Sangeet Paul Choudary, bestselling author of Platform Revolution and Reshuffle, and senior fellow at UC Berkeley joins the show for a second time to explain why AI’s greatest impact will not be automating today’s work, but rebuilding the structures beneath it.
Jim and Sangeet discuss AI-native companies, the future of knowledge work, American and Chinese AI models, creative “exhaust,” changing career paths, and how to compete when the rules of the game are constantly being rewritten.
We’ve shared some highlights below, together with links & a full transcript. If you like what you hear/read, please leave a comment or drop us a review on your provider of choice.
Links
Highlights
The Cutting-Room Floor Is the Real Asset
Sangeet: If value chains are moving from output to outcome, then the creative exhaust is going to be worth more than the finished work. Think about writing a book. For every word that makes it in, there are probably ten or twenty words in wasted drafts, and maybe five hundred more sitting in notes that are genuinely valuable but never see the light of day. They only exist to help you get to the output, so we treat all of it as waste. But that is exactly where the learning sits.
For the last sixteen years I’ve captured everything I read, every thought, every connection, in a tool called Workflowy. Every day I’d add to it. Every year I’d take a week off just to move ideas around and see how they connected. I could never really do anything with it, until now. I took the whole thing and turned it into a knowledge graph that links all those ideas together. Now it can tell me where the gaps in my own thinking are, the connection I’ve been circling for years but never actually made. What I used to think of as exhaust turns out to be the most interesting part.
Jim: The analogy I use is what was left on the cutting-room floor.
Sangeet: Correct.
Jim: And I’m a prodigious note-taker. For forty-five years it was all by hand, and now we’re digitizing it. Some of what I’m seeing, I would never have thought of in that particular combination, even though they’re my own notes. The AI can basically go in and say, “Dude, you’ve been saying the same thing for forty-five years, how come you never acted on it?” And out of that come whole new ideas, whole new books, whole new ways of learning.
Sangeet: Exactly.
Don't Think Electricity, Think Agriculture
Sangeet: I think the shift to AI is very foundational. It’s not something you can compare to the shift to electricity. I’d say it’s more like the shift to agriculture, or to wood, or to minerals.
Jim: That’s the argument I’ve been making. Don’t think electricity, think agriculture.
Sangeet: Exactly. Because the fundamental input into the value chain is changing. And when the input changes, an entirely new production system has to be built on top of it. But most of our industrial mindset assumes the input as given and treats the production system as the thing you act on. Now we have a new input, and we haven’t even begun to rethink what the production system should be. That’s really what China is working on, treating both energy and AI as new inputs into fundamentally new production systems.
The point isn’t whether China is right or the US is right. It’s that we’re still incredibly early. We’ve just discovered agriculture. We’ve worked out how to make the ground a bit more fertile. We haven’t yet figured out that one day we’re going to make clothes out of cotton. We’re that far from the whole value chain.
Transcript
Jim O’Shaughnessy
Well, hello everyone. It’s Jim O’Shaughnessy with yet another Infinite Loops. I’m delighted to welcome back Sangeet Choudary, the best-selling author of Platform Revolution and the one I really want to talk about, the newer Reshuffle: Who Wins When AI Restacks the Knowledge Economy. Welcome back.
Sangeet Paul Choudary
Thank you, Jim. So glad to be here.
Jim O’Shaughnessy
I have certain guests who, when I see them on my schedule, I get excited, and I so enjoy talking to you. But for people who might not be familiar with your work, let’s do a bit of a recap of your thesis so we can level-set for our listeners and watchers to understand where you think we’re going.
Sangeet Paul Choudary
Absolutely. The key thesis I talk about in Reshuffle is that when we think about the impact of AI on knowledge work, we often tend to think of AI in the same way we’ve thought of previous forms of mechanization, which have typically impacted relatively more codified, standardized workflow. AI is fundamentally different, because work that was previously considered tacit and bound inside human cognition can now be broken down and taken over by machines in a way that was previously not possible. So one of my key points is that the traditional way of looking at the impact of mechanization on traditional forms of work does not apply when we think about the impact of AI on tacit knowledge work.
What I mean by tacit work is fundamentally this. Take the example of solving any market analysis problem. There is no fully codified workflow. You make up your workflow as you get into the problem. You define it, and you learn from the information you’re gathering. It’s difficult to explicitly codify the workflow, the number of iterations, what good looks like. That’s why traditional forms of mechanization did not apply to tacit work, but today AI does. You can see that with the market analysis example: you can get a really good market analysis report using Claude, which is getting better every few months.
So my key point is that when we looked at previous forms of mechanization, we used to think about the impact primarily in terms of automation or augmentation, because you were taking a predefined way of doing work, looking at different steps in that predefined way, and inserting a machine at specific steps. It was either automating that step or task, or augmenting the human on that task. A lot of us are still taking that same model when we think about AI and applying it. That is what I call the task-centric, automation-based way of looking at AI.
The real story we’re missing in that process is that when previously known, modularizable, non-definable forms of work suddenly become redefined and remodularized in fundamentally new ways, the structure of work changes. Who performs which work changes. Take the example of consulting. The pressure on consulting firms is not simply from model providers. The pressure is also from the fact that model providers now change the division of work between what customers could have done in the past and can do now versus what consulting firms can provide them. When customers take more of the work and model providers absorb some of the other work, the consulting firm in between has to rethink not just what to do with the work that remains, but what new work to go toward.
So the real idea of Reshuffle is that what AI does is unbundle our traditional work structures, whether it’s a workflow or an organization. These are specific configurations, what I broadly call bundles of performing work. It unbundles all of that. It redistributes what machines can do and how humans can reorganize what they want to do based on improving machine capabilities. With that, it creates the possibility for fundamentally new structures to emerge: new workflows, new organizations, new industrial boundaries. So the real impact of AI will not be in speeding up today’s work. It will be in reconfiguring everything that supports work, which is our workflows, our jobs, our organizations, our value chains. All of those are poised for reconfiguration. That’s the real idea of Reshuffle. When AI restacks the entire knowledge economy, the game will change, and hence the winners and losers will change as well. That’s the whole idea of who wins when AI restacks the knowledge economy.
Jim O’Shaughnessy
Yeah, and we are very simpatico on that idea. I think many people are asking the wrong questions and framing it wrong. One of the reasons they’re framing it wrong is because they’re captured by their priors. They’re captured by the old workflows, the hierarchy, the meetings, the departments, the job descriptions. I started thinking about this a while back and called my little thing the great Reshuffle. What I saw coming, and it really is, is changing the routing. AI is really going to change the way things get routed to different nodes. And there are going to be problems. There are always problems. But it could create what I would call a new distributed value mesh network.
If an earnest CEO who really wanted to understand what was going on hired you as a consultant, what would you tell them about how he or she should work to design, maybe from first principles, around the new coordination that AI will allow?
Sangeet Paul Choudary
The place where I would start is how you think about strategy. Today, the typical way you think about strategy is in terms of answering two questions: where to play and how to win. What’s the playing field going to look like, and how do you think about the winning game? The first thing I would talk about is that the playing field within which your firm is playing is fundamentally changing. I’ll give a few different examples. If you’re in the knowledge services industry, the professional services industry, clearly the playing field was structured around the assumption that access to certain forms of knowledge work was expensive and proprietary, and hence you could hold it internally and license it or charge for access to it.
So when access to certain forms of knowledge work suddenly becomes very cheap, because AI does perform certain forms of knowledge work, within limitations of course, but it dramatically reduces the cost of accessing certain forms of knowledge work, how does that then change what is differentiated versus what is commoditizable in your industry logic? Or take it all the way away from professional services to a very physical industry like material sciences, plastics, and compounding. Traditionally, the cost of compounding was that if you had to identify or discover new compounds, you had to go into the lab, run a lot of tests, and come out with compounds. Today AI can generate new compounds from scratch. So the value then shifts to how you match the compound to the process, to the demand profile.
Because alongside that, the number of use cases at the front end is also dramatically expanding. As the supply of compounds increases and the demand for different use cases increases, how do you match the two things really well? That’s where the new constraint sits. So the first thing you would want to look at, if you’re a CEO, is that you have to follow the scarcity, because you can compete on the basis of what’s scarce, not what’s becoming abundant. What’s becoming abundant is going to get commoditized. Margins associated with that particular activity are going to get compressed. So you need to think about where your new scarcity lies, and how you can create defensible business models around owning that scarcity. That’s on the playing field side.
At the same time, you have to think about AI strategically in terms of your organization. Today a lot of CEOs are just focused on: we need to get AI adopted, we need to roll this out, we need to get teams on board. All of that is a good starting point, but it is only that. What we should really be thinking about is that as AI gets adopted in our organizations, there are three things changing simultaneously. One, the capabilities themselves are changing, because Claude, OpenAI, all of these are improving on a daily basis. With every new release, what the machine can do and what humans can do changes. So the capabilities from a pure model perspective are changing.
Which of those capabilities actually get adopted in value-creating workflows in the organization is constantly changing as people learn how to adopt the technology. And as people adopt the technology, how they change their own work is constantly changing. So if you’re a CEO, the biggest thing you need to think about today is how you manage this constantly evolving capability set, because you’ve never had such a constantly evolving capability set before. You had a top-down capability set where you said, we’re going to hire these people, map them into these skills, and keep training them on this basis, so the skills you have access to constantly change but you managed how those capabilities evolved. As you deploy AI in the organization, it’s going to evolve on its own.
And not just the machine capabilities, but how they’re adopted, and more importantly, how your people are changing their work around it and how they are discovering new capabilities or moving to new things. That’s going to change what you can do with this evolving capability set. So there are two things here. One, the nature of differentiation and competitive advantage in the playing field is changing. Two, the capability set with which you can innovate from inside is changing. It’s really at the intersection of these two things, given that they are changing for every company across industries. So your competitors are not going to look the way they used to look in the past.
Because once certain forms of knowledge work that were protected by your industry boundary become available outside that boundary, new competitors can come into your industry where they could not have in the past. I’ll give a very simple example. Think about what’s happening to customer support. Traditionally, customer support was something you performed in a contact center. It was a very closely bound industry boundary in terms of who runs a contact center and what kind of tools are provided to them. Today customer support can be performed by AI agents alongside human agents. So you have companies like Microsoft and Salesforce, completely outside those industry boundaries, now entering the support industry, because they can access the capability of providing support without having to invest in training humans to build that capability from scratch.
So when you think about these aspects, the fact that previously scarce capabilities are becoming commoditized, that ways of differentiating are changing, that who can get access to and enter your playing field is fundamentally changing, and that your capability set internally is changing, that creates a very volatile and dynamic competitive playing field, both in terms of what you can do and where you can play. I think understanding and resolving that is the key challenge for CEOs today.
Jim O’Shaughnessy
Another thing I think we face: the authors of the book The WEIRDest People in the World, I don’t know if you’re familiar with it, make the argument that cultural lag is a real thing, and that it takes humanity a while to adapt to a new way of doing things. I think that’s true. But I also think it provides an asymmetric arbitrage opportunity. The asymmetric arbitrage opportunity is: are you going to be willing to move rapidly to a truly AI-first corporate or institutional structure in everything? If you are, it’s my thesis that you’re going to be able to eat the lunch of the players using the old playbooks. Is that one of the reasons? This is why you see the current top of the hierarchy, let’s look at movies, books, a variety of things, they’re not happy at all.
And yet, much like the movie industry, they tried to suppress VHS for years and years. They spent millions of dollars, they lobbied against it, and of course they lost. And then the wonderful irony is that VHS sales and other such sales ended up being a much larger part of their revenue, the very thing they fought all that time. Do you see something like that happening with traditional dominant companies today versus the upstarts?
Sangeet Paul Choudary
Yeah, I think so. It’s that particular element of a company native to a new technology versus a company migrating to a new technology, and the arbitrage opportunity for the company that’s native to it is not entirely new. What’s new is the rate of change, the speed at which this is happening, which confers a compounding advantage to the native player versus the migrant player. If I were to qualify this, take a different technological shift. Take the shift from desktop to cloud, which happened 15 years back. You had a company like Adobe, which moved to the cloud very successfully. There are Harvard Business School case studies about how well Adobe managed that shift, how well its CEO moved, how the CFO agreed to move from licensing to subscriptions. And yet it struggles today.
It has struggled ever since to compete with the native player, which was Figma. The reason it struggled to compete with Figma is that Figma was architecturally organized around the properties of the cloud. Adobe was using the cloud only as a channel. Figma was using it to redesign and reimagine the product architecture, the offering architecture, and the workflow supporting it around the capabilities of the cloud. To give a simple example, Adobe was built on a file-centric architecture. You had to create the file, and in the past you would send it over email. Now you could be on the cloud and have others coming in and commenting on the file, but it was still centered on the file-centric architecture, and the file was tied to the design division.
So the file-centric architecture preserved a certain logic. Figma, because it’s cloud-hosted, is created on an element-centric architecture, which means every design element the organization uses is managed in a central library. It can be reused across design files. So design files are very often reconstructed using new elements and common elements across the organization. There are two important unlocks because of this, and this shows the difference between the native player and the migrant. The first is that when you manage design elements in an organization-wide library, you’ve moved the value from design only to design and governance. Governance no longer sits inside the minds of designers, which is what was happening within Adobe. Now it’s managed organizationally. So the budgets you can access are fundamentally different.
And when you can represent design in terms of a cloud-hosted graph and not a design file, that allows multiple different players, marketing, sales, product, all of them, to work alongside the designer on the same file without forcing the designer to keep ensuring it is well validated and verified, because all that verification is happening centrally. My point is that Figma realized design was not an isolated activity. It was an activity that fed many workflows across the organization, and the cloud natively allowed that workflow spanning to happen across the organization. No matter what Adobe does, it cannot move in that direction. Ironically, Figma is now facing the same issue, because Figma is sort of an AI migrant.
It’s got this nice file structure, element-based, but then Anthropic, for instance, can infer connections between elements, can identify fundamentally new ways in which design files are being manipulated. It can do a lot more with the entire design system than a migrant like Figma can. So we see this cycle repeatedly: when a new technology comes in, if you have the opportunity to rethink the logic of the product, the workflow, your customer’s workflow, your organization, everything around this new technology, the starting point is whether you can really define what the new unit of value is. If you look at Figma versus Adobe, for Adobe the unit of value was the file; for Figma it was the element. So you could do everything at the level of the element.
Even if you opened up APIs for the ecosystem, your partners could be the same, but with Adobe they could only work on the file. With Figma, they could work at the level of the element and integrate at that level. So it fundamentally changes how companies organize, how they collaborate, how they compete. That is why there’s a huge arbitrage opportunity for players built ground up with the new technology. There is also a significant opportunity for players who are migrating, as long as they have several structural constraints that do not simply get wiped out. As an example, in professional services, you could be a copywriter at one end, or an audit and tax consultant at the other end.
As a copywriter, you don’t necessarily own structural constraints, whereas as an audit or tax consultant, you have the right of final sign-off, which is regulatory protected. So there’s a structural constraint that, even if the work gets taken over or commoditized by AI, you are able to hold on to the value at your end. It’s important to know under what conditions incumbents can retain value because of structural constraints, and under what conditions the AI-native player can fundamentally impose a game that the incumbent simply cannot play. All of those factors become very important when we think about who has that arbitrage opportunity.
Jim O’Shaughnessy
And one of the things, this is my view, I’m not attributing it to you, but I think one of the reasons we’re seeing all of what I view as desperate attempts for regulatory capture on the big frontier model designers and providers is because they understand that if they can’t get regulatory capture, if they can’t make the legal aspect of who gets to do what, then the playing field is truly a different playing field, because you’re not going to be able to be saved through the traditional things like copyright, all of the various legal procedures and protocols that we’ve built up over long periods of time. They’re all under attack, right?
Sangeet Paul Choudary
Yeah. I think there are two different issues, and let me know if I’m approaching this the way you intended it. When we think about copyright and other issues around value captured by AI models, there’s this element of ring-fencing what was previously a public good and extracting it the way you would a private good. AI enables you to create a transformation value chain where the end product is completely untraceable back to the input. Copyright law was built in an era that did not assume that kind of transformation, and essentially because of that it’s difficult to protect what’s happening here. This is not my idea. This comes from a well-known book called Power and Progress.
This has played out over the centuries, where every new technological shift, or many new technological shifts, has created opportunities to extract value by ring-fencing a public good and making it private, essentially because the regulation stayed stuck to the previous model whereas the technology imposed a fundamentally new model. In one way, that’s what we’re seeing here. The other thing that’s interesting in terms of how Anthropic and others, the model providers in general, are calling for regulation is because they also, I believe, realize that the American models and the Chinese models are playing a fundamentally different game. In the American model game, the model is sort of the end. So that is where value is supposed to be captured in the long run.
In the Chinese model, I believe the model is the complement that has to be commoditized in order to capture value at the complementary layers, especially when we think of physical AI and new production systems. The value there to be captured is in the energy system, in the production system on top of the model, which is the physical production system, robotics, and so on, and the overall coordination system in terms of how you build fundamentally new value chains, fusing the knowledge value chain and the physical value chain. So the fact that it’s a fundamentally different race and a fundamentally different game that the Chinese model providers are playing is also driving a lot of this concern around regulation.
Jim O’Shaughnessy
And which model, the Chinese version or the American version, do you think wins?
Sangeet Paul Choudary
I think they’re fundamentally different games. If we believe that eventually the knowledge value chain and the physical execution value chain are not two distinct value chains, the way we’ve been talking about digital and physical for the initial part of the digital era. When the digital era started, we always talked about the digital and physical value chains separately, and eventually we realized they are not really separate. You essentially create value by bringing the two things together. I think the same thing applies when it comes to AI. There’s the knowledge work value chain, and then there’s the physical work value chain, which converge.
I believe the way the model providers in the US in general are looking at it is purely in terms of what AI does to the knowledge work value chain, which is a fair way to look at it, because traditionally most value in today’s industries has moved to knowledge work. If you look at any industry, if you look at mining, or the example of plastics I took, the physical work gets fully commoditized and the value moves to the knowledge work. But if you look at the Chinese model in general, the two things are not fundamentally distinct. I take the example of Shein in the book, which clarifies this really well. If you think of Shein, the fast fashion player, the reason it’s interesting is that it creates a learning architecture that brings the physical and the knowledge value chains together.
It constantly senses new ideas and new trends coming up on social media. Then it converts that into a design using a network of designers, converts it into a small batch of production, tests it in the market, and with all of that it’s constantly training which batches work and hence how its physical production system and its design system should work together. So this idea of creating a learning architecture that spans both the knowledge work and physical execution is, I believe, where things are headed. But because in the West we’ve largely assumed that the physical side of things is commoditized, we tend to see the impact of AI only within the knowledge work value chain.
Jim O’Shaughnessy
Yeah. And I should note that we also have a publishing company, so we’re very concerned about copyright and are looking for new ways that the technology itself can unlock to preserve the author’s right to their work, to preserve the filmmaker’s right to their films, and so on. I think that is going to be a really important topic, because if you can’t cut the creators out, you have to come up with a protocol that works with the new technology. When everything can be digitized and the end product looks nothing like the inputs that created it, you’re going to get a lot of very unhappy people. So I think this is not just a business question, this is a societal question.
Sangeet Paul Choudary
Right, right.
Jim O’Shaughnessy
Because another question I would have for you: I sort of think that distributed value networks, mesh networks, where there is no central authority. So, for example, China had the Gutenberg technology way before good old Gutenberg came up with it.
Sangeet Paul Choudary
Right.
Jim O’Shaughnessy
But it had a very centralized command-and-control government that suppressed it, because the emperor wasn’t interested in people printing things that might be critical of the emperor. Whereas Gutenberg worked because Germany was not unified. It was a collection of city-states competing with each other. So China had the tech first, but Gutenberg was operating under a distributed network protocol, the protocol won. And I think another one of your points is that the coordination layer is going to be vitally important to determining who’s the winner and who’s the loser. Could you explain that a little bit, because you’re much better at it than me.
Sangeet Paul Choudary
No, for sure. This is one of the things I felt people were missing about what AI can do. The way I typically think of it is that when we think about the impact of AI on the underlying economics of what’s improving, we tend to think about two effects. Predictive AI is collapsing the cost of prediction. Generative AI is collapsing the cost of generation. What happens when the cost of something collapses is that it gets done more, in many more places, and hence it unlocks second- and third-order effects, because those previous constraints associated with those activities go away. Those things get done much more. The third thing these two framings miss is that the cost of translation, or coordination, is also going down with AI.
What I mean by cost of translation is that when a foundation model is trained on the world’s knowledge, it can translate between different domains in a way previous technologies could not. Previously, if you had to translate between different domains, you had to either rely on bespoke integrations, or rely on some form of power structure that was willing to invest in doing that translation so it could extract rents from it. It could be a broker or an intermediary, or it could just be what we call platforms, where they specify the interface, manage some form of the mapping, and then once they get sufficient traction, they force the ecosystem to follow the standard. So the translation happens through that mechanism.
Essentially, what all of that leads to is that there were two primary ways to coordinate multiple actors. You either had to get the actors to agree and reach consensus, and the example I take is the shipping container. Once trains, trucks, and ships agreed on the same format of the container, logistics got unlocked globally. So it’s coordination with consensus, or it’s coordination through enforcement and power, which is what, say, a Facebook or an Amazon or any of these large platforms do, initially heavily funded by venture capital, which allows them to gain scale, and then they enforce the coordination. My point with AI is that it allows coordination without consensus, because it collapses this cost of translation.
The reason that’s important is that today’s power structures are structured around proprietary formats, proprietary ways of doing things, proprietary workflows, proprietary interfaces. Essentially those power structures create walled gardens. If you take any industry, whoever owns the choke point, in some industries that’s the R&D choke point, in others the distribution choke point, whoever owns the choke point enforces the standard, the schema, and the way of working on everybody else. What AI can do is take outputs at every step of a value chain, learn from them, and translate them between different walled gardens in the same industrial system. The example I give is construction. In the construction industry, you have players like Autodesk and Bentley with different design formats.
That’s on the architecture and design side. Then there’s the downstream construction and engineering side. All of these tools and technologies don’t talk to each other, and the players providing these tools create their own islands, or walled gardens. What we’re seeing now with new AI-native construction players coming in is that none of them are trying to do design faster. Nobody’s attacking Autodesk straight up. They’re trying to dismantle the Autodesk logic by creating these translation layers, where they tell the user: no matter where your different teams are working, in which formats, we’ll help you see a unified view of the project. Through that, we’ll create a coordination-without-consensus layer.
Once enough demand moves in that direction and starts using this AI layer to make decisions, they would then want to flip it and enforce the interfaces on the various format providers. That’s the idea of coordination without consensus, where you can dismantle today’s power structures and create an alternate way to span the different formats. I think the new power structures it yields could be a combination of protocols, because you need completely open protocols to get different formats to work together, but then a different form of concentration at different points. It could be just one node that you capture, it could be a combination of different layers, but it would be different from today’s power structure. That’s where, going back to your point about protocols, that’s a fundamental Reshuffle pattern we’ll see as well: enabling coordination without consensus, breaking down today’s power structure, and creating alternate structures going forward.
Jim O’Shaughnessy
Yeah, and that’s where I got really intrigued in that part of your book, because I agree wholeheartedly. But if you study history, it’s really nothing new. By that I mean the distributed value architectures and protocols. The example I always use is the Rothschilds. The Rothschilds were five nodes without a central authority. The father sent each of his five sons to the capitals of Europe of the 19th century, the financial capitals. They had an amazing sub-layer of courier network. They had rented boats, carrier pigeons, ciphers that only they knew. But it was not routed to a central authority to make decisions. It was routed to the brother who could take the greatest advantage of it. And it’s an apocryphal story.
I think historians go back and forth on this, but because of this topology that the Rothschilds had, that was their key advantage over others. They had a better information network than government information networks at the time. Of course, the canonical story is Nathan Rothschild, through this courier network, knowing that the Brits won the battle against Napoleon. Then he took advantage of human psychology, went into the stock exchange, and sold British gilts, or bonds. Everyone thought Rothschild knows things ahead of other people, so everyone just rushed to sell. Of course, he had his agent buying from the plunging prices. And they literally controlled finance in 19th-century Europe until the telegraph destroyed their advantage. I see AI in a very similar role here.
I think it’s hard to articulate just how fundamentally transformational a technology this is, because you’re going to be touching a lot of third rails. That’s going to be messy, and it’s going to be a very difficult fight. What about the idea of measurement capture? You don’t have to control the entire network if you can control what gets counted. Could AI itself, as the new router of all this information, could we see a system where that just gets copied and copied and everyone’s using the same routing system? Or am I off base with that?
Sangeet Paul Choudary
No, that’s an interesting thought. There are a few complementary thoughts I have. One is that we’re clearly moving in a direction where firms are not going to compete on the basis of outputs. At the risk of using the cliché, outputs to outcomes, everybody’s talking about it, but what does that mean? I think that has something to do with what we’re talking about in terms of measurement capture as well. We’ve optimized firms around creating the best output, the most output, and scaling that output. We don’t necessarily know what it means to optimize against outcomes. I believe optimizing against outcomes is about owning a proprietary learning architecture, which helps you own the outcome better than every other firm and hence optimize your entire internal production architecture to improve that outcome better than anybody else.
What I mean by owning that learning architecture is this. When we think about learning, after 10 or 15 years of digital technologies, we still think about learning as what happens on the customer interface, personalization, that the system knows more about you, it’s learning about you. We’ve not necessarily seen learning to the same extent across the entire production architecture. Shein is one of those examples where, because it’s a proprietary architecture, you are fine-tuning every element of it. What should be designed, which kinds of design elements work well with which kinds of production factories, which production factories work well with which markets.
You’re fine-tuning every coupling in the value chain in terms of what works well with what. In new value chains, if you look at something like electric vehicles, you do have that learning architecture, because you’re simultaneously fine-tuning what the battery can do, what your charger can do, what your charging network should look like, and hence what the customer experience looks like as a combination of these three things. So there’s a coupling between battery capabilities, charger capabilities, charging network spread, and so on. Over the last 50 years, we’ve constantly moved away from inefficient vertical integration to modularization, first with outsourcing, then with cloud capabilities. Whichever way you think about it, it’s constant modularization, decoupling, making things more flexible and plug-and-play.
But when you move from outputs to outcomes, you need to own the entire production system in order to guarantee the best outcome. So we’ll see a new form of reintegration happening, structured around defining what should be counted as the right measurement at every stage in the value chain, toward ensuring you are able to guarantee the best outcomes to the end customer. That’s the way I would think about measurement capture: the ability to win the outcome game is determined by your ability to own the learning architecture, which is determined by your ability to dictate what should be measured at every stage, because that’s what guarantees the end outcome in the best possible way. I’ll throw a related idea against it as well.
There’s a recent HBR article I wrote that came out last week, where I essentially make the point that we’ve traditionally seen work and learning and skilling and all of these things as decoupled, in the sense that you go and perform your work, then you go somewhere else to learn, and then you come back and perform your work. With AI, what’s happening is that for the first time we’re working with a technology that’s constantly working alongside us. Hence it’s also measuring how we are working, and transforming how it behaves so that it can give us highly personalized feedback based on how we’re working. So we have created this.
We are moving in this direction where learning, measurement, all of that is going to happen in the flow of work, and hence which kind of measurements matter and what learning path they take you individually down is also going to be determined by AI. So I don’t know if that directly addresses it.
Jim O’Shaughnessy
Yeah. Actually, I’m looking at a question, because I believe pretty deeply in this distributed value chain, where, historically, if you look at all wealth creation, not extraction, but creation, it happens under these circumstances. The Greek city-states, all the way up. For politics and philosophy and all of that. It’s not just business, it’s societal, it’s political, it’s everything, the Gutenberg example. In your answer, I asked you this question that I’ve written down. I’m going to read it, because you answered it very much the way I’m looking at the world. I say: if AI is the new coordination layer, and the distributed architecture is the historical engine of value creation, is the highest value produced by controlling the layer, or by designing a layer that can’t be controlled?
And what you’ve just said is you’ve got to create a proprietary layer of knowledge, reinforcement, learning, all of that. So it sounds like we’re both directionally thinking the same way, right?
Sangeet Paul Choudary
Yeah, and I would agree. When I use the term learning architecture, there are two things I try to reinforce, because people have preconceived notions of what learning means based on how we’ve used the term in the last 15 years when it comes to business models. The first is that people just assume learning happens in the market, not across the production system. I think that’s going to be truly unique and new with AI. That is why, again, I believe the Chinese model of betting on commoditizing the model and capturing value from the production system, the energy system, and the existing coordination system on top of the physical knowledge production system is where value is going to be.
So that learning architecture piece is something I believe is not fully understood to the same extent in the Western paradigm of looking at technology. The second piece is essentially the other reason I use the term architecture. When we use the term layer, and I’m just trying to refine what we’re going back and forth on, the term layer is borrowed from the past several years of modularization we’ve gone down, where we’ve assumed certain layers are going to be commoditized and certain layers are going to be valuable. I believe that’s going to continue happening. That’s just how value migration works.
But in the past, it was possible to be just one layer and not do anything else, or at least be visible as one layer, even if you had adjacent control positions and adjacent points. The classic example being that Amazon has Prime as the visible layer, but it has that price-matching algorithm that runs at the back end which helps it retain the Prime position. But if I take the Tesla and electric vehicle example again, it’s not just what we call a layer today, but the couplings between layers and a proprietary new learning architecture that helps you understand how those couplings behave. As the technologies at each of those layers are improving rapidly, owning those couplings is where advantage accrues.
So those are the two reasons I stress the idea of the learning architecture in a way people have not used it with digital technologies, at least.
Jim O’Shaughnessy
Yeah, and again we are aligned, because we think about it at the architectural level as well. We started at O’Shaughnessy Ventures to be AI-first in every aspect of everything it does. And if you’re 66 years old, that requires a lot of open-mindedness about the way things are going to get done. Luckily, I’m very open-minded about these things. I definitely think you are absolutely right that people are not thinking big enough in many circumstances, because I think you’re spot on about thinking at the layer level. That’s a whole different conversation from the architecture level, and the ability for learning to happen in the workflow, to be reinforced, and all of that changes the way everything is done. My concern is, again, back to cultural lag.
We have been socialized to believe that I’m going to get this job, this is what I’m going to do, and this is how I’m going to be judged, by my outputs, and the whole deal. That is not what the future institutions look like to me at all. And I don’t think there’s necessarily a sense of urgency, like those of us who were early to machine learning and seeing what it could unlock. I hear a lot of people saying, oh, we’ve got to get going, we’re late. I think in the grand scheme of things, we’re still pretty early. Am I off base? Do you think I’m right there?
Sangeet Paul Choudary
I think we’re at that phase where we’ve seen the new capabilities that are emerging, we’ve seen how that could lead to new firms and new architectures. We have not yet seen all possible architectures in all possible contexts through which the capabilities can be used to create advantage. What I mean is that if you think of any of the previous shifts, and I think the shift to AI is very foundational. It’s not simply something that can be related to a shift to electricity. I would say it’s more foundational. It’s probably similar to a shift to agriculture, or wood, or minerals.
Jim O’Shaughnessy
Because that’s the argument I’ve been making. I’m saying, don’t think electricity, think agriculture.
Sangeet Paul Choudary
Exactly, right. Because the fundamental input into the value chain is changing. The previous input used to be very different, and an entirely new value chain has to come on top of it. So both agriculture and minerals are two things I would look at. Most of our industrial production mindset assumes the input as given and assumes the production system as the object of action. But now that we have a new input, we have to rethink what the new production system should be. And I don’t think we are thinking about that. This brings me again back to the China point, where they are working on both energy as a new input and AI as a new input into fundamentally new production systems.
I think it’s less a point of whether China is right or whether the US way of looking at AI is right. It’s more about the fact that we’re still very early. We’ve sort of just discovered agriculture. We’ve figured out how to make the ground more fertile, more arable, and so on. We haven’t yet figured out that we’re going to make clothes with cotton at some point. We’re too far from that whole value chain at this point.
Jim O’Shaughnessy
Yes, and again, I agree completely. The thing I’m struggling with as well, though, is this notion of, if you have this distributed, non-centralized mesh network protocol, a viable mesh should be forkable, it should be auditable, it should be hard to fake and hard to capture. So what happens if AI becomes the ultimate mesh capturer? By that I mean, we have the centralized model that we’re using in America, we have the commodity model that the Chinese are pursuing. It’ll probably be a combination of both. But what happens, if the AI is getting smarter and smarter, are they also compressing toward a middle? That’s a criticism I often hear about AI, that it doesn’t, unless you specifically say, hey, I want to look way out in the tail. What it does is compress all of the knowledge into a standardized common ground.
Sangeet Paul Choudary
Right.
Jim O’Shaughnessy
I’m being inarticulate about this, but the idea is: is it the ultimate enabler of these hard-to-fork, auditable, hard-to-fake, and hard-to-capture mesh networks, which I kind of think it is? I’ve been thinking a lot, and this is what I was looking forward to in getting your view, because I like to think I might be wrong. What happens if I’m wrong?
Sangeet Paul Choudary
So I’m going to again offer a complementary perspective here. I think what I’m saying and what you’re saying, if you bring them together, we could have an interesting theory. Let’s take the creative value chain. It assumes there’s a certain creative process that’s valuable, and the object of action in that creative process is the output that has to be created, whether it’s a book or a song or whatever that is. In the process of working on that object of action, there’s a significant level of creative waste it is willing to deal with, which is not going anywhere. It just goes out as creative waste. It does not have any place in that value chain. It’s all the unwritten drafts, all the messy notes, and so on.
I believe that if value chains are moving from output to outcome, the creative exhaust is going to be more valuable than the object of action, because the creative exhaust is where the learning architecture sits. Hence I believe everybody who has built a life thinking about perfecting the output should now go back to thinking about what to do with the creative exhaust. I wonder whether the protocol plays out most interestingly at the level of the creative exhaust, where my creative exhaust, what I waste, can be refurbished and reused by somebody else it perfectly applies to, who is at a certain point in their own creative exploration. I’ll try to give a couple of examples to illustrate that, and also share how I’m thinking about my own work with that context.
But the other thing, so there are two points I’m trying to make. One is that if we believe the input is fundamentally changing, which it is, which is what is scaring the creative industries, if we keep focusing on the creative object as the output, as the focus of production and the focus of improvement, if we stay focused on the output-centric model of creative work, I have a feeling that’s a loss. We are moving in the direction of: where does this output move into, what problems does it solve, how can I think about that outcome and create an architecture, a learning architecture, to solve that particular thing?
It’s easier to think about that when we think about electric vehicles, because we’re saying, okay, whether it’s battery technology or charging network, all of these are working toward one outcome, driving range. So you have to keep improving them. You don’t improve only the vehicle. You improve every component that was previously assumed as given, in order to get to the outcome, and the learning architecture is where the advantage is. It’s difficult to apply that directly to creative work today, because we don’t necessarily think of it in the same way. But I’ll try to share what I’ve been thinking about in terms of my work, in case it helps us, first, bring it closer home to something I’m applying this to, and second, help us see if there’s a way to bring it back to creative work.
When I am writing books, for every word I write in a book, there’s probably 10 to 20 words I’ve written in wasted drafts, but there’s probably 200 words, or more, 500 words, that I’ve captured in notes that are very valuable but are never seeing the light of day, because they’re only helping me improve, only helping me get to the output. Over the past 15 years, 16 years now, I’ve captured everything I’ve read, every thought I’ve had, every idea I’ve connected. I’ve captured them in a tool called Workflowy, which is sort of like something before Notion and Roam came in.
It was a way to organize your ideas and do it in a slightly hierarchical but flexible way. I curated that over the years. Every day I would sit on it, put in what I was reading, put in what I was thinking, connecting those ideas. Every so often, when I was in a curatorial mood, I would go into it and move ideas around based on how I thought things connected. Every year, at the end of the year, I would take a week off to just play with the Workflowy for the year and see how it connected with previous years and so on. So that was just me geeking out on something I enjoy doing anyway.
So far, I’ve not been able to do anything with it, but now I’ve been thinking, why should I be looking only at the book? I should be looking at all these ideas that have been there at the back end. So what I’ve done is taken all of that and created a knowledge graph that links all of those ideas and concepts together. Now that the underlying knowledge graph has been created, it’s really powerful, because it can tell me which conceptual gaps the ideas have, because I just never read that. So it’s able to identify what I should be looking at next. It’s able to identify what’s unique about that contribution, and it’s able to show it to me.
For my curation on whether a certain thing should be added to the knowledge graph or not, I can take the same knowledge graph and run an entire case study of something happening in a certain industry through it. It calibrates the knowledge graph to say these concepts are playing out in this format, in this timeframe, in this industry. I’m having a lot of fun with it, because I’m constantly learning where the gaps in my learning are. It’s also opening a whole range of ways in which what I had previously seen as exhaust and waste, and had never seen the light of day, all of that is being converted into a knowledge graph that can be converted into something usable and productizable and interesting for different stakeholders downstream.
So that’s the way I’ve been thinking about it. If you think of the creative process, or the production process in any knowledge work for that matter, the value of that, corresponding to the learning architecture, lies in these learning iterations that lie scattered all over the place, not in the finalized output. I’ve always felt that the finalized output was more of slicing and dicing and providing a certain narrative. But the narrative is only one way of bringing together all those ideas. So if I take back my unbundling-rebundling thesis, the narrative of the book is a bundle. If you unbundle it, there are many concepts, and they can be rebundled in many ways.
But because the book has a certain constraint that forces that narrative, whether it’s attention span or storytelling, that constraint prevented many other components from entering that bundle. So how do you then reorganize all of those components? That’s where a knowledge graph, which can then be transformed into many different products, becomes even more interesting. So that’s been my way of thinking about it. This is something I always think about in terms of where my own work is going. But in general, I think some version of this applies to the knowledge economy and knowledge work in general.
Jim O’Shaughnessy
Again, listeners and viewers are going to get bored, because we’re just going to agree on everything. I think you are absolutely right. The repurposing of, the analogy I use is, what was left on the cutting room floor.
Sangeet Paul Choudary
Correct.
Jim O’Shaughnessy
And I am a prodigious note-taker. For many years over the last 45, it’s been by hand. But we’re digitizing all of this, and some of the things I’m seeing are, I would have never, even though it’s in my own notes, I would have never thought about it in this particular new combination. And it is, for me at least, incredibly exciting and indeed incredibly generative. Because the AI’s ability to go in and say, dude, you’ve been saying the same thing for the last 45 years, how come you never acted on that? That jives with your ability. You can recombine, and then whole new product lines, whole new ideas, whole new books, whole new ways of learning come out of that.
Sangeet Paul Choudary
Exactly.
Jim O’Shaughnessy
But let’s, okay, so we’re both very lucky in that we have the ability to have access. We have our own hardware installation, our own multimodal AI. At OSV, you obviously have access to all of the frontier models. You probably have your own bespoke models as well. What advice can we give to the listener out there who doesn’t have the total access, but some access? How can they start to change their own path? Especially, I’m thinking of younger people. I think the days of “I’m going to go get my degree in X, Y, or Z” are painfully antiquated. What would your advice be? I’ll hire you for a second here as a consultant to our younger listeners. Where would you point them to be able to maximize their ability to thrive in this new economy?
Sangeet Paul Choudary
Okay, so I think the larger question here has different answers when you think about a younger audience versus people like me who are in their 40s, or even people who are in their 30s, who have had a certain run which gives them, not so much a buffer to experiment with, but more importantly, it has helped them cultivate judgment. It has helped them understand what good looks like. And it has helped them see what output to outcome looks like, because even though you don’t measure the outcome, you sort of know the outcome you’re looking for and you tacitly understand it, so you fine-tune the output to get there. I want to come to the younger part, but to the...
Jim O’Shaughnessy
Actually, you know what, you’re going to the right place. Start with the people like yourself who are in their 40s, or even 50s. They have this long history that they’ve habituated in themselves.
Sangeet Paul Choudary
Right.
Jim O’Shaughnessy
And they’re going to need to be able to change that. So let’s address those folks first, but then get to the younger people who might not have to avoid all the bad habits that we’ve developed.
Sangeet Paul Choudary
Absolutely. There are two very different scenarios between the two. There are people who have been through it. I think the biggest thing to unlearn is that they were told a story, then they saw that story play out, and they now believe in that story. That’s what the youngsters don’t have to worry about to the same extent. They had to worry about a lot of things, but not this. The youngsters already don’t believe in the story they have been told, because they’re seeing it’s not working. But those who have been through some sort of career for 10, 15, 20, 25 years, they were told the story, they saw the story play out, they thought they’d figured out the rules of the game, and now they’re seeing the game change.
So the first thing is that everybody wants to believe the game is changing for others, but not for themselves. That there’s something special about what they’re doing, and not just about what they’re doing, but about how they mastered the game, and that it cannot be taken away from them. I think that is one of the first challenges that anybody with experience has. I’ve gone through that cycle as well, thinking I’ve mastered a certain game, and now, even if AI can do certain things, I should just keep getting better at my game.
It takes a lot of unlearning and questioning, some level of shock to your system, and a lot of reflection for you to come to a point where you realize that, well, it’s all good if the game is not changing and if I’m special, but shouldn’t I be preparing for the reality where the game is changing and I’m not special? And if that doesn’t come through, that’s fine, it was just insurance money. But I should be preparing for that so I’m not caught unawares. So that flip does not necessarily happen for everyone. It happens in different ways for different people.
Some people might realize they’ve sort of come to the end of the road of the previous game, and then realize it’s just not worth moving toward a new game. They’re at a place where they want to stop playing games. That’s possible. Another group might be: we’ve come to the end of the road, but you know what, I can see four more years of playing this game, and I’m going to double down and make as much as I can playing this game, and I’m going to stop after that. So those are very viable scenarios, and those are the classic board-level gotchas.
Every time I speak to a board, the underlying question in the room is: I’m 65 or 75 or whatever, and I’m going to be around in this board seat for three more years and in any board for five more years. Should I really care about AI, if you don’t know when your specific prediction is going to play out? Yes, it’s going to play out, but is it two years or 10 years? So should I then care about it? All those are very viable responses as well.
But if you’re somebody who has taken pride in their craft, if you’re somebody who’s always tried to improve at their game because of your internal ambition and to see the results of that ambition, and you’re not just trying to stay in the game just to stay in the game, I think you should take a hard look at what the new game that’s going to be played is. You should just assume the game is going to change. If you really care about your work, if you really care about getting better and doing better, you should fundamentally assume that everything you know is going to be taken away, and you have to rethink what that is going to look like.
Maybe it’s going to happen in two years, maybe in five years, but instead of betting on the timeline, I would just bet on the outcome, and figure out how to make the most of the lag period while betting on the outcome. So that’s how I would think about people in their 40s, 30s, whatever. Unlearn what you felt you knew. Get over the fact that you felt you had figured out the game and that your game will not change, because it’s actually going to change. Get over that and start afresh.
For the youngsters, I think it’s a lot more challenging. And here I’ll say it’s not just those who are young, but anybody who’s been structurally locked out of an upwardly mobile knowledge economy career and who’s been constantly trying to work their way back into it, whether they got kicked out of a job and entered the services industry, or became an Uber driver or whatever, and they now see that path being locked out even more.
So in general, people who are getting into what would traditionally have been called an upward graph of a knowledge work career, and especially those who are young, won’t necessarily have the on-ramp that allows them to build judgment, because the friction of working with an untrained human in certain cases will be too high compared to the lack of friction of working with a sufficiently trained, highly trained AI. Because of that, firms are going to make choices where they simply don’t go down a path where they need it. So a lot of those on-ramps that helped you fine-tune your judgment, that helped you learn from people who had judgment, all of that goes away. So it’s very easy to say AI will do the execution, humans will retain judgment.
But while there are many issues with that, the bigger issue is who will have that judgment. There’s going to be a significant portion of humanity who might not even be allowed the opportunity to develop that judgment and hence have a unique edge. I don’t have very good advice to give to people who are young. I think the younger you are, the better off you are in a way, because you’re learning that this is happening. I think the transition generation, which went into college in 2020, was stuck at home instead of being at university because of COVID, and then came out of COVID to see ChatGPT launch.
I think that transition generation has it most difficult in a way, because they’ve been told a certain story, they’ve worked their first 12 years of schooling toward getting into that story, and suddenly they’ve seen that whole story break apart. I think if you’re much earlier on, it’s an important time for both students and their parents to realize that the traditional way of optimizing for outputs, optimizing for measurement of output, which is what the education system is based on, is not sufficient. You have to figure out a way to constantly create your own learning architecture, and the only way you can do that is through small experiments, through being entrepreneurial, through gathering feedback.
The answer, again, is not as simplistic as: hard skills are going away, soft skills are important, specialists are not going to be important, generalists are going to be important. This is how we try to bucket solutions, but it’s not quite as binary as that. The only thing we can be entirely sure about is that if you have an iterative, learning-based approach toward your work, if you assume uncertainty as the constant, what that means is you have to have an aim on where things are headed, place a bet on that, think in terms of placing that bet, and constantly update your model of where things are headed based on what you’re learning on a daily basis. That’s the only way you can build your own learning architecture.
You have to constantly assume that where you play is uncertain. It’s not going to be the same as it is today, so you have to bet on that. I would say that’s something that’s true for everybody in jobs as well. Today your company is dealing with uncertainty. Everybody has a view on where AI is taking things, and again, it’s not just AI, it’s many forces. But in the midst of uncertainty, you can’t look at your company as the unit that’s going to guide you on the way forward. You have to bet, you have to learn, and you have to give that learning back to the company to signal that you’ve got it figured out even better than they have.
I think that whole flip, from “I’m going to rely on somebody else to guide my career” versus “I’m going to be at the frontier of betting and learning what’s happening, so that everybody else who’s confronted with uncertainty can see that I’m one step ahead,” I think that flip is going to be really important. So I know I didn’t answer the youngster question really well. It’s a very difficult question. But I believe there’s some version of this that applies at every level, where you have to unlearn the old game, just assume that at some point it’s going to go away from you, and then get comfortable with betting on uncertainty, running experiments, and, based on your learning, constantly changing your roadmap of what you’re betting on and how you’re getting there.
Jim O’Shaughnessy
The thing that is ironic is that, in order to become a great success in the new world, you have to select against the basic coding of the human OS. In other words, what’s one of our biggest biases? Confirmation bias. We seek information that proves, in quotes, that we are right, where we should be seeking information that proves or suggests we are wrong. And high agency versus low agency: expecting tutelage and a coach who’s been through the game to be able to help you is kind of a norm. And yet what you are articulating, and I think correctly articulating, is that we’re going to have to learn to try to unlearn these deeply embedded biases in ourselves. When you were talking about the person who thinks, my game is A-plus and everything else, I’m the exception.
I used to have a GIF that I would put up on Twitter to people who were saying something like that. I said, “This must be your screensaver.” And it was a thing that says, “You are the only exception.” I know that I am not the only exception. I know that I am just as fallible and just as outdated and everything. But it took a lot of time for me to habituate this self-interrogation. For you to be able to continue to iterate and succeed, you have to be able to exit an old mental model. That is really easy to say and it is really difficult to do. So I definitely agree with you that the younger people, honestly, they’re going to be native to this environment, and so they probably will excel.
We have some young, as you know, we do the fellowships every year. And oh my God, the brilliance that is out there among people who are still teenagers. You can literally see it in the way they construct their proposals. They are intuitively doing many of the things you are advocating. Whereas people, well, certainly people of my age, a lot of people who are in their 60s, are like, yeah, I have no incentive to be involved in that. Like your board member. I’m only here another three years. So their incentives are to either be reticent or even to actively push back. Whereas the younger people have no such incentives, and they’re doing fine.
I think you nailed it with the people who really need the advice you just gave, which is people who are already successful in their careers. They got successful in a certain way, you would say, by focusing on the outputs as opposed to the outcomes, inputs, excuse me. But that itself would be incredibly useful. Just that conversation would be incredibly useful, because one of my worries that I’ve expressed repeatedly is there is going to be a cohort of people who, through, and I underline this part, no fault of their own, are just not calibrated or designed for this new world. And I would love to come up with a way where we can do everything possible as a society to ameliorate, to in some way make that blow land less hard.
Because you see all the back-and-forth right now. I did not anticipate the anti-AI response. Well, I did a little bit from the old hierarchical guard, that’s kind of a given. But from younger people as well. Because it’s like, whoa, all the rules are changing, and now you want me to change. And unlearning is hard. The ability to unlearn, I think, is much more important, really, than the ability to learn and then habituate. You’ve got to constantly be unlearning. All models are wrong, some are useful. If you take that attitude, you can switch models relatively easily. You can iterate, you can change. How, in your career, have you changed? I mean, in AI time, your book was published in 2025, that’s decades ago in AI time. How has your own outlook and your own process changed from when you actually published the book to now?
Sangeet Paul Choudary
Well, there are a few different points here. One is that when I wrote the book, I wanted it to be timeless in terms of concepts, if not in examples, because I knew it would not be timeless in examples if I wrote it based on 2025 AI. So I rely a lot on historical analogies and so on. First, over the past year, I’ve tried to constantly calibrate those concepts and see if, with everything that’s happening, those concepts play out reliably. It’s helped me fine-tune where it applies and where it does not apply. It’s leading to an updated version, an updated edition this year, which will talk about things with a lot more nuance.
The first-order effect of that was just that I would have to launch an updated version of the book much faster than I normally would. But then that would be the same output fallacy, right? Because that would mean the faster AI moves, the faster I need to keep writing books, and that’s not going to work. What I’ve tried to do with this new edition, which I’m still working on, is that before I write anything, and this is not why I built the knowledge graph, but now that I’ve built the knowledge graph, I’ve actually calibrated the book through the knowledge graph to see where it’s consistent with what the knowledge graph would predict and where it’s not.
The fun of having this knowledge graph for me is that on a daily basis I’m seeing exactly what you talked about, things that I’ve talked about in the past, things I’ve curated in the past, but where I’m not seeing a certain connection. I’ve seen those connections on a daily basis. Some of them get me excited, so I want to dig down further. So it’s sort of curiosity on steroids. If you’re naturally curious, then having this, it’s one thing to have an LLM that’s constantly finding interesting stuff your way, but you don’t always want that.
For me, the way the knowledge graph has transformed the way I work is that I can just wake up in the morning and ask it to throw me three surprising things, or something that appears a lot in my work but that I’ve not connected to something that appears very little in my work, or it helps me see a lot of gaps in my own conceptual frame. The reason I like it is that I want to keep thinking, and I want to use AI to help me keep thinking even better. When I think, the gaps are very often that I waste time trying to figure out unknowns.
What the knowledge graph helps me do, because it’s structured around knowing what I already know, is that it has pre-scored concepts where I’ve done a lot of my work. It’s able to identify unknowns that are just at the level of the right delta, that stretch me a little bit to connect a well-known concept to an unknown concept, to see it happen in a fundamentally different industry, to connect an unconnected historical precedent to it. So I’ve created it as a way to help me think, which I don’t think I get with an LLM. Because what happens with an LLM is I start with a question, it gives me responses, and in sifting through those responses I learn how to ask better questions. But it’s a very tedious way of learning how to think.
At the end of a long session, I may come out of it having learned something new, but because it doesn’t know what I already know, it doesn’t know how my mind is structured, it’s not able to push me in the right way to increase my thinking. So it may sound a little theoretical as I say it, but only after having created that knowledge graph did I realize how important this idea of identifying the unknowns becomes. The second thing, take away AI and all of these things, one of the things that has really liberated me ever since I decided that my previous game is not going to be worth playing anymore and there’s going to be a fundamentally new game, is that it set me free to rethink things.
If I’m going to find a new game from scratch, I cannot optimize it around anything except longevity and fun. It should be something that is organically fun for me, something I have the incentive and the interest to keep playing long enough. Hence it cannot be optimized around a well-structured game, which was very much optimized around certain markers of success that the game led to. So here, if I don’t know those things beforehand, the only way to keep winning is to stay in the game and keep changing with the game. That has set me free, because if I start with, is it fun, is it something I want to continue doing rather than keep getting burnt out, it helps me take a different approach to my work.
One of the positive outcomes of having to rethink your game has been that the only way you can play a game in a constantly changing playing field, which is what AI and other things are doing to us today, is to ensure you’re in the game long enough and that you’re enjoying playing it. So there are two ways I can respond to what’s happening. I can say I want to play the old game to its maximum for as long as it works, or I want to reinvent myself toward the new game. If I’m doing that, I might as well play it for the next 20, 30 years. In order to do that, I should have fun while doing it.
So what all of this has forced me to think about is, if I’m rethinking what it means to be somebody who has traditionally used, say, books and certain very specific forms of narrative to address this overall point of how things are changing around us, how should industries think about this change, how should companies think about strategy, that overarching question that has always guided my work, which is how does technological change force companies and individuals to rethink how they should compete, if that question continues to be of interest, I would rather create the surrounding infrastructure that helps me stay focused on answering that question, if it’s still relevant, and do it in a way that I continue to enjoy for the long term. So, long story short, the key thing I believe has been really refreshing about having to just stop and forget what I was doing in the past, and the thing I’m doing, is that it gives me the opportunity to build it fun-first, joy-first, rather than speed-first or accuracy-first, all of which actually impose a lot of overheads. I think if I can keep focused on that and the underlying capabilities keep improving, then the notion of scale would also change. You should be in a position where, if you are a fun-first or joy-first company or person or career, you will have the ability, as the underlying capabilities improve, to keep scaling your impact with that. So that’s the way I think about it.
The temptation always is to get faster at what I’m doing today, but I think that just traps you in playing a game that’s going to end soon.
Jim O’Shaughnessy
That’s one of the things I’ve been so excited about, literally because of all these journals. I have been waiting for AI since I was 22 years old in 1982, because the possibility space, learning new things, recombining things, to me that’s incredibly fun and invigorating, and, as you would put it, joyful. I think that the old idea of the nine-to-five job, the whole structure of the way our economy works, is going to change, and I personally think it’s not going to be easy. So I’m not Pollyannish or Panglossian about the whole thing. There’s going to be a lot of disruption, and that’s kind of a given in any kind of transformational technology.
If you can look at it the way you’ve just articulated, I think you’re going to have a much easier go of it than if you’re a type that really believes something because that’s the way it’s always been done here. Because you’re not going to survive in the new environment, right?
Sangeet Paul Choudary
Yeah. I know there’s a certain level of privilege associated with building a fun-first career with constant learning and constant experiments, because we have to think about that alongside this whole point of how there’s growing wealth concentration, there’s growing opportunity polarization, and things of that sort. So that is the context within which it would apply to many people. But the one thing that remains immutable to a large extent is that if you just focus on the fact that the old game is going away, and you either get mad about that or you figure out how to play it better or play it faster, that’s a clear path toward irrelevance.
So given every set of constraints that you personally have, you have to figure out how to get to the new game. Everybody will have their own set of constraints. Some will start from a place of more privilege than others, and that’s to some extent how things have always been. But the one thing I think is going to be common to everyone is that if you keep simply focusing on playing the old game, you’re not going to get to a win that’s going to be significant, because winning the wrong game is actually losing.
Jim O’Shaughnessy
Yeah. So if I were to summarize, it would be: you’ve got to relearn, unlearn. Because so many of us have habituated ways of doing things that we’re not even aware of them.
Sangeet Paul Choudary
Right.
Jim O’Shaughnessy
And I’ve undertaken a series of new projects, one of which is writing my first fiction book. And boy, when I was starting, I’ve written four non-fiction books, and the correlation between fiction and non-fiction I found is basically zero. So I’ve had to learn it, but it’s been really fun to learn how to do it. And I couldn’t have learned as quickly as I learned without AI. I did the traditional thing and read some books on how to write fiction, and I even read a couple about how to write a screenplay. Then I watched some YouTube videos, and I’m like, I don’t think I’ve learned anything at all. So I just went right to our AI platform and started, and then I designed a very mean but very truthful editor that the AI simulated. And it basically came back, “This is horrible, this is awful.”
But it’s interesting, because it’s a whole different conversation. It’s easier to take that advice from a machine than from another person, because from another person I actually experienced it. The first thing I ever wrote for publication, my wife graduated summa cum laude with a degree in journalism, and I was the editor of my high school newspaper. I thought, oh, I know how to write all this stuff. So I wrote it and said, hey, would you take a look at this? This is back in the 90s. I actually handed her the paper that I wrote, and it came back a sea of red. And I got mad, and it became, how can you say this to me, and all of that. I was recounting it with her. And by the way, she was entirely right.
All of her edits made it a vastly better piece of writing. But the point is, we were chatting about it, and I said, so I modeled you into the AI, and they’re wicked mean to me, but I don’t take it personally. And I just think that your underlying factors of curiosity, the joy of experimenting, and the ability to understand that everything’s an iteration, you’re not going to get the hole-in-one every time on the golf course. It seems like selecting for those, again, is selecting against things that we’ve tried to beat out of children in the previous era. Like what was much of public education in the 19th and 20th century?
But no, installing a correct-answer machine in the brain and driving curiosity and all of that kind of stuff out, because your society wanted you to follow orders. Society wanted you to be able to understand those orders and then execute them. So I definitely think this touches not only business, not only society, but the way we learn. And I think it’s interesting to me, because I think we have to be open to every conceivable permutation here. Most of them won’t work, but we’ll learn from what doesn’t work. Your knowledge graph is a great example. It isn’t the large language model that is really giving you all of these super new ideas, it’s your knowledge graph. And people sometimes forget that part of it.
Who knows, maybe at some point in the future AI will get creative and will get all of those things. I have no idea, and I lack the expertise to even comment on whether that is possible. But for now, that isn’t the way it works. It’s a tool. It’s a tool in the hands of a deeply curious person who is comfortable with uncertainty, the best tool that’s ever been invented, certainly in my lifetime. It’s hard for me to articulate how excited I am by it, because it really leads to a whole new paradigm. How do you deal, because you’re dealing with external institutions, and I don’t have that friction, how do you overcome a very deeply skeptical board? Back to the board. Where the CEO answers to the board, and the board is likely made up of people close to my age whose incentives are very different, as you pointed out. Three years and out. How do you convince them, you really have got to take this seriously?
Sangeet Paul Choudary
Yeah. There are quite a few things there. And I want to reiterate the summary you brought in, which brings together a lot of the points we’ve been talking about. It’s a combination of uncertainty in the playing field, with constantly changing and ever-improving capabilities that are available at your disposal, that allows you to innovate in the center and manage that uncertainty to your advantage. That is what the new game is. The difference between a business and an individual is that an individual has emotional constraints and capacity constraints at a different level. That’s why fun-centric becomes very important. But for a business as well, you have to think about what your constraints are, within which you can operate. You have to know that what you’re really going for is, there’s uncertainty outside.
There’s a new capability set internally that creates the opportunity for you, within the unique constraints of your business, within whatever identity you have created for yourself, first, to potentially question that identity, but second, within those constraints, and I wouldn’t say within those constraints, but in cognizance of those constraints, to think about how you can leverage these capabilities to do something fundamentally new, to create a fundamentally new way to compete in the playing field you define. In general, my point to any of these so-called skeptical boards is, by default there’s a certain filter in terms of what comes my way.
So I don’t necessarily speak to a company that’s at step zero. It’s typically to companies at step four out of five, where they’ve kind of done the adoption, they’ve tried to accelerate things internally, they’ve bought into the hype but not seen the results, and then they’ve realized that what I’m talking about is how to change that game, how to think about strategy in a fundamentally different way. So I don’t necessarily have to go in and tell them that AI is important. They sort of know that, everybody’s talking about it, they’ve seen the value internally. They just haven’t yet figured out what that means for them in terms of whether they should continue doing what they’re doing or fundamentally flip.
So that is where I essentially make the case that you can place a bet on the fact that things are not going to change externally. You’re going to have better capability, you’re already adopting it, so you’re going to improve how you do what you do today. But you have to start placing bets on what if things change externally. And to place those bets, you have to start by saying, what are those scenarios? At the end of it, you could come back and say, it’s not going to happen in my relevant timeline, and that’s perfectly fine, because the difference between having a view on what’s going to happen and betting on it is the risk you are willing to take against the timeline.
So it’s a perfectly valid statement to come back and say, I’m not willing to place that bet, because I don’t believe in that timeline. You’re accepting that risk, and that’s fine. That’s how allocation works in general. But you have to do the hard work of thinking through what all those scenarios are, and what your firm looks like in those scenarios. And be very cognizant of the fact that you have to place those bets. You cannot place those bets by omission, the fact that you did not place them because you did not run the scenarios. So you have to do that. At the end of it, if you still believe it’s not going to play out, that’s just the best bet you’ve placed.
There are going to be many situations where companies optimistically, proactively place bets and they don’t play out. And there are going to be situations where companies do not place bets even after seeing the scenarios. But not placing bets because you didn’t do the hard work of seeing the scenarios, didn’t do the hard work of projecting what’s going to happen, didn’t think through how these shifts are playing out, that is the failure case. Because if you do the hard work of developing the scenarios and placing the bets, you’ve already taken some steps to say this is going to happen, that’s not going to happen.
Which means you’ve already perked your sensors to keep sensing whether those bets are going to play out, which means you’ve already started fine-tuning a learning architecture and a feedback loop to see if that’s going to happen. With that, you’ll keep tweaking your scenarios, you’ll keep tweaking your bets. But if you never did that, you would just be an ostrich with your head in the ground, not thinking about what’s going to happen. And so there is no architecture, no feedback loop, nothing created. So that’s what’s important. It’s less about “be the fast mover, do this, do that.” It’s okay to not be the fast mover if that’s deliberate, but it’s not okay to not be the fast mover if that happens out of omission.
Jim O’Shaughnessy
Yeah. And again, our alignment on how we look at this is strikingly similar. Spending a lot of time thinking about failure modes is, in my opinion, not only required but incredibly instructive. Because if you’re a naturally optimistic person, you might have a hard time thinking about all the ways things could go wrong. And the other thing AI can do is simulate cognitive diversity. I’m constantly playing with coming up with the most pessimistic model I can possibly design, but making it very smart. And what’s really interesting when you do that is your blind spots are revealed. We all have blind spots, every single one of us. Having this kind of red-teaming of our own ideas, I just think that is such a valuable thing to have.
And yet a lot of people look at me like I have three heads when I’m like, no, you’ve got to think about all the, not only that, you’ve got to give it to an intelligence that is going to find those blind spots that you have. So, well, listen, I always love talking to you and admire your work. You’re working on the new version of the book that comes out when? Later this year. Perfect. I would love to have you back on again when that comes out, because a topic we didn’t even touch is, what is going to happen to books, your books in particular? Are we going to move to a new way of packaging a book so that it can be constantly updated, so that it can be interrogated?
I think all of those things are going to be super useful, and I would love your view on them.
Sangeet Paul Choudary
Yeah, absolutely, I would love to. It’s a topic I’m constantly thinking about. I would encourage whoever is listening to go to reshufflebook.com. It’s the companion site to the book, where what I’ve tried to do is extract the concepts from the book and show them as a navigable graph, so you can see which concept connects to what and work on applying that. I’ve tried to create different narratives, all of which are again generated by the knowledge graph, because once you map all the concepts, it generates different narratives. So for different personas, whether you’re thinking about it in terms of jobs or in terms of how you design the workforce and the organization, there are different tracks in which you can go through the book.
So that’s one attempt to think about how you unbundle and rebundle the book in new forms. But the larger thing I’m really going for, and I’ll just throw it out here and would love to discuss this again, is that the book was not an optimal bundle for a lot of things it was used for. It was just forced on a lot of formats. What I mean is that I still believe a big idea with a persuasive narrative requires a book. It cannot be done with questions to an LLM, where everybody’s going to get a different version of the idea in a different narrative.
But a lot of how-to advice, which has so far been packaged in a book, is much better accessed in the context of your work in an LLM and so on. So we’ll have to really rethink the relationship between the jobs we wanted to do and the formats that were forcing those jobs on us. There’s a lot of discussion there, and I look forward to that topic.
Jim O’Shaughnessy
Yeah, the user interface is going to change, there’s no doubt about that. Sangeet, always a pleasure talking to you. As you might recall, our final question here is: we make you emperor of the world. You can’t kill anyone, and you can’t put anyone in a re-education camp against their will. But what we can do is hand you a magic microphone, and you can say two things into it. And those two things are going to incept the entire population of the world. They’re going to wake up whenever their morning is, and they’re going to think the two things you say now were their own ideas. And they’re going to say, unlike all the other times where I had these ideas, shower ideas or waking-up ideas, I’m going to act on these two. What are you going to incept in the world’s population?
Sangeet Paul Choudary
Assuming that these ideas can be personalized by the individual, I would want to create the drive for every individual to figure out what they are insanely curious about, which they would want to do just for the love of doing it, which they would want to think about just for the love of thinking about it. I think what a lot of us are lacking is that by outsourcing a lot of our thinking, we are not seeing the incentive for getting into training our thinking machine, our brain. So creating an organic incentive mechanism for that, which is why I said it would be personalized. Everybody wakes up figuring out, here’s what I’m insanely curious about, I’m going to start figuring out how to go after this, and here’s how I’m going to.
At the same time, I would want everyone to think about it as, if I’m so curious about this, I want to do it as long as I can. So how can I do it in a way that I create the conditions for me to constantly think about it? Because if my conditions don’t support this kind of curiosity, and if I’m constrained in other ways to do things that are more short-term, that comes in the way of doing it. So I don’t know if that answers the question directly, but that’s great. Leave that with people.
Jim O’Shaughnessy
The way I would rephrase it is, we’re incepting everyone to wake up thinking, I am insanely curious about this, how can I go about satisfying that curiosity, and sustaining it, and building it into something I can continue to do. Okay, so that’s your first one. What’s your second?
Sangeet Paul Choudary
I think the second piece is the sustaining part, in terms of how you sustain it. It’s about being very clear that you don’t simply, it’s easy to be curious about something, it’s very difficult to sustain it. And in order to sustain something, it’s not just about the fun, but about having the discipline to sustain it. So yes, it’s important to find what you’re insanely curious about, but you should be disciplined enough, you should care about your curiosity so much, that in order to protect it and do it on an ongoing basis, you create the discipline and the conditions around it that can help sustain it. So the second piece is really about getting to that point where you know this is not just a passion, it’s a discipline. And how do you bring those two things together.
Jim O’Shaughnessy
I love them both. What’s the old saying about curiosity being the cure for boredom, and there is no cure for curiosity? Sangeet, thank you so much for your time and your ideas, and I can’t wait for our next conversation.
Sangeet Paul Choudary
I look forward to it. It’s been such a pleasure. Thank you once again.
