Making sense of chaos with Doyne Farmer

One of the real privileges of doing a podcast series like this is you get to interview some of the real legends in the field of complexity science. We've had W. Brian Arthur on the show before where he talked about the early days of the Santa Fe Institute and the development of complexity economics. Well, today we're joined by another legend, which is Dauin Farmer. Dauin Farmer is the director of complexity economics at the Institute for new economic thinking at the Oxford Martin School. He's the bail gifred professor of complex system science at the University of Oxford, and he's an external professor at the Santa Fe Institute. Dauin's career is massive and it goes way back to even before complexity science became a discipline because he played a key role in the formation of chaos theory. A theory that's very related to complexity science but also quite different. Dauin's going to talk about the early days of chaos theory. He's going to talk about how that got him into complexity science and how eventually that brought him into complexity economics. And he talks a lot about complexity economics in his new book, which is called Making Sense of Chaos. Particularly he's going to talk about the importance of multi-disciplinary research. And he's going to talk about what complexity economics tells us about how we can face the climate change challenge that's before us. But before he gets into all of that, he's going to talk about the time he spent at the Roulette wheels in Vegas. This is simplifying complexity, a podcast where we explore the underlying principles of complex systems. Systems that seem to defy our rational view of the world, like economies, ecologies, or even you or me. I'm forensic engineer, Sean Brady, and I'll be your host. Dauin Farmer, welcome on the show. Very happy to be here. So you began your complexity economics career if I can call it that way way back with sort of the beginning of chaos. Can you go back and briefly talk us through that? I mean, sure. I got here through a rather circuitous route. I started out studying cosmology and physics. I was imprompting from my childhood friend, Norman Packard. We decided to try and beat the game of roulette using physics to predict where the ball would go. So we built what turned out to be the first wearable computer was even a conceivable computer that we used to time the motion of the ball. And we used to use physics on our careful studies of the dynamics of rolling ball on a circular track. We predicted where the ball would come off the track and successfully beat the game of roulette. Though we never managed to up the stakes to a level where we got rich due to fear of broken kneecaps and inadequate bank role and a lot of hardware failures. So, I think our first trip into the casino was probably in 78. And we, I think, went to the casino in 78, 79. I'm not quite sure of the chronology anymore. It was a long time ago. And you used to time the speed of the ball. And you discovered that when it actually comes out of the track and it starts bouncing, that doesn't have a massive impact on the part of the roulette wheel where it ends up. And it does have an impact. It just wasn't so massive as to completely obliterate our edge. In effect, there were about eight numbers that were more or less excluded. We're very unlikely to come up. And so if we bet on numbers on the other side of the track from those, we did reasonably well. And the computer you used is press what you told and that would time the quickly the ball was going around. And which is built into our shoes that we would depress when the ball went past. Actually, for the ball, we just used the right shoe. But we could effectively make a very simple operating system because the whole thing had to run on three kilobytes of program memory and 128 bytes of read, right memory. So had a very compact floating point package, simple operating system that took inputs from switches and our shoes. And as outputs had buzzing solenoids that would signal what number are also convey information about what was going on because we had to set parameters and do various tasks in order to get ready to play. And this was possible. I didn't quite appreciate it until I read it in your book that it was possible because the ball starts moving or the ball is thrown into the track where it starts spinning before the bets are closed. So you do get this, what, this 10 second window where you're able to collect your data and make decisions accordingly. That's right. The key is they don't close the bets until a couple of seconds before the ball leaves the track. And then that brought you to Santa Cruz. Well, I was already at Santa Cruz when I was doing that. I went on living in Santa Cruz. I just made we made 11 trips to Nevada during that period. And you started in what was then a brand new display that point in term called chaos or became to be called chaos. Yeah, I think doing roulette primed me for thinking about that, roulette's an example of sensitive dependence to initial conditions. It's laws of physics, but the relationship between what happens and where the balls end up is sufficiently complicated that it's used as a random number generator. So that economy had me already thinking about this question. Then when I heard about chaos, I got very interested and decided I wanted to do my research on that. And we formed a collective together with Norman Packard, Rob Shaw and Jim Crutchfield, we called ourselves a dynamical systems collective and began doing research with the focus on trying to understand whether we could develop methods for identifying. And then we started identifying chaos in real experiments. Can you just briefly explain for the list of what's chaos and how did that relate to the more deterministic approach to mathematics or modeling at the time. So chaos is what fascinating is both deterministic can be fully deterministic, but also generates random behavior. And nonlinear dynamical systems whose geometric structure is in some sense complicated enough to make that happen. And so it happens when nearby initial conditions become separate exponentially through time. So that two points that are initially so close together that they are indistinguishable from a practical point of view under finite resolution. Become randomly related at some point in the future. And it happens in some sense because the underlying dynamics acts like needing bread dough that nearby points get stretched far apart and then fold it over on top of each other. And then that repeats again and again so that you end up with behavior that's what's called mixing it effectively takes any little ball of initial conditions and eventually mixes them throughout the whole dynamical system so that they're practically indistinguishable. So that you can't tell in advance exactly where they're going to go. You famously have the weather works. Is that probably one of the best examples that's often cited for chaos. The weather is one of the best examples and it illustrates all the basic points. I mean you might have wondered given that the forcing of the sun energy that the sun in parts to the earth is imparted in a very regular way. Does more or less the same thing every day changes gradually as the seasons change. The weather on the other hand is quite a regular and responds on time scales that are both shorter than and longer than those changes and can be extremely intermittent. The calm pertains at a time and then you can ever hurricane and all that is caused by the fact that the weather and the underlying Navier Stokes equations that determine what the weather does are highly nonlinear and very chaotic. And can't was really started by Lorenzo and that he wasn't meter artist. That's right. Well, Poincarei already had the idea in the end of the 19th century and there was a little thread through there that occasionally surfaced itself and a few papers that didn't receive much attention at the time. Lorenzo himself published a paper on it in 1963 that attracted very little attention until it was discovered by some mathematicians who had been thinking about this question independently in the early 70s. And then as the implications caught on it became widely known. And when you were at Santa Cruz and you got exposed to case, you went on to your PhD on the subject. chaos. I looked at situations where you can have precise timing that's preserved even when the system is otherwise chaotic. So it can preserve periodicity in the middle of chaos and well looked at other things like that. And then when you left, uh, Santa Cruz, what happened next? Well, I went to Los Alamos where I became an Oppenheimer Fellow, then I joined the theoretical division and I went on to start the complex systems group at Los Alamos in 1988. And what for you? I mean, people talk a lot about the other relationship between complexity science and chaos. What is the the key sort of relationship for you and the. Well, chaos is one example of what's called an emergent phenomenon. Where something happens at the high level that's not at all obvious at a lower level, where the high level phenomena has qualitatively different properties than the low level building blocks. So in the case of chaos, it could be a very simple innocent looking equation that because of its non-linearities, gives rise to chaotic behavior that you could not have predicted from just looking at simple equation itself. So it's an example of complex systems, which are systems that have emergent phenomena. Now, chaos means that you have a deterministic dynamical system that actually destroys information. So the information on the initial conditions gets lost through the action of the chaotic system. Even more interesting in complex systems are adaptive complex systems that create order and structure and the ability to process information in a way that is often not at all obvious. And starting in the late 80s, I became more interested in those kind of systems. What was your your early steps into the. Well, I began by organizing a conference on cellular automata, another conference on evolution games and learning, and began working on models for the origin of life, and for the system level information processing in the immune system. And then spent a lot of time thinking about machine learning and questions in cognitive science. Now, would you say now, darling, that you specialize very much in complex economics? Is that your main area? And if it is, how do you get to it? Well, yes, that is my main area. I got there because when it was at Los Alamos, I was still doing a lot of work on predicting chaos because, although chaos on one hand says long term predictions impossible, there are circumstances where you can actually make better short term predictions if you realize that the random behavior is coming from low-dimensional chaos rather than just intrinsic redness and complication. So I was working on that and making predictions for things like ice ages and sunspots. And whenever I would talk about the work, somebody in the audience would always say, have you thought about applying this to the stock market? And I got tired of saying no and decided to quit my job at Los Alamos and start a company that became one of the first quantitative trading firms. And that got me into thinking about finance and economics and reading the papers in that field. I came to feel that the ideas that were being used by the academy to understand what happens in financial markets, well, they were missing key ideas. And so when I left prediction company after eight years, because we kind of reached a level of success that made me happy, I decided to devote myself to applying complex systems ideas to finance and then later economics. You're based at Oxford now. And what are the key areas you focus on? Well, nowadays I'm really thinking a lot about the economy more generally, trying to develop ways of modeling the economy that are very different than in mainstream economics. And also thinking a lot about technological change and the climate transition, how we can guide the economy to make a transition away from fossil fuels as quickly and smoothly as possible. And what are the key sort of, from a perspective of your neoclassical economics versus complexity economics? What are the key learnings ideas that neoclassica economics is missing in that transition? Well, neoclassical traditional mainstream economics begins by assuming that people have utility functions and derives models where each agent, each decision maker in the economy could be a household, could be a firm, even a government, each decision maker makes decisions that maximize that agent's utility. So if you will, the utility is a kind of scorecard for what each agent likes or doesn't like. And the agent reasons about the world in some way and understands that they're not alone in the world. There are other agents who are also reasoning about the world. But the assumption at the end of the day is that all the agents make the decisions that will maximize their utility under whatever framework they've been put in. So that's the basis for essentially all of mainstream economic theory. Now, complexity economics takes a very different approach. It assumes that the agents are making decisions based on some decision rules, some rules of thumb and heuristics, maybe even simple learning algorithms that allow them to make their decisions. They may keep a kind of scorecard that influences what those decisions are, but the decisions are never perfect. And the system is dynamic in that the agents process the information they have from the economy. They make decisions, they take actions, those are integrated into what the economy does, and then the agents look at the economy again, the information's now changed, they make another round of decisions and the process repeats itself. So it's a very different framework than that that is used in mainstream economics. And when you apply that to climate change and the transition, what do your big take away? So we've done several different things relating to climate change and the transition. The most tangible ones now actually are based on relatively simple analysis of historical data. That is, we, Francois, La Fonde and myself looked at historical data on performance in particular cost of technologies through time. And where you can see that through time, many technologies improve, they do so very differently. Everybody's familiar with Moore's Law, computers tend to improve it about 40% per year. They get cheaper, more energy efficient and get faster all at once. Solar, photovoltaic cells improve at a rate of around 10% per year. Similarly, lithium ion batteries have done that more or less, hydrogen electrolyzers are doing that. But then other technologies like fossil fuels, actually we're paying, once you adjust for inflation about the same for oil as we did 140 years ago. And similarly for natural gas and actually similarly for anything we mine out of the ground. So some things get cheaper, some things don't. Once they set in motion or trajectory for improvement, they tend to hold that for a long time. And so we developed some good methods for making predictions and for knowing how good those predictions are. So and we tested them on 50 different technologies, making 6000 forecasts where we pretended to be in the past and predicted up to the present, not knowing what the present was going to be. And so tested these. And then we put together some scenarios for the energy transition, taking these into account and looking at what we think are the key technologies, which are solar cells, wind turbines, lithium ion batteries and hydrogen electrolyzers. And mapped out a scenario for the future. And then based on our forecast for the cost of the individual technologies made a forecast for what the transition was going to cost. And showed that it's likely that the transition is actually cheaper if we do it quickly. And that we can display fossil fuels and make energy cheaper than it's ever been. So it's a transition we want to make even if we weren't worried about climate change. faster better? Faster is better because as we build out these technologies and deploy them, we learn. We improve. We learn how to manufacture solar cell modules more cheaply. We actually learn how to install them more cheaply through time. And in fact, those two processes have proceeded at about the same pace. So we gain experience. And if we deploy them faster up to some point, we gain experience faster. Now, there's some point at which we do it so fast, we can't keep pace anymore. But we're not close to that. It's a sad, we must be quite a distance away from that at the moment. Well, actually, solar photovoltaics have come down in price by a factor of about 10,000 since they were first deployed in the Vanguard satellite in 1958. So we have made enormous progress in getting better at manufacturing and deploying solar cells since then. And why are fossil fuels different? And it's purely because they are what they are and you can only get them out of the ground roughly the same sort of race or the same sort of scale is, or what's driving that? No, well, the first answer would be, I don't know, because it's a mystery. On one hand, oil drilling equipment is way, way more sophisticated than it was a century ago. On the other hand, oil becomes harder and harder to extract because we extract all the easy deposits. But it would be very, very surprising. I mean, as it happens, those two things have just about canceled each other out. But why it should be that they've more or less exactly canceled each other out and that that's happened for everything else we dig out of the ground. That's a little hard to see why that would, what is the mechanism? But there is some mechanism that for whatever reason is causing those two things to roughly cancel out consistently over a century for not just oil and natural gas, but coal, but about a hundred other minerals. What is it about the neoclassica that pushed it doesn't take this into account or doesn't take into account the savings on the transition? To be honest, there is nothing that says they couldn't have done that. My new book is mostly about economic theory. This is really about data analysis because we don't have a good theory for why this is happening. But a mainstream economist easily could have done what we did. It just, they didn't. Now, let me mention we're also doing some other things in progress that we think are going to be very useful. And that is we're building an agent-based model or agent-based models of the power and energy sectors of the economy. So models for electricity, oil, natural gas. And in those models, the agents are companies like ExxonMobile or Pacific Gas Electric. And every year, those companies make decisions about how much they have to invest about whether that investment is going to be into a new oil field or a solar farm. And we track the process of deployment. We simulate the markets therein to make an estimate of what price they could have sold things at, how much they would be able to sell. Therefore, what their profits and losses are going to be. And then repeat the process to figure out how much money they have to invest. So that's an agent-based model. It's very different than the kind of models economists have made in this arena. The economists have focused on models called integrated assessment models or family of models that basically try to answer the question, given a certain target for global warming, what is the most efficient way to get to that target? So it's a different question. We even said ask, what's likely to happen and how can we nudge it towards making it better? We like our approach because it's more testable and we can ground it more directly in data so we can get more better answers at the end of the day. And just those two models, the more classic economics model where they say, you know, this is the target. Do they have to make a range of assumptions on how they believe companies will need to behave and interact to produce that? And is that the fundamental difference between your model where you're essentially defining the local interactions between well-offing investors or invest in this and we do this and this and then you roll all that together in the model to pick up the emergent behavior? What's the fundamental difference between the two of those? The fundamental difference is, first of all, the mainstream models are not none of them are cast at the level of individual companies. They make broader assumptions that oil can be produced at such and such a price or that solar energy can be produced at such and such a price and they integrated all together to try and show example of what the energy system could be in the future. They typically involve a lot of optimization assumptions because they try and find the best possible path. We don't try to find the optimal path. We just try and find the likely path and then make it better. And because we're not optimizing, our models can put in a lot more institutional structure. We can model the economy much more realistically than they can. And that we think gives us an advantage. What does that particular model tell you of what are the big things coming out of it? I mean, we're working on it. So we can't say yet. Let me just say my hypothesis is that most companies have not realized how quickly the energy transition is going to happen and that if they don't get on board with making it happen, they're going to go bankrupt. And that's purely down to the speed at which the transition can happen, which will make their technologies obsolete. Is that fair? Yeah. Again, this is hypothesis right now remains to be shown in our model. But people who make decision most companies have yet realized that all the data points at a very fast transition because technologies like solar energy and wind are growing at rapid exponential rates. So it's not just that their costs are going down. They're increasing globally at rates of 20 to 30% per year. And so we're doing some other work that we're wrapping up extrapolating that and our forecasts are that things are going to happen fast. Exponential change can be misleading because it's small, it's small, it's small, and then it suddenly gets big. And we're about to see those technologies suddenly getting big. The next five years, the increase is going to be very large. And we'll see fossil fuels really start to decline in a meaningful way as a result. And is that the fundamental problem we have as humans? We just cannot, we cannot imagine that that exponential growth. We just see it as this linear thing that's likely to continue. And then we get dramatically car chart. Well, it's one of our problems. We have a lot of other problems. But yeah, exponential growth can be hard to wrap ones at a round. Going back to the work you did on the changes in technology and the predictions of the changes in technology. Can you give you a few examples of what technology you looked at where you went back and tried to project into the future to see if it got right? And one of the most surprising ones of it. We got there to the database of technologies. Basically anything we could get our hands on where we had data about what it cost to do something at some given point in time. And how much was being produced through time. So for example, my postdoc at the time, Baylor Nage, wrote Gordon Moore and Gordon Moore, a centrist, the data he had personally collected on the cost of transistors and hard disks and DRAMs and other electronic components. You know, we had data on what it cost to produce beer on ethanol and lots of other chemical processes. Solar, wind, we just gathered everything we could get our hands on. And what we saw is that technologies are very heterogeneous. As I said, things you take out of the ground just don't improve through time. Electronics related equipment improves it 40% per year. So there's a big gulf between those. Transistors now are a factor of a billion cheaper than they were when Gordon Moore made his forecast in 1965. The number of transistors in the world is now comparable to the number of synapses in human brains in the world. Not neurons, they've surpassed a number of human neurons, but they've pretty much reached parity with human synapses. So your new book, Dying, talk to us about your new book. Yeah, my new book's called Making Sense of Chaos, Better Economics for Better World. In my book, I try to show the world what complexity economics is. Talk about some of the examples we have where we've been able to do better than mainstream economics at making predictions or understanding things. I also kind of take a step back and narrate what it's like to be a complex system scientist who has moved from one field to another, the excitement of scientific discovery and looking at how these fields have developed and succeeded over time. It's kind of a scientific adventure. I'm the subject of, which I think is fascinating. We've talked for a lot of guests about this, but what is it? like to be a complex system scientist? What is that for you? Well, on one hand, very exciting. I've had the good luck to live through several scientific revolutions in my time, chaos, complex systems, and machine learning being the three big ones. It's been very exciting. At times, it's been really challenging because when you're doing something that's very different from what everybody else does, it can be hard to have it accepted. You can sometimes feel like you're fighting an uphill battle. But it's been enormously satisfying to see these ideas take hold and become widely known and be adopted by many people. The big battle that I'm engaged in now is trying to get complexity economics to take hold. It's strongly resisted by the mainstream. It shouldn't be too surprising. I was a bit naive, to be honest, but I had quite anticipated the amount of pushback I was going to get. And this, in some sense, is the hardest of the revolutions that I've been involved in. I think it's so because mainstream economists realize that if complexity economics succeeds, it's doing things completely differently than the way they're used to doing them. It is competing with them very directly. It's likely we'll see that it's complimentary in the sense that it can do different things well than their models do, but I think they're afraid of it because they recognize that it's competes with what they're doing and might make some of their expertise obsolete. It's incredible. We had Brian Arthur on the show and he was talking about the same challenges for 30 years ago. The same challenges existed in terms of trying to engage with mainstream economics. I've heard you speak on another podcast about a subject that I think is fascinating in complexity, where you talk about the way science is, well, this is my description, but science is all about learning more and more about tiny areas. And you've said that the stuff you've been interested in is follows between the gaps or the chasms between the traditional disciplines. Can you talk about the role you see and the importance of the role you see for complexity science in dealing with that problem? Through time, science and the pursuit of knowledge has become more and more siloed. Back in Newton's day, where did he publish the philosophical transactions of the Royal Society? So philosophy included everything back then. Through time, disciplines have evolved and they've evolved to become narrow and narrower. So the task of the typical academic to get tenure is to become on an expert on something and no more about something than anybody else in the world. And the easy way to do that is to narrow the scope down of what you know to something that's very focused. A problem with that is then there's a lot of knowledge that a lot of things we really have to put information together from different areas. There's been a widespread recognition that we need more interdisciplinary work, but it's hard to do because the normal pressures in science are pushing everybody in the opposite direction. Now, I've had the luxury through my career to be involved in interdisciplinary science almost across the board and have a lot of fun learning about lots of different fields, learning from my colleagues in those other fields and working with them and really making progress by putting knowledge together from very different areas and synthesizing rather than reducing. It's been tremendous fun. I've managed to navigate my way through my career doing that and it's been a real pleasure. Why do you think you were drawn to that? Well, I've been curious since I was a tiny boy. I was always the kid asking why, you know, what I learned how to read. I spent several years just basically in the library just reading voraciously because I wanted to understand everything. And I'm still in many ways that little kid trying to understand as much as I can understand about the world. And I just can't resist the temptation to go find out about new things. I like to stay on this deep part of the learning curve. And what do you reckon is the biggest success story of complexity science, shall we say? The one you tell me about the difficulty of getting the economics communities to listen. But what do you think? You're through good luck or good time or whatever complexity science has got something really clean across the line that has resonated and stuck. Well, there's lots of small things. Some bigger things include, say, one of the main methods in complexity economics is agent-based modeling, simulating systems where people make decisions. So we've seen that in epidemiology go from being something fringe to being the center of the field, similarly in modeling traffic and modeling inventory management. Many other areas, it's now the tool that's widely used. It has commercial success. It's just out there embedded in the world. It's still not part of economics, but the company that I've formed macrocosm, we're devoted towards doing that and hope to really show that in a practical way, we can do answer practical questions and economics better using agent-based modeling than other methods. And we've already had some success. We predicted the economic impacts of the COVID pandemic on the UK in our second quarter of 2020. We forecast that the hit to the UK GDP was in annualized terms going to be 21.5%. You know, with a dust settled, it was 22.1%. And we didn't just get that number right. We got all the other numbers right. We got them right economic sector by economic sector. We got the time trajectory of the recovery right. And we did it by using new methods that don't involve the assumption of equilibrium that central and economics. With a simple agent-based model that just every day looked at each sector and said, does this sector have, is there demand for what this sector does? Is there enough labor to produce the good this sector makes? And does this sector have the inputs it needs to make what it makes? And just by doing that, iterating it day by day, looking at tracking how inventories were, we were able to make really good predictions. So we think we have proofs of principle now that do very well, but we're now trying to scale that up and show that we can really do it day and day out on a practical basis. I mean, back to your broader question, machine learning is, at least in part, has grown out of the complex systems community. And it's transition from being something that was fringe and that was not substantially better than other things to being totally mainstream now. And if you were to finish, if you were to dream of what next in complex economics or whatever else for yourself, what does that look like doing? I think we can show that we can just do a much better job of predicting what the economy is going to do next, conditional on the policies that we adopt. So I think our goal is to really model the global economy at a fine grain level. My dream is to do for the economy what Google did for traffic modeling, to have a kind of Google economics where we can really make micro and macro predictions at the same time. We can aggregate up those micro predictions to get macro predictions and where we can answer a broad set of questions to provide better guidance so that we can see the cause of fact. If we implement this policy, this is the likely outcome. The world will be more like this instead of like that. And to really be able to do this day and day out on large scale. I thank you very much for being on the show. Thank you. I'm Sean Brady and I'll see you in our next episode.