Wardley Mapping and the Art of Strategy – with Simon Wardley

Danny Buerkli: My guest today is Simon Wardley. Simon is a former CEO, scholar, and most famously the inventor — or discoverer — of Wardley Mapping, the most wildly useful way of thinking about strategy and much else. Simon, welcome.

Simon Wardley: Fantastic to be here. I’ll have to push back on the “scholar.” I’m a terrible academic — so lousy. I stopped at my second master’s, so I’m not anything like a scholar. The mapping stuff was all accidental discoveries. They drag me in to teach at a couple of places, a day here and there, which is extremely kind. There are some academics who look into the maps, but they’re academics. I’m just a guy who wandered along.

Danny: Well, I’ll push back on the pushback — which is why I said scholar rather than academic. These are two separate things.

Simon: Fair point.

Danny: With that out of the way, we’ll be referencing mapping in this conversation a fair bit. To the listeners: if you haven’t seen it, it is a visual method and benefits from some visual aids. Pause and follow one of the links in the show notes, then come back. If you haven’t read anything about Wardley Mapping, trust me, it is very much worth your time.

Simon: First of all, that’s really kind. There’s an entire community around mapping. There’s my site, wardleymaps.com. Chris does Wardley Maps, Ben does Learn Wardley Mapping, and there’s a whole community built around this stuff. If people just search for it, they’ll find something interesting.

Danny: Yes. And it’s one of the things that, in a generous — and strategic — act, you Creative Commons-licensed much of your work.

Simon: Pretty much all of it.

Danny: Which means you find a lot of really good material out there, including by some of the people you mentioned. To start us off: which Sun Tzu translation is your favorite, and why?

Simon: Oh, no. They’re all translations.

Let me go back to how I started getting into mapping, because it’s connected to this idea of translations. I was running a company, and we were being very successful, but I had absolutely no idea what I was doing. I was completely clueless — I was the “fake CEO,” as I called myself. I started reading every trashy book I could find. It was getting nowhere.

I ended up in this bookshop in Charing Cross, talking to the bookseller — a wonderful person. She asked whether I’d read Sun Tzu’s The Art of War, which I hadn’t. She persuaded me to buy two different versions of the book — partly because she was a fantastic bookseller, partly because they’re all translations. It was in reading the second version that I noticed a pattern between the two.

Sun Tzu fundamentally talked about five factors that matter in competition. When we talk about competition, we’re talking about groups of people seeking something — knowledge, resources, whatever. You can do that through various forms: you can fight others (conflict), labor with others (collaboration), help others (cooperation). All of that is competition.

When you’re in competition, having a purpose or moral imperative is a pretty good thing to do. Then you need to understand your landscape — the environment you’re competing in — and we compete over many landscapes: technological, economic, social, political, legal. Then you need to understand the heavens, as it was called — the climatic patterns, how the weather changes the landscape. Then you need principles for how you organize and structure yourself. And then finally you’re into leadership and gameplay.

I only picked up that pattern in reading the second version. So I can’t say either is my favorite — it’s the combination. I won’t pick a favorite. I like them all.

Danny: You define strategy as “the art of manipulating an environment to gain a desirable outcome,” which I’ll note is very similar to Stuart Russell’s definition of intelligence — acting to achieve your objective given what you perceive. Why that definition of strategy?

Simon: What a good question. The realization that those factors matter took me into a journey of trying to understand my environment, particularly through mapping it. I discovered that everything I had which was called a “map” wasn’t actually a map, so I had to develop a way of mapping. I came across John Boyd and the OODA loop, which dovetailed nicely with many of these ideas.

That brought me to: what is gameplay? Gameplays are the methods and techniques you can apply to a landscape — and they’re highly contextual. A principle is universally useful, like “focus on user needs.” A climatic pattern, like “everything will evolve, driven by supply-and-demand competition,” will happen regardless of what you do. The gameplays are the things you choose to do to move the environment into a situation more favorable for your purpose.

Open sourcing is a great example. There are evil ones too — fear, uncertainty, and doubt to slow a change down. All highly contextual. They depend on the environment and the stage of evolution of the components in it.

I should describe what a map actually is, since people can’t visualize it. All maps have three things in common: an anchor (such as magnetic north), the position of pieces on the map (north, south, east, or west of this), and consistency of movement — if I’m going north, I’m going north. Geographical maps are a bit easier than economic and technological maps because the tectonic plates don’t move that fast. If I want to go from Beaminster, where I live, to Paris, it’s highly unlikely Paris will have moved during the journey. If I took a hundred million years, then Paris would be somewhat different — if it exists at all.

The problem with economic and technological maps is that the landscape moves much faster. So how do you create those characteristics? For the anchor, I focused on the users. In my first map, I had consumers; I could have had regulators; for space, you could have astronauts. I put users at the top of the map.

Then I ask: what do they need? Take a simple tea shop. A consumer needs a cup of tea. A cup of tea needs a cup, tea, hot water. Hot water needs cold water, a kettle. The kettle needs power. You create a chain of needs. Within an organization we call that a value chain; across organizations, a supply chain. Now you have anchor and position.

That’s still not enough. You have to capture consistency of movement, even though the tectonic plates are shifting. You do that by measuring in terms of change itself. I go through all the nodes on my graph and ask a simple question: how evolved is this component? There’s a common pattern for how things evolve: from the genesis of novel and new items, to custom-built examples, to products and rental services, to commodity and utility services. That pattern actually holds for pretty much all forms of capital — data, knowledge, even values. The terms differ slightly. For values, you’ve got the concept of something, then emerging values, then converging where you’re getting agreement (like a product), and eventually accepted, settled matter (commodity).

How do you get it perfect? You don’t — all maps are imperfect representations of a space. To create a perfect map of Paris, it would have to be the same scale as Paris, which is completely useless. So maps are imperfect representations, but they’re vehicles for discussing and communicating our understanding of the space. That’s what they’re good at.

The strategy question is simply: how do you change this environment, applying context-specific patterns to achieve your purpose? That’s where the definition came from. Is it the right definition? No. It’s my definition.

Danny: It’s a useful one. Tyler Cowen has this good line — context is that which is scarce. I think of these maps as providing the context for many strategic decisions, which is often otherwise missing. Without it, you end up in cargo-cult territory.

Simon: One thing worth pointing out: management practices take a long time to change — thirty to fifty years on average. Look at the work of a friend of mine, Kent Beck, extreme programming, the concepts of agile development, things like test-driven development. We’re talking twenty-seven years. If I get a conference full of people and ask, “Who does test-driven development?”, about 70% will put their hands up. If you dig in, it’s two or three percent. A massive collapse. It’s the same with agile — people say “we do agile”; when you look, they don’t. It’s still spreading, but you’ve already got twenty-seven years for it.

1999 was when I picked up Beck’s book, Extreme Programming Explained. The Agile Manifesto was 2001. People are still learning. So thirty to fifty years. The fact that I turn up to a conference twenty-one years after I started mapping and 5%, 8%, sometimes 10% of the audience say they’ve heard of mapping — that’s just amazing. These things take time to spread.

Danny: Quick detail: is your forecast for mapping being ubiquitous still 2035? I think that was your estimate at one point.

Simon: Did I say 2035? Terrible. I said thirty to fifty years. From 2005, 2035 is when it would be becoming widespread. We’re allowed fifty years as well — between 2035 and 2055, somewhere around there. Or dead. Or it’s reached its bottleneck, found its niche, never going any further. Give it another nine years and we’ll have a better idea. I might turn up to events where 60% of people say they do mapping and three or four percent actually do — and I’ll be in the same boat as Kent. That would be wonderful. Hopefully, someone finds a better way of mapping.

Danny: We’ll see about that. One quick technicality, because I think it causes a lot of confusion: what is the difference between evolutionary movement — Genesis to custom-made to product/rental to commodity/utility — and maturity?

Simon: Do you mean adoption?

Danny: Yes.

Simon: Early adopters to laggards. So back in 2004 I started playing with the idea of mapping. I thought it was going to be super easy. I’d stumbled on the idea of the users and the chain, and I thought I needed this movement piece. I’d use maturity — wouldn’t that be fantastic? I knew Everett Rogers’s work on diffusion of innovations, which was made famous by Geoffrey Moore, who cut out that chasm.

I would go through components and ask how mature each was. The problem was: I’d go to an audience and say, “Mobile phone — who has one?” Everybody would say they had one. Widespread. Some have two. I’d say something daft like, “That means it’s pretty evolved — it must be a commodity.” And everyone would go, “Oh, no, no.” Right. “Gold bars — who’s got one?” Nobody. “So gold bars aren’t a commodity?” “Well, no…” What is going on here?

I started collecting data. I knew electricity evolved — Parthian battery, Hippolyte Pixii, Siemens generators, eventually Tesla, Westinghouse, eventually utility electricity. I thought surely it would just be how widespread something is. I could find no connection. It was driving me mad. I’d read reports saying “compute is widespread, it’s a dumb business,” and you’d think: hang on, it’s very much a product at the moment — what happens when it becomes a utility? Something was missing.

I’d almost given up, but I’d collected — initially over 6,000, eventually 9,223 — publications. I noticed something odd: an old advert that said, “This room is equipped with Edison electric light bulbs. Do not try to light with a match.” My favorite was an image of an early phone with instructions telling you how to hold it — which bit to speak into, which bit to listen to. We don’t tell people that these days, but there must have been a time.

I started pulling out themes — the wonder of something, operation and maintenance, feature differentiation, or when we stop talking about the thing and just talk about how to use it. These publications fell into a lovely pattern: Genesis, custom-built, product, commodity — based on how certain we are about something and how ubiquitous it is in its final market.

The point is: things diffuse. Compute diffused from early adopters to laggards. Then a new wave of compute diffused from early adopters to laggards. Then multiple product waves. Then eventually a new wave — utility — which was cloud. And of course you get the early adopters of cloud while everyone else, the laggards, is still on compute as a product. In fact, many of the early adopters of cloud were the laggards of the previous phase. It’s difficult to tell.

So evolution is built from many — could be ten, a hundred, a thousand — waves of diffusion. One problem: we can only define the applicable market once something has become a commodity.

Danny: Right.

Simon: Once it’s a commodity, we can go back and look through history and say exactly where it was. Until then we have to cheat. I have something called the “cheat sheet” with a long list of characteristics for each stage. You can only tell the past with accuracy. Once it becomes a commodity, you can see the past. Otherwise, we have to tease out weak signals. The maps are imperfect. I don’t have a crystal ball.

Danny: Sticking with mapping as a method — it’s a tool, it has limitations. One you’ve highlighted is that the evolutionary flow only emerges under conditions of competition. That may not be the most important limitation, since competition shows up in many places and forms. What are some non-obvious limitations?

Simon: You have to have competition — I don’t know how to map without it. As you say, that can be in any space. Secondly, you have to accept that it’s an imperfect representation. If you could precisely say where something was on the evolution scale before it became a commodity, you’d be predicting the future — you’d have a crystal ball. So you have to accept uncertainty. Evolution itself is ubiquity versus certainty, and we only become certain once something becomes a commodity. Uncertainty is inherent.

There’s a U-shape: it’s easier to say something is in Genesis or clearly in commodity. The tricky bit is what’s in between. We know this is totally novel; we know that’s a commodity. The bits in between are hard.

One of mapping’s strengths is that organizations normally run on stories, and we tell everyone that great leaders are great storytellers. The problem is that when you challenge somebody’s story, you challenge their leadership, which is why they get defensive. Putting the story on a map lets me challenge the map — not the person. So it’s good for negotiation and communication. You can bring the ideas of many people together onto a map, which is why solo mapping is generally not a good idea. People feel uncomfortable, and I don’t know how to get past that — they try to create a perfect map. You have to get used to the uncertainty, accept it’s imperfect, and invite challenge. Others will have bits of the puzzle you don’t.

So one weakness: a tendency to solo map. I tell people not to. They still do. The other weakness: a map is a representation, not the territory. The map is missing bits. It’s better than nothing, but it isn’t the territory.

If you can cope with the uncertainty, the imperfection, the challenge, and the fact that it’s a representation rather than the landscape itself — then hopefully you find it useful.

Danny: Your point about the U-shape — it’s easier to tell us something is really novel or that it’s fully commoditized than the bits in between. The reason that matters, presumably, is that the distinction between custom-made and product is very meaningful. You had a really interesting point about the dispute a few years ago between Clayton Christensen, of disruptive innovation fame, and Jill Lepore —

Simon: Oh, yes.

Danny: — the historian. I believe Jill Lepore’s charge was that disruptive innovation was a lot of nonsense, didn’t compute.

Simon: Well — that it wasn’t predictable.

Danny: Yes, exactly. Your point was that we’re actually talking about two different kinds of disruption, and if you look at it on that continuum, it starts to make sense.

Simon: Sure. When I think about how things evolve — Genesis to custom-built to product to commodity — there are various forms of substitution. You can get one product substituted by another product. And you can have a product substituted by a commodity, a utility.

When it comes to predictability, there are very simple things where we can say what and when, roughly. There are lots of things where we can say what but not when, or when but not what. And then there are things we can say almost nothing about.

When you talk about product substitution — product replacing product — that turns out to be highly unpredictable. Christensen said that, I think, Nokia would beat Apple. He was wrong. But that’s not because he’s daft, it’s because it’s highly unpredictable. Product-to-utility substitution — crossing a boundary rather than staying within one — is much more predictable. You can give an idea of what will happen: change of practices, punctuated equilibrium, an explosion of new higher-order systems. You can talk about Jevons paradox. There’s a long list of things you can say. There are even signals for getting a sense of when it will happen.

Then suddenly this whole debate appeared — Lepore versus Christensen. One side: disruption is predictable. The other: it isn’t. I was looking at the map and thinking, well, you’re both right and both wrong, because there’s more than one form of disruption. Product-to-product disruption is highly unpredictable. Product-to-utility disruption is much more predictable. The debate was interesting, and I think they were both right and both wrong — which is a great way to be shouted at by fans of either.

Danny: You have an impressive record when it comes to — I don’t know if “predictions” is quite the right word — pointing not only at what we may see, but interestingly at when we’ll see it.

Simon: I cheat.

Danny: Tell us how you cheat.

Simon: I only say things that are shifting from product to utility. I talk about cloud, the shift of compute from product to utility. Serverless — the shift of runtime from product to utility. Machine learning — the shift again to utility with things like large language models. Then I talk about the common identifiable patterns. Every time we go through this change, we get coevolution of practice. With cloud we eventually called it DevOps; Andrew and Patrick coined “DevOps” several years later. With serverless, several years later, we called it FinOps.

That’s how, in 2018, I could map out the space. I knew machine learning was evolving and was going to cause higher-order systems and new practices. It wasn’t particularly interesting unless I combined it with the coevolved practices of serverless, particularly with the no-code stuff. So I converged those and came up with conversational programming. A friend of mine then presented it at AWS re:Invent — he built an AI system you had a verbal conversation with, and it built other systems for you.

Danny: Also known as vibe coding. Well, it’s the same thing.

Simon: It is. But this was 2016, 2017. The underlying components have evolved much more, and everybody now has access to it. I’m “cheating” by talking about coevolution, punctuated equilibrium, Jevons paradox — all of which apply to the same situation.

There’s a wonderful piece of work I just found by the World Economic Forum on technology convergence, using the evolution and mapping ideas; I talked to them as well. They highlight which components are industrializing and how they combine to create new opportunities. In the mapping world that’s like, “Surely everyone does that,” and then you discover, as the WEF points out, no, they don’t. Which is a bit of a shock, but it’s great they’ve done it. So yes, I cheat.

Danny: In 2014 you said intelligent agents would take off between 2030 and 2034, I believe. Are we early?

Simon: I don’t think we’re quite there yet. I had it pegged for 2030 to 2035 — within a five-year range. We have a lot of the components. We’re doing a lot of experiments — Open Floor and that sort of stuff is quite interesting. This actually dates back to a talk I gave in February called “Any Given Tuesday,” about a future where you’d have agents organizing your life for you. I think we’re getting close. I might miss by a few years. The components are largely there. We haven’t quite industrialized this yet — we’re still in the learning phase, doing lots of interesting experiments.

Danny: Right. On a similar theme — why are data centers moving to space, and why was that predictable?

Simon: You’ve been looking at my history. Quite some time ago I started mapping space. I did this back in 2014, then again in 2017. I’ve done it more recently looking at things like operational capability in space.

The early maps were focused on the network going into space. I’d meet some of the telcos and satellite companies and point out that the telcos had basically converted themselves into real estate companies — they own the land the towers sit on, often not much more, and were trying to work out how to maximize potential from customers. I was pointing out that the network would go into space, and people said, “That’s gibberish.” Okay, well, we’ll see.

In 2017 or 2018 — I did another one. Around 2019 we came up with the idea of “inner to outer space.” If much of the network goes into low Earth orbit, you’re communicating to the outside of the building, and the question is what goes on inside the building too.

The problem with getting the network into space was launch capability. NASA had been encouraging the industrialization of launch capability for some time, and our ability to build small satellites has gone lockstep with launch capability — you have to balance the two.

Then the question is: if the network is in space, what else can follow? Usually you have network, compute, and storage. Why not put compute in space? There are tricky problems, particularly cooling — it sounds ridiculous, but cooling is a major problem in space. That’s where the idea of data centers in space came from. Back then, that was even more ridiculous than the idea that the network would be dominated by space. People would say, “Isn’t that science-fiction nonsense?”

What follows is shifting manufacturing. There’s a whole set of supply chain issues, but the energy cost of shifting from the Moon to low Earth orbit is actually better than from the ground to low Earth orbit. So if we’re talking about manufacturing in space, we may well be talking about Moon mining. That’s not going to be humans on the Moon with banners and hammers; it’s autonomous robotic systems. This goes back to the WEF technology convergence work — the components are all coming into play.

So in 2018 when I talked about data centers in space, it was simply: once the network is established in space, what’s the next thing to follow? Compute and storage.

Danny: Got it. What is SpimeScript, and are we already seeing it?

Simon: You’ve done your homework. This one connects to space as well.

Back in the late 1990s I had a big interest in 3D printing. The first patent was at the Battelle Memorial Institute back in the 1960s, I think around 1967 — vats of liquid and stereolithography, using lasers. I’d taken a massive fascination in 3D printing in the late 90s. That’s actually why I joined Fotango, the online photo service — I was interested in the distribution of images, because in manufacturing, as we make things more digital, we get the distribution of images-to-manufacture.

I gave a talk at EuroFoo in 2006 with Bre Pettis, who had his cupcake printing — he went on to create MakerBot. So Bre came along, way before MakerBot, and I gave this talk about the future of 3D printing and printed electronics, because Kate, a friend, had worked with Sirringhaus on inkjet-printed transistors and that sort of thing.

The talk had a map in it because I was fascinated by the fact that CAD is very digital but lacks chemical, electrical properties and so on. It’s a digital representation of physical shape, and we have code that interacts with the electronics. The description of the device is itself code, and there’s something manufacturing it. We were very much in 3D printing as a single material, but Washington State University had already done a hybrid printer in 2004 — they made a junction box that had both physical and electronic form. So you’d get hybrid printers.

The interesting thing: if I want a watch, it could be mostly mechanical or mostly electronic. What I really want to do is describe the function of the watch and have a compiler decide what should be mechanical, what should be electronic — and the printer prints it out. And it has to write the code, too. So you end up in a world where I describe the function of something and an advanced compiler decides what should be physical, electronic, or code, and then builds it.

There’s a wonderful book called Shaping Things by Bruce Sterling — I hope I’ve got the name right; if not, find it anyway — about future materials. He had this idea of “spimes” — objects that exist in both physical and digital form. So I said: this is SpimeScript. We’re going to have a new language we can parse. Recently, because I was doing some stuff on space, I came across a group in the US doing something called compiled manufacturing — twenty-odd years later, but it is basically this concept.

When you think about manufacturing in space, we won’t be there straight away — we need robots and other components, we’re not just going to teleport stuff. But eventually we’ll get to the point of describing the function of what we want and having a system create the physical, electronic, and code forms. That’s SpimeScript.

Everyone gets very excited about large language models. The blast radius, combined with practices from serverless, is very interesting in terms of conversational programming. But changing the way we manufacture everything hits every value chain and supply chain. So I think it’s going to be much bigger. We’re getting closer. Not there yet.

Danny: What’s your prediction for when it kicks off?

Simon: I’ve now heard of the first group really doing something in this space. So we’re very much in the Genesis phase. A lot of it depends on how much communication has sped up. I don’t see that communication has sped up too much — it might have, but I haven’t seen the evidence.

Danny: Are LLMs not speeding up the pace of communication and therefore industrialization?

Simon: Things like the postage stamp did. The telephone did. The telegram did. The internet did. The printing press did. LLMs are different — there’s lots of information being spread, but is it speeding up our ability to communicate? I’m not sure.

Danny: Can I push on this for a second? I would have thought yes, because — and I haven’t thought about this particularly deeply — LLMs let you get to the frontier of knowledge much faster than you would have otherwise. So you’d imagine ideas would travel much faster than before.

Simon: Maybe. But I’m one of these people who has to wait and see some data and evidence. The counter-effects could be that yes, people can quickly get to an answer — LLMs are fast at getting you an answer — but it’s not necessarily a representative answer, or even the right answer. There are issues around critical thinking and challenge. They’re good at settled matter, but they tend to take material that’s in the emerging-converging stage and portray it as settled. They’re quite good in Genesis — they’ll say “nobody knows this stuff.”

So is it effectively speeding up communication? Not yet. It’s certainly speeding up our ability to generate code. That’s absolutely true. But that’s not necessarily a good thing either. It can be a good thing if you engineer the system with trust — guardrails, an agentic runtime, all of that. You constrain it, make sure it’s using tools, deterministic methods. You ring-fence it. We’re still learning those practices.

Danny: Right.

Simon: Typically it takes thirty to fifty years to go from Genesis to the edge of commodity — maybe 50 to 70. The previous cycle was 50 to 70. I think we may have accelerated to 30 to 50. Some will argue it’s gotten faster — 20 to 30. I don’t think we’re there, but we’ll see. If SpimeScript is just getting started, then you’re talking thirty to fifty years to become more of a commodity. Once it starts to appear as a commodity, ten to fifteen years to spread and have real impact. So you could be talking about 2030 — picking a figure of about thirty or forty years for it to really start making a difference.

Danny: That’s not that far away, in fact.

Simon: We don’t have crystal balls, unfortunately.

Danny: But the orders of magnitude matter. It makes a difference whether we’re talking about two years, thirty years, or three hundred years. The fact that it’s thirty rather than three hundred is really meaningful.

Simon: The other thing is whether we have all the components we need. If they’re not there, we have to wait for them to evolve. The compiler being able to compile down to physical, electronic, and code — well, we certainly have function-to-code; we’re developing intent-to-code now with LLMs; we have to do the same on the other side. If there are gotchas, it takes longer. Have you ever seen the film Strange Days, where people live in virtual worlds? People often think, “I’ve got a VR headset, the Strange Days world is almost upon us.” But it isn’t — there are huge numbers of components still in Genesis. So thirty or forty years doesn’t seem unreasonable.

Danny: On other predictions — you once had a heat map of different technologies and predictions of when they might kick off. It included social-cultural disruption, kicking off right around now, 2025 to 2030. Then it disappeared. Why did it disappear from your graph?

Simon: First of all, incredible job that you found that. I mapped technological and economic systems, but I mapped social and political ones too. I had weak signals I used to identify roughly when technologies were going to start industrializing — because as I said, we can see the past easily once something has become a commodity, but until then we use weak signals. That’s where the intelligent-agents stuff came from. It’s not perfect — it’s weak signals, layers of uncertainty — but it’s not bad.

I could also apply the method to social systems. I used a particular weak signal — the same one I used for technological and economic systems — and a very strong pattern came back.

Danny: What was that weak signal?

Simon: It’s to do with publication types — the same one I used for technology and economics. I just applied it to the social area, out of curiosity. A strong signal came up. That’s interesting. I didn’t mean to include it in the graph, because in that graph I was talking about a “point of war” — a point of industrialization. “War” is an uncomfortable term; I’m terrible at naming things. There was a signal for a point of disruption — I think it was 2025 to 2030. Unfortunately, the first time I published the graph I accidentally left the social one in there. Somebody immediately commented, and I thought, oops — that was really an experiment. So I removed it.

And then you’re going to ask: what does it mean?

Danny: Indeed. What does it mean, Simon?

Simon: To be honest, I don’t know. This was back in 2014 — there was a strong signal for 2025 to 2030, that there would be social disruption of some form. Do human systems follow the same patterns as technological and economic systems — which are ultimately human systems too? Probably. Is there any specific meaning to the social signal? No. It wasn’t a specific region or location — it’s just a signal. So we’ll leave it at that.

Danny: We’ll leave it there, I’m afraid.

Simon: You’re probably going to hit me with the China stuff now.

Danny: Indeed. You’ve read my mind. First the LLM stuff and then maybe the China stuff — in many ways, it’s the same.

Simon: There’s a connection here. I did a piece of work looking at competition between the US and China.

Danny: This was the 2017 Clash of the Titans report you’re talking about.

Simon: Yes — though the work was done beforehand. I think it was published in January 2017 or around there; the work was done the previous year. It was a whole bunch of Wardley mapping across social, technological, and economic spheres, comparing the US and China. It was when I was visiting China and the US.

There was a range of industries where it was clear the US was going to lose its position. So I wrote this report. I presented it in Washington, DC. I was with a friend, Abraham Jackson. When I was presenting, it was — his phrase, I think, was — like someone had passed wind in a lift. People were not happy. I was quietly ushered out of the room. I was firmly told that Silicon Valley would out-innovate China. I was trying to point out all these issues. There’s real American exceptionalism.

The interesting thing — I think Donald Trump had just announced he was running, and he talked about trade wars or tariffs. I pointed out in the report that that was completely the wrong way to play the game. Obviously he’s played it that way. I’m not a fan of Trump, but he has a use. I talked about this in 2018, 2019: it’s a bit like Anthony Eden. In the UK, a lot of people blame the loss of empire on Anthony Eden. It was going on well before he was on the scene — he had almost nothing to do with it, but he’s a good psychological cushion. The same is true with Trump. In the shift of power from West to East, despite many things I dislike about what he does, he has the potential to be that role — the person others go, “It was his fault.” That’s a really useful psychological cushion for change.

But enough said on that.

Danny: You’ve called LLMs a non-kinetic form of warfare. Can you say more?

Simon: Whenever you build a system, you have functionality and you have structure. Tudor — a chap I work with — runs a company and has done wonderful experiments where they get seven teams of people to build a system. All seven have the same specification, the same language, the same environment, the same tool, the same effort. The environment they use can self-describe the architecture, creating a graphical representation. The architectures look completely different. Why? Because in the process of developing, they’re making lots of tiny decisions all the time, and those decisions get embedded in the code.

Whenever you’re using an LLM to build a system — and remember, we’re shifting to a world where LLMs are a utility, with coevolution of practice, new things built on top, Jevons paradox, an explosion of new systems — we’re going from, say, large enterprises with 30 million lines of code to a billion lines of code. Reading code? Forget it. It was broken when we were at a couple of hundred thousand lines; it’s well and truly busted now. We’re going to get agentic systems building this stuff. Our architectural diagrams are intent, beliefs, wishes — same as before. The actual decisions are made by whoever writes the code. So now the LLMs are making those decisions and imposing their values into our systems.

It’s actually more than that. I also map cultural systems, and one of the interesting things about a cultural map is the concept of values. Values are quite tightly connected to memory — our collective memory of a space — which we talk about in terms of symbols, rituals, and heroes, what we call art. The values our culture expresses are quite tightly tied to art. You control art through tools, medium, and language.

For example, if I control the printing press (the tool), paper (the medium), and the written word (the language), I control what books can be written and how. In the UK, the story of Robin Hood — well, what if it were the wonderful book about the brave Sheriff of Nottingham trying to stop that marauding horde led by that crook Robin, who’s stealing from others? That’s not how we tell the story, but if I control tools, medium, and language, I can change the way the story is told.

The interesting thing about GPTs and ChatGPT-style large language models is that they’re becoming the tools by which we code; the language is changing from deterministic to non-deterministic prompts; and the medium is changing — it’s not just code as symbolic instructions in text, we’re now using images to build things. All of this happens in a large language model world controlled by a few providers. So these systems have the potential to control how you reason about a space. And they’re making decisions and embedding values based on what they’re trained on. None of this is transparent.

Does it matter? Consider the trolley problem. A self-driving car comes along — kill one person or five? Hopefully kill the one, if that’s your only choice. What if the one person is a member of the Gold Club — a digital subscription. Do you kill the one or the five? Hopefully still the one — it doesn’t matter if they’re a Gold member. But what if the LLM, the GPT, is encoding the system and making that decision? Whose values are taking effect? Not yours anymore. That’s the point.

That’s why I describe them as a non-kinetic form of warfare. We’ve previously used art — Hollywood, video games. Hezbollah produces a triple-A video game, Holy Defense. It’s not because they’re big gamers — nations and organizations use games to influence values and behavior. We’ve used art, statues —

Danny: The way we’ve done with movies for —

Simon: Movies.

Danny: — a very long time. Gaming seems weirdly underappreciated given how many people spend how much time on it.

Simon: And how much money it makes.

Danny: Right. But the principle is the same. On LLMs, one response would be to say, well, we want to mandate open source — properly open source, not just open weights — and we can have a long discussion about how to get there. But the strand I want to explore: there are two possible worlds here. One in which open-source models are competitive, at least for the most demanding tasks. And one in which they simply aren’t.

The jury, for the moment, very much seems to be out. It is possible that we’ve figured out a way of building this artifact that consumes such vast amounts of capital — energy, resources — that it may be quite hard to match the very best models. If we find ourselves in that world, what do we do, given the diagnosis you’ve just made?

Simon: China has interestingly mandated that all its systems should be trained on socialist values, which is an interesting way to open a door — and they indicated several years ago they were going to be much more open and transparent in terms of the training world.

You’re touching on the question of sovereignty. How do we maintain sovereignty? Normally, when we talk about sovereignty, we take a map of a space, draw where our border is, where we need to protect, where we conflict, where we collaborate. Most of us don’t have maps of the technological-economic space, so we don’t do that. It’s a bit like generals not having a map of their landscape and talking about sovereignty — you end up with conversations like, “We’ve got to protect the trees.” What trees? Which trees, where? You get the same in the institutional sovereignty space. “It’s about data.” Well, what data? “It’s about the location of stuff.” Surely it’s about the training of the models, not where the model is hosted. And about whether it’s open or not — and that’s open in terms of weights, not training data, so you can’t even investigate. Most of the conversation on digital sovereignty has been pretty poor for the last eleven years.

So what do you do? Given there may be an advantage with huge models — you don’t know that, given the technology that may come around the corner, given you’ve got multiple players — your best bet is to hedge. Rather than getting trapped in one, spread across two or three. Better to use the stuff coming out of China and the stuff coming out of the US than be tied into one.

Then there’s the idea we should build our own. That’s very much like the early days of cloud. A friend of mine, Rick Clark, worked on OpenStack — they were going to create an open-source equivalent to Amazon EC2. They took a lot from Eucalyptus in terms of fundamental ideas, but they completely messed up on day one by differentiating on APIs, which made no sense. We’ve had people say we should build our own cloud. It’s not impossible. Look at China. Fifty or sixty years ago, China was primarily luggage and footwear. With concentrated effort — the work of Deng Xiaoping — moving up the supply chain, look at them now. China is in an amazing position.

So it’s never too late. People are very dismissive of Europe’s capability. With the right focus, even the UK’s capability — with the right people and enough time — we shouldn’t be dismissive. China proves the example. But right now, if you’re looking to maintain sovereignty, I’d be encouraging the adoption of multiple systems, not getting tied into one.

Does that make sense?

Danny: Totally.

Simon: We’re in a tricky position.

Danny: Exactly. There are probably not many fantastic options out there, and recognizing that seems important — without dismissing the longer-term possibilities. But you do want to acknowledge the situation you’re in rather than pretend a comparatively small investment in Model X or Y, or European provider X or Y, is magically going to fix the mess overnight.

What accounts for the Chinese government’s seemingly dramatically higher level of situational awareness? Have you taught them?

Simon: When I was over in China I was investigating this, because it became quite clear they were very good at identifying how systems are industrializing and targeting that in their plans. Very skillfully done — very impressed. Unfortunately, during my trip, a message came down to all the people I was meeting that there shouldn’t be conversations with Western researchers; basically all my conversations got shut down. So I never got to the bottom of it.

Friends of mine teach mapping at Peking University, bizarrely. I’m not saying they use it — they’re certainly not. But whatever it is, within the system there are clearly high degrees of situational awareness across the technological and economic spaces. I’m really impressed by the way China plays the game. We should learn from China. But the problem is, I don’t know how they do it.

Danny: You’ve said China will “win the AI battle.” Why?

Simon: I think we’re getting very tied to a specific architectural model in the West. When I was looking at China in terms of research, and the constraints — the chip restrictions imposed on them, which encouraged them to develop their own industries and capabilities. That’s one of the reasons I argued against trade restrictions: invariably you encourage innovation within the constrained system. It’s much better to bolster your own innovation.

It was a combination of factors — the research, the engineering talent, some of the constraints. I came to the position that, over enough time, as I’ve seen with other industries, China would be in a stronger position.

Danny: What do you like about Werewolf, the game?

Simon: Another weakness of mine. Everybody knows Werewolf now because of The Traitors, the TV programme, which is Werewolf, which is based on Mafia. What do I most enjoy? I like being the dungeon master, the host. It appeals to my theatrical side, and I love watching the interactions of people working through the problem. It’s a fascinating game. You have to be careful that it doesn’t end in fights and arguments. Generally I find it quite pleasant, peaceful, and amusing when played in the right way. If you’ve ever watched The Traitors, go play the game. Well — not yet, because I need to buy another set of cards. Once I’ve got mine, please go buy them.

Danny: It’s a great game.

Simon: It is.

Danny: I think you produce your own energy — solar in your house, I believe. You may have a vegetable patch. Are you a prepper?

Simon: My total energy bill, including running my car and everything else — I’m all electric — is that they pay me money every year, because I produce that much. I have a vegetable patch; I like to grow some food. Not enough. I might have some preparation stuff set aside. So yes, I like to have contingency. You could call me a prepper, but not “underground shelter and two years of supplies.” Just caution.

Danny: And I don’t mean this pejoratively — there’s an obvious connection between your work and that fact.

Simon: Within the UK government, where I used to do work — never as part of a department — I would talk about the issue of supply chains. We’ve gone through Brexit, COVID, Ukraine, and now Iran. In all of those cases, understanding the supply chain and where the connections are matters, and often we haven’t had it.

As soon as I heard Israel and the US had gone into Iran — literally an hour after it happened — the first thing I said was Strait of Hormuz. I couldn’t believe there wasn’t a military plan; this must have been gamed out. The next thing I said: do we have a good supply of urea? Urea is used in AdBlue. AdBlue is used in logistics trucks — many of these trucks won’t start unless their reservoirs are full. There are long chains of components and dependencies. Helium — I think 30% of the world’s supply goes through the Strait of Hormuz. That knocks on chip manufacturing, medical devices, and things like quantum experiments because of the cooling side. Fortunately quantum computing isn’t that big at the moment; if it were, it would be an even bigger problem. Whether it’s urea, helium, or any other material, there are all sorts of dependencies.

One of the most interesting pieces of work I saw was from Hungary — VAT data. They had transaction-level reporting of VAT records, so they could graph the entire economy and work out who’s selling to whom, and from that work out the bottlenecks. Huge amounts of GDP were dependent on a very small number of companies. That alone should set alarm bells ringing. We lack that sort of information — we don’t have a department of the supply chain. There’s resistance to some of these ideas; we seem to have a reliance on “the market knows best.”

Danny: Speaking of public and non-public things, we talked about gameplay earlier — context-specific moves in service of a strategic objective. You have a very long list of gameplays. I don’t know whether it’s 60 or 100 —

Simon: Several hundred. Well over 100.

Danny: Several hundred. I think some gameplays you’ve never published.

Simon: Because I have to have a job too.

Danny: Fair enough.

Simon: It’s as simple as that. I made everything Creative Commons — pretty much everything. The map itself, the process of mapping, the climatic patterns — things like “everything will evolve,” inertia to change, coevolution of practice — all open. The principles — all open. Where you’re into gameplay, you’re thinking about the map and how to manipulate it in your favor — what I talk about in terms of strategy. I think I made about 60 open, but not all of them, because I may want to play the game too. I don’t want to reveal everything. People should create their own — I want to give space for others to discover patterns themselves.

The way I learn patterns is pretty simple. I use the mapping when we’re building something — a premortem challenge. We map it, question what we’re trying to do. Once we’ve built it, we do a postmortem and learn — take the map and see what happened, and often pick up patterns from that. I make a chunk of them available, but not all.

Danny: You’ve shared a lot, very generously, in public over the years. Thank you for that.

Simon: Pleasure.

Danny: For anyone who hasn’t read your book — warmly recommended. It is highly rated, but in my view still underrated. There’s a huge amount of alpha in there, as some people would put it. With that, thank you so much, Simon. This was huge fun.

Simon: It was fantastic. Some of the most interesting questions as well — they brought up things I thought had been forgotten. Always fun.