into the ocean there. They needed to be relocated to the United States. These were people who had cooperated with the United States, and now they were looking for asylum.

The United States had to relocate hundreds of thousands of people within a period of a year. That relocation can be considered quite random or exogenous because it wasn’t, “Where do you want to go? Would you prefer to go to New York or to San Francisco?” No, it was like, “We have a small town in Iowa that has a church that is willing to take 10 people. Okay, here we have 2 families of 4 and this couple. Boom, Iowa.” They would be relocating people at an enormous speed because this was a humanitarian crisis they were trying to solve. There were different takers, and they were being allocated like that.

After the war is over, what also happens is that the United States imposes an embargo on Vietnam, so they cannot trade. What these economists, Parsons and Vézina, did is they looked at the stock of Vietnamese people that was exogenously allocated through this process, and then they looked at trade data starting in 1995, when the embargo was lifted. They said, “Well, if this state got more Vietnamese people because of this exogenous allocation, did they trade more with Vietnam after the embargo was lifted?” The answer is yes. There is an effect there. They show that, in this case, the relocation of these people brought knowledge on how to trade with Vietnam, and they had relationships that they could use to develop that commerce.

There was also an example—you use this analogy of monkeys in a forest to talk about not only the geography of knowledge but also the geometry, the structure, the topology of knowledge. Talk to me about that.

Yeah, exactly. The second law of knowledge is about diffusion, and diffusion has 2 subprinciples. One is about diffusion across geography or across social networks, and the other is diffusion that is constrained by the geometry of knowledge itself.

One way to explain that idea is to say, “Look, the economy involves multiple activities, like multiple industries or multiple products, and you can think of each one of those activities as a tree in a forest.” You have multiple trees in a forest. Let’s say this is a tree that represents shirts, and nearby you have a tree that represents blouses. Those are very nearby trees because shirts and blouses are similar products. If you produce one, you can produce the other. Then maybe far away you have a tree that represents natural gas, and on the other side you have a tree that represents tractors or combustion engines, and so forth. You have a kind of geometry or geography of knowledge itself.

In that representation, if knowledge is represented by the activities that an economy can produce, and these activities are trees, countries are collections of firms, which are collections of monkeys that live on these trees. Let’s say you have a garment company, and all of your monkeys—all of your knowledge—is being harvested from the fruit that is being grown in the tree of blouses, the tree of shirts, the tree of linens, the tree of socks, and so forth. If you have an electronics company, you are in a different part of the product space. Economic development is the process by which countries move into new trees, but the ability of monkeys to jump from one tree to another depends on their distance.

What we have shown—and after we did it, it has been shown by hundreds of papers—is that the probability that you would enter an activity that you were not specialized in in the past depends on how many monkeys you have around in other trees. This geometry really matters, and the story that I use to illustrate that, which I think is the most telling, is the story of Vespa.

Everybody knows Vespa; it was even in a recent Disney movie. It’s an Italian scooter that is iconic, but people don’t know that it was created not by a motorcycle engineer but by an aircraft engineer. His name was Corradino D’Ascanio. Corradino is the equivalent of Theodore Wright, but for Italy. He was in charge of the production of aircraft for Italy during the Second World War as well. He was working in a factory owned by the Piaggio family, which was also specialized in the manufacturing of aircraft.

When the war is over, 3 things happen that change the destiny of the Piaggios and of Corradino. The first thing is that Italy is no longer allowed to manufacture aircraft. That’s forbidden. The second thing is that the factories had been bombed. Aircraft factories are primary military targets. They’re not hospitals. They’re not schools. They’re weapon factories, so they got bombed. All of that capital, let’s say, was destroyed or whatnot.

The third thing is that the bridges and the roads had also been bombed and destroyed. It was a war zone, so people in Italy needed a way to move around. They wanted to arrive at work without being all muddy. In that process, people started thinking about vehicles that they could create to satisfy that need.

First, Corradino goes to work with Innocenti. Innocenti was a competitor that wanted to make a motorcycle based on tubular metal, and Corradino wanted to use sheet metal. They didn’t get along; they had a falling out, and he went back to Piaggio. Then they created a motorcycle in which the mudguard is the body, the engine goes in the back, so you can sit with your feet together, and the wheel in the front comes on and off the same way as it would come on and off in a helicopter.

Now, you might think that’s a nice anecdote: These guys were manufacturing aircraft, and now they cannot manufacture aircraft anymore, so they make motorcycles. But if you go to Japan or Germany, you see the same example. Miles and miles away, you see companies like Kawanishi going into light-vehicle manufacturing after not being able to produce fighter aircraft anymore. Heinkel in Germany—the same story. They produce the Heinkel Tourist when they’re not able to produce aircraft. In their case, they were even able to produce jet aircraft during the war.

What that tells you is, when push comes to shove and you have to get out of your industry—in this case, these guys had no option; they could not stay aircraft manufacturers—they had to jump into something else, and they all jumped into something similar. If they’re all jumping to something similar, it means that they’re moving along the same map. It’s a map in which motorcycles were close to aircraft, and aircraft maybe was not close to blouse manufacturing. So, they all ended up in the same parts because the diffusion of knowledge is also constrained by the geography of knowledge itself.

There are signals for knowledge diffusion. Prestige is one signal. There was an experiment where, if children see a particular person as prestigious, they’re more likely to transfer knowledge. But what I want to get to with creativity, though, is that a lot of people think it’s completely serendipitous. YouTube started as a video-dating website, and the way microwaves were invented took quite a divergent and weird and wonderful path. We might conclude from that, “Oh, it’s just a random walk through epistemic space,” and that’s not true at all.

There are these analogical jumps that you can make when you perform creative actions, and I think the examples you gave just pointed to that. So the jumps are still very much dependent on the history, but they are obvious stepping stones that can be taken from the previous position. Does that make sense?

Yeah, yeah. I would agree that, for example, a video dating website and a video streaming website are quite related. They are trees that are close by in that product space.

What is interesting about this story is that, since this principle has been so well established in the economic geography literature, we now talk about strategies that involve, for example, targeting related or unrelated activities. We know, for instance, that migrants are better at developing unrelated activities, while locals are better at entrepreneurial activities that are related to the current economic structure. So we even have ways to connect these 2 stories. Migration is going to be something that is going to help you jump far in the product space, like the migrant monkeys might land in a tree that is not near your trees and might help you develop a new area.

Yeah. Actually, you spoke a lot about migration in general. First of all, migrants have a lot of choice about where they can go, because you started the book talking about NEOM in Saudi Arabia, and there was Yachay. Exactly. These were places that were designed to be epistemic centers where they could do lots of innovation and invent lots of things.

Actually, it didn't really work because it didn't respect the ecosystem. They were so disconnected that no migrants would want to go there, and it can't actually be self-sustaining in terms of the creation of new knowledge.

The people who have a choice tend to be quite strategic about how they exercise those choices. There are a lot of statistics about the role of migrants in innovation and the fact that people who are highly creative or highly innovative tend to be migrants. There is this statistic—I think in the United States, if you count Nobel Prize winners after the 1970s, about 70% of them were born outside of the country, or they migrated at some point in their lives. There's something like that.

60%.

Yeah, the higher the level of education, the higher the propensity to migrate. That's something that we know for a fact. But migration is not random. These are people moving to centers of knowledge where they know that they're going to find the complementarities they need to develop the things that they're looking to develop. So being attractive to high-skilled migrants is a very good signal for an economy.

I was talking with people at the innovation agency of the UK, Nesta, the other day, and one of the things I was saying was, “Look, if you want to look at the impact of migration on innovation in the UK, what you need to count is the number of superstars that you attract,” because in the end you worry about that tail end of the distribution.

I know migration is a topic that is quite controversial right now around the world, and I don't make a call in my book for a single way of thinking about it. But I do think that I show evidence that migration is not all equal. If you're able to attract people who have a high level of skill and talent, the probability that they would be net contributors to your economy and to your society is going to be much higher. So it's not, “Let them all in and we'll figure it out later.” But how can you become attractive so that the best people in the world are fighting to be working in your cities, in your universities, in your companies?

For example, if you look at genetic diversity in Africa, it's much higher, and that actually is a proxy for where the cradle of evolution was, because there's so much diversity. Similarly, you can look at the diversity of innovation and knowledge as almost a proxy for where you are on the maturity curve in evolution.

There was this really interesting concept. By the way, as a point on what you just said, I think there was a statistic that something like 50% of the large companies in the US were started by migrants. I can't remember the exact statistic, but as you say, when you have skilled migrants, it should be obvious that you have this kind of diversity acquisition, and you need to have diversity to try new and interesting things, to not get caught in the basin of attraction that you're currently in.

This seems obvious to me. There is a lot of nasty talk about migration at the moment. But migration is clearly a very, very good thing. Maybe you want to comment on that first.

No. Yeah. No. Migration definitely is an important vector for knowledge diffusion, and that's one of the best-established facts in the economic geography literature. But we have to keep in mind that those facts are always established using data that focuses on very high-skilled individuals.

They're looking at migration by looking at people who patent—what fraction of the population patents—or people who publish papers. That's a larger group, but still it's a very elite group of people.

We have other examples, like one of the Vietnamese boat people. In that case, I put that example in the book because it's a non-elite example of knowledge diffusion. But the impact per capita of the migrant might be less than that of, let's say, an inventor who has 20 patents and is going to produce 20 more in the next 10 years. So you do have that differential aspect.

Now, I wanted to go into your diversity in the context of Africa, because I was in Rwanda a couple of weeks ago, actually. It's a fantastic country, honestly. I was very impressed by what they've been able to accomplish. The country is clean, people are nice and respectful, and they're really pushing forward and so forth.

But one of the things that you notice when you go to a developing country—I do a lot of development work, and I travel around the world quite a lot—is that there are many people doing the same thing. So, for example, in Rwanda, taxis are mostly motorcycles. They're moto taxis. These guys are going around on motorcycles. They're carrying a couple of helmets, and you put on a helmet, sit on the back, and that's your taxi. There are parts of Kigali where you have tons of moto taxis.

So it's a lot of people, but if you were to add, let's say, the diversity of that knowledge based on the different activities that they do, you wouldn't add that much because it's the same activity. And then when you go to other places that are knowledge-intensive, what you have is that everyone is specialized in something different, and they tend to be complementary.

So it goes again into this idea of whether knowledge is extensive or intensive. If you put 100 moto taxi guys, the amount of knowledge that you add from 1 to 100 is not that much. But if you then put 100 people who do different things and who are complementary, you could get an amazing amount of knowledge. Whether it is extensive, whether it adds up, depends on whether there are complementarities and differences.

Very cool. Now, in Chapter 8, you started talking about Bretton Woods, which was this thing after the Second World War. Essentially, to rebuild Europe, they set up these institutions and kind of did financialization to encourage growth. That was seen as such a success that financialization was actually seen as something that could be used to aid development even when it's not a war-torn country. Tell me about that.

The story of, let's say, 20th-century Western institutions that is told more commonly is that after the Second World War was over, the United States had an interest in helping Europe recover. They generated different funds and institutions to help do that. There is a German Marshall Fund, the World Bank was created, and the IMF was created, all in that context. The World Bank started lending money to the Netherlands, to France, and to different countries.

Those development efforts did really well because the money that came into Europe at that moment translated into investment that helped rebuild Europe. Europe started to pick up, started to grow, started to rebuild, and so forth. Then people said, “Wow, development is a really easy business. You just throw money, release the constraint of finance, and it just happens.”

Now, the thing is that Europe is extremely knowledge-rich. They were throwing money into a continent that was extremely knowledge-rich. Sure, the bridge was not there, the hospital had maybe been partly destroyed, and so forth, but the knowledge was still around. So if you liberated the financial constraint, you were going to get that rebuilding happening, and the GDP was going to start to return to what it was supposed to be.

Then they started to think that you could develop other regions of the world through a similar model, and the results were not the same. Many things happened. On the one hand, people started to think, “Well, maybe it is about institutions and so forth,” and there were lots of efforts to try to connect financial support with institutional reforms. “Okay, we're going to make you this loan, but you're going to have to do all of these reforms within your country.”

And what happened is that a lot of countries would do those reforms, and the medicine still did not take. Now, there are people who have said that the thing is, they would do these reforms only in form; they would not be real reforms. It would be a kind of mimicry of the reform, and therefore that’s why it doesn’t take.

But there are also people who think that, in reality, it’s not just about institutions; knowledge also plays a role, because the demand for the institutions that you need to develop comes from the more knowledge-intensive members of your society.

So what I do in that chapter is build a little bit of a parallel, because when it comes to economic development, I think the 2 dominant ideas are that it’s about knowledge or it’s about institutions. I’m happy to believe that both play a role, and that sometimes one is the horse and the other one is the carriage. But there have been periods in which you can clearly see a change between the 2.

So I tell a lot of stories about China and how knowledge-intensive workers in China helped demand institutions of intellectual freedom and entrepreneurship that helped develop Zhongguancun, the main innovation district of Beijing. I tell the story of the printing press, which is an excellent example of a technology that enabled institutional change about 60 years later with the Reformation.

And just quickly on China, you said that the China story was quite different. So they didn’t have the institutions of the West. What happened there?

The story of China is a story of a country that institutionally was rather constrained and was very poor, but it had pockets of knowledge, even all the way back in the 1960s and 1970s.

The story that I use to start my description of the Chinese growth miracle, and in particular Zhongguancun, is that Chen Chunxian was a physicist who developed the first fusion reactor in Beijing in the 1970s. He did his PhD with the Soviets, who had invented tokamak technology—this idea of confining plasma in a magnetic field. That is a Soviet technology that he had learned during his graduate studies. He brought it to China and built a fusion reactor, so you cannot say that was a low-knowledge-intensive worker. Building a fusion reactor is probably a little bit difficult.

As a consequence of that, he gets invited to be part of a committee that goes to the United States to see how the Americans were building fusion reactors. He goes to Princeton, he goes to Boston, he goes to Stanford, and so forth. He was expecting that the Americans had these huge factories where they were building the components for their fusion reactors. What he realizes is that they didn’t have these huge companies. Even though the US was so much more advanced than them, what they had were these small companies run by professor-entrepreneurs.

Professors adjacent to these universities would have a company, and the company would be 10 or 12 people who would be building some component, some electronic device, or something that would then be sold to the plasma fusion lab. He goes back to China and starts advocating for that. He goes back to the US to learn more about this because he becomes obsessed with the idea that professors should be allowed to be entrepreneurs and to have outside activities that would help develop their ideas beyond academic research and so forth.

He goes through hell. He gets ostracized, and everybody is watching him because if Chen Chunxian doesn’t succeed, everybody else who had the same idea—maybe with a different application, maybe they were not doing electronics for a plasma fusion lab—would be thinking, “If they cut his head, they’re going to cut mine next.” People who, for example, went on to create Lenovo were watching that, because if he did not succeed, why bother to try?

Eventually, through a set of serendipitous encounters, Chen Chunxian is able to succeed. One of the people championing him, a very smart woman, is married to a journalist who would write a briefing for the Politburo. Together with Chen Chunxian, they write an article about this successful experiment in entrepreneurship in Zhongguancun that makes it to the higher spheres, where Deng Xiaoping’s people were then in charge.

Remember, China had a big institutional change that started in 1978. They grabbed onto this as an example of where they wanted to go. They protected him from the middle management that was oppressing him. When Chen Chunxian survives that, there is this wave of entrepreneurship that gets released.

But the point of that story is that the guys demanding those institutions of entrepreneurship are not guys who are selling oranges at a red light. These are guys who are building plasma fusion reactors. That demand for institutions also needs to come from somewhere, and this is a very good example of that demand coming from knowledge-intensive workers.

Knowledge also generates a demand side for the institutions that you might need to continue to develop that knowledge.

Yeah, and in a sense, there was this incredible amount of latent and locked-up knowledge, and it became released as a result of this process.

But there are folks who are negative about China. They say China just conducts corporate espionage and copies the West and whatnot. That doesn’t really jibe with what you’re saying—that there are incredible amounts of talent and expertise in China, and possibly, going forward, they’re going to become the innovation center of the world. What would you say to people who have that kind of common perspective of China?

It’s tough. Now we’re getting into more colloquial territory, but I live in France, and I have conversations sometimes that surprise me. I’ve been to China many times, so I’m very bullish on China. I say China is not a country. China is a planet.

The West doesn’t understand that because we tend to think in terms of countries, but China is like all of the Americas plus Western Europe put together. It’s a planet, and therefore the diversity that you have internally is enormous in terms of creativity, innovation, capacity, and so forth.

But there’s a lot of skepticism, and I think it’s a kind of—I wouldn’t maybe say xenophobia in an industrial context, which I think is unjustified. These people who usually have those attitudes have never set foot in China, or sometimes even outside the continent of Europe themselves. They might not have the experience that would have helped them open their eyes to these other countries and how they work.

I think China is a country that, for better or for worse, is here to stay. It is growing faster than everyone else, has been for a long time, has taken so many people out of poverty, and is a force to reckon with that you want to be friends with. That geopolitical side of trying to be too defensive about it, I don’t think, is conducive to the growth of knowledge at the global scale.

Yeah, and I think it’s just wrongheaded. I hope the thing that people get from reading your book is that knowledge is not just a bunch of blueprints and things written down on a bit of paper; it’s about whether you’re capable of actually recreating something.

This goes to my theory of creativity. It’s not possible to create something unless you’re in possession of the entire generative process. Understanding is the generative process of new knowledge and new artifacts. If you’re not in possession of that, then you can’t create anything.

So, in a sense, it’s doing them a disservice if they are creating things that work. But I wanted to get to something we just touched on a little while ago. Certainly during your PhD, you were doing all of this sophisticated graph analysis. You might have been using something like Gephi, so you were looking at all sorts of—

Oh, okay, right.

So you were trying to create beautiful partitions of graphs that made sense in a parsimonious way, and it’s a very difficult thing to do. I guess all of this is leading to you trying to come up with analytical or statistical representations of the accumulation of knowledge, and you created this interesting link with complexity and entropy. Can you tell me about that?

What makes the study of knowledge difficult is this idea that it is nonfungible. When you look at the laws of knowledge that have been studied the most and for the longest, that principle is not accounted for.

When you look at learning curves, eventually, it’s learning how to type. You don’t have knowledge about different things; it’s about one thing. When you look at Moore’s law, it’s transistors. There’s nothing there about other technologies. You might have the same law in other domains, but it’s a within-knowledge or within-domain law.

When we look at relatedness—the story of the monkeys and the trees—now we have some specificity of the knowledge. The monkey is in this tree, and this tree is close to these other trees and not those other trees. So nonfungibility is playing a role.

Then you have another question, which is: If knowledge is nonfungible, if it’s kind of like the letters of an ever-growing alphabet, you need to figure out a way to count the letters that a country or a city might have from that alphabet as a way to score their game of Scrabble, or as a way to get an estimate of the value of the knowledge that they have.

And how do you do that? Should all letters count the same? In Scrabble, they have different values. A Q gets more points because it's harder to use, because it's rarer, and so forth. So how do you do this?

What I did—and I think originally it was a little bit serendipitous, but eventually we developed all of the formal models to validate it and prove it—is that there's a smart way to estimate the number of letters that a country will have in the alphabet by extracting vectors from matrices of specialization that are carefully normalized.

If you grab a matrix that tells you which country exports which product, and you normalize it in such a way that you take away the effects of country size—because larger countries have some extensivity to things—products that have larger markets are going to make that column very heavy, with a lot of big numbers. You have to take that into account because there's a lot of heterogeneity in geography.

If you normalize things properly, you can then extract a vector that we can show through a deductive model is a monotonic estimate of the number of different letters that a country would have in the infinite alphabet. What's interesting about that vector is that, once you estimate it, it's very good at explaining future economic growth.

It tells you that the countries that have more letters than you would expect based on their income are going to grow. For example, when you run that model right now, what it tells you is that the country to bet on right now is not China, because China is already quite rich. The room that they have to grow is less. They're slowing down; they're going to be growing at 4%.

India should be the next rocket, and Indonesia and the Philippines are the next rockets. But not other countries that are similarly poor in other parts of the world, because for those other countries, we don't expect that growth. We don't observe the letters based on the products that they export or the industries that employ individuals, and so forth.

Yes. I don't think we explained this concept of the infinite alphabet, so we should do that as well. In a sense, when you apply this analysis, you could look at somewhere like Kuwait, and it might have quite low diversity and complexity. Then, for a place like India or the UK, you would look at that entropy almost as a proxy for the productive future potential of that economy.

Exactly. It's a measure of potential. This complexity is a measure of potential because if you have an idea of how many different letters an economy has, you also have an idea of how they can be recombined. They would be able to find new products or find new combinations, and that's why it predicts future economic growth.

In economics, you have this idea of convergence clubs. It's not that economies simply converge because they're poorer and therefore capital is going to generate larger returns and these economies are going to grow faster. That convergence is going to be conditional on meeting some conditions.

In this case, we can show that convergence clubs are associated with the same levels of complexity. When you look at India according to this measure, you say, “Well, India in terms of income is here, but in terms of complexity, it's at the same level as Turkey.” India should be able to converge to the level of income of Turkey, and it's within that club.

When you look at Liberia, Liberia is very poor, but in its club there are not a lot of rich countries. There are none, actually. Therefore, its complexity is in equilibrium with its income at that moment. When you look at a country like Qatar, you get the opposite. Its income is extremely high given its complexity.

If they were to run out of petroleum and gas, then their income would have to come down because they would be in equilibrium with countries that are more middle-income, maybe countries that have about 7K, 10K, or 12K of GDP per capita.

Can there ever be too much complexity? Again, to use the language-model analogy, we should also talk about the relationship between embeddings—natural-language embeddings that use this distributional hypothesis, “a word by the company it keeps.” I guess you could describe what you're creating as something similar to embeddings, based on different types of products being sold in the same location and economic space.

But the point I was making is that you could just argue that more complexity is good. In a way, what that's saying is that there are more paths. There are more paths in innovation space that are reachable through combinations of the letters that exist. But can too many paths be a bad thing?

No, yeah. I think one of the things is that having a wide space of options is one thing, and the ones that you are exploring are another thing, which is a strategy. I would say that a country like the UK, China, the United States, Singapore, or Switzerland has a wide set of options, and that doesn't mean that it's exploring all of them.

There might be a lot of bad options to explore. Let's say Breaking Bad is a business model: that's a bad option. You can use your skills as a chemist to go down that path, but the same skills that Walter White had as a chemist can also be used to find cures for diseases, which is another application.

The options are there, and the choices are the ones that can be good or bad. But not having chemists in your country, I think, really constrains your capacity.

Absolutely. We should also factor in knowledge decay. In a sense, the reason why this is such a good proxy for future potential is because any bad strategies would have already been deleted by knowledge decay, because we would have let things die out.

We're trimming and pruning as we go. If these convergent strategies still exist, they probably exist for a good reason, because they have utility.

No, yeah. I think knowledge also gets replaced by other representations; sometimes they encompass the previous one. There's an example that I use in the last chapter of the book to illustrate disruptive innovation. It's about the way in which we calculate pi: calculating pi using polygons or calculating pi using Newton's series formula.

The idea of calculating pi using polygons was something that people used for more than 1,000 years. The last person to use it extensively was, I think, a Dutch mathematician who dedicated his life to calculating pi to about 32 digits, and it took him 20 years to do so. Then Newton was able to do that in a matter of days using this formula.

You had disruptive innovation in an example in which the new curve didn't even start worse; it already started better and went to enormous heights. In that context, I think it's fine that we forget how to calculate pi using a polygon with a million sides inscribed into a circle, because it is useless.

It's contained within a better way of doing things. We have a superior knowledge representation that replaces that inferior one.

So, based on our conversation today, do you think large language models have knowledge?

I don't think of knowledge as an individual phenomenon. I think of knowledge as a collective phenomenon. A baby born on an empty island isn't going to grow to be smart, even though it has the capacity to develop language, because that's a genetic gift that we have; it's not something that is socially learned.

I do think that LLMs contribute to this ecosystem of knowledge. Some people would have said in the past, “I have all of the knowledge in the world in my library.”

I would say books do not have knowledge, but books contribute to that collective capacity to know, learn, communicate knowledge, and so forth. So if knowledge is a collective phenomenon, the question is not whether LLMs have knowledge or not, because none of us have knowledge; knowledge only exists in that larger context. The question is whether LLMs, within that larger ecosystem, are increasing our collective intelligence or our capacity for collective learning.

There are obviously 2 schools of thought. One is just what you've said: that it's a cultural technology, and the locus of cognition, thinking, understanding, and action comes from us. The other school of thought is that they are agentic, cognizant machines in their own right. I just don't think that's true.

For me, what matters is that they're useful. I'm more pragmatic when it comes to that. Are they useful? Yes. Are they perfect? No. Nothing is perfect. A lot of things are useful, and I think they fall into that category.

I learn a lot by talking with an LLM, for example. I moved to France 5 years ago. There are a lot of things I don't know about the local rules that I can more easily consult using an LLM than a Google search. I want to know about tax law in France, and there are a lot of details. I explain my situation, I get those answers, and then I have an accountant.

When I meet with my accountant to discuss that, I'm much better informed and I can ask better questions. So I think I'm becoming smarter through those conversations because I'm learning faster thanks to those interactions. Is it because the LLM has knowledge? Is it because I have knowledge? Or is it because we're wiser when we're together?

So then there's the question of how efficient your book is at knowledge diffusion. I absolutely loved reading it. I actually read it in 24 hours and gobbled it all up. But it's overwhelming. I want to read it again several times just to get all the value out of it.

But in the afterword, you said that, I think it was about 10 years ago, you wrote a book—I think it was called “Why Information Grows.” You've done your PhD, and you were using very abstract thought experiments. Now you've had time to think about it, and you've told it using stories. You found really interesting examples that exemplify all the different ideas, and it makes it so much easier to understand. Tell me about that journey.

There was an excellent point that was made by Annabel Huxley. She works at Penguin, and she was the one who basically paraded me around London for the last 3 days. She's actually—funny fact—part of the Huxley family, you know, Huxley like Darwin's bulldog. So I felt so honored to go around the London science scene with someone like her.

She said one of the things about your book which is interesting is that your book is about the fact that knowledge is extremely specific. The book starts with the story of Charlie, which we haven't told, but it's a story that shows that knowledge can get extremely nuanced. But then you go and communicate all of your ideas about the principles that govern the growth, diffusion, and value of knowledge using these very nuanced ideas.

But these nuanced ideas, in this context, are not just decoration, because they help exemplify that exact point: that knowledge is extremely nuanced. The story of Ibuka absorbing knowledge on how to produce magnetic tape and the story of Samuel Slater bringing cotton-spinning manufacturing to the United States are full of details, and in those details is where knowledge hides.

That's something that I thought was quite interesting because, honestly, I had not realized it until she mentioned it that way: that the book made that point about specificity, and that therefore the specificity of the story, like that detail, was not a simple literary resource but was a way to hammer home that point.

Beautiful. Professor Hidalgo, this has been absolutely amazing. Thank you so much for joining us today.

Thank you. It's been a delight.