Mental Models for complexity | Scott Page and Shane Parrish | The Knowledge Project #55
What this covers
Shane Parrish speaks with Scott Page the Professor of Complex Systems at the University of Michigan. They discuss Scott’s book the Model Thinker and the power of mental models to improve your understanding of the world and solve complex problems.
For more information see: https://fs.blog/knowledge-project/scott-page/
#TheKnowledgeProject #ScottPage #ShaneParrish #MentalModels
Source description (no synthesized summary yet).
Scott Page argues that understanding complex reality requires multiple mental models applied together, not single frameworks, because individual minds are insufficient for making sense of high-dimensional problems—wisdom comes from assembling ensembles of models and diverse perspectives.
- Single models can only explain partial variance in complex systems; ensembles of different models collectively capture reality better
- Cognitive diversity—people with different mental models—outperforms individuals because they see different parts of problems
- Complex modern decisions (drug approval, policy, business strategy) require perspective-taking across disciplines and paradigms
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Individual human brains are not capable of solving major global problems like the obesity epidemic, climate change, or world peace because the complexity and dimensionality of these problems exceeds what any single person can hold, but collections of people with diverse mental models can make meaningful progress by aggregating their different ways of understanding the world.
“you yourself are not gonna sort of solve the obesity epidemic you yourself are not gonna sort of create world peace you used to help or not getting sort of you know solve climate issues right your brain just is gonna be big enough but collections of people we're creating it larger and some of them I'll actually have a hope with addressing”
Systems can contain logical inconsistencies at the aggregate level even if the components follow consistent rules—for example, if everyone follows a success formula, that formula becomes self-falsifying because it no longer provides an advantage once widely adopted, which is a fundamental aggregation problem in complex systems.
“there's a circular reasoning in there there in the sense of everybody followed that formula it's not clear that everybody would be successful right and so the thing is oftentimes these systems can contain feedbacks within them right that make them logically inconsistent at the level of the whole”
Markov models are particularly valuable for forcing clarity about assumptions because they assume that if transition probabilities are fixed and every state can reach every other state, the system will converge to a unique equilibrium—which means if someone claims their system is path-dependent or complex, they must be either creating new states or changing transition probabilities, forcing them to be explicit about their mechanisms.
“what that model forces you to do then is if you want to argue the world is complex if you want to argue for path dependence if you want to argue that a policy interventions gonna make make a difference in some line you then have to you have to either be saying I'm creating a new state it didn't exist before or I'm fundamentally changing these transition probabilities”
People can improve their predictions and problem-solving on new but similar tasks by reading written descriptions of mental models used by successful problem-solvers on old tasks (tacit knowledge can sometimes be codified and transmitted via language), though some skills (like tennis) require direct practice and cannot be transmitted through descriptions alone.
“I have another set of people read the mental models from the first set of people and then play not the same games but different games that have the same difficulty and compute how long it takes them and what happens is they're just a lot better and what you what you see is that this is a this is something where it's it's not tacit knowledge is actually learn about knowledge”
Nature cannot be 'carved at its joints' into independent domains; rather, the world is characterized by complex overlaps and feedbacks that require multiple overlapping (not parallel, independent) models to understand.
“the simple models we teach people because we thought like Plato Plato's famous quote about carving nature at its joints right I think there was a belief that we could carve nature at its joints and then for each one of those little pieces mmm you sort of apply this model and right here some people will sometimes say oh the many model thinker it's like the parts of the elephant and I'm like no no it's almost exactly wrong in the sense that you want each model you know there is a sense in which yeah different models look at different parts but you you need that overlap right because you can't carve nature at its joints that's what we've learned over the last 50 hundred years right is that it's complex the world is a complex place”
Mental models are frameworks that people use to make sense of the world by mapping reality onto mathematical or logical structures; the key challenge is that reality is messy and complex while mathematical models are clean and logical, so the work of mental modeling is connecting those two domains.
“what a mental model is is just a framework that you use to make sense of the world”
Derek Saville's intuitive response to the 2008 Iceland financial collapse—that Iceland is smaller than Fresno and therefore can't matter to the global financial system—was wrong because it failed to use the appropriate model (network of interconnected financial obligations) and instead used a simple supply-and-demand model, illustrating how model selection matters enormously.
“someone comes into his office and says Iceland just collapse in two models sort of model sort of human said one is you can think of the international financial system as a network of you know loans and deposits across banks and across countries another model is just a simple supply and demand model and so Munger has this wonderful quote about you want to sort of array your experiences on a latticework of models and when Eric doesn't the situation those are his two most complicated you know network of loans and promises to pay and simple supply and demand and he looked at the person who worked in his office Iceland is smaller than Fresno go back to work”
Repeated games with high stakes move decision-making toward rational models, while infrequent or low-stakes decisions suggest rule-of-thumb or heuristic models; organizational procedures are typically even less rational because they often encode historical decisions rather than current optimization.
“Colin camera you know richard thaler people study behavioral economics would say if it's repeated a lot that should move you a little bit more towards irrational behavior because people should learn and if the stakes are huge that should move you towards rational behavior”
Power law distributions (where most events are very small but a few are catastrophically large) arise from three distinct mechanisms: preferential attachment (positive feedback where success breeds success), random walks (where firms or species can fluctuate randomly until failure or large growth), and self-organized criticality (where systems aggregate to a critical state where small perturbations cause large cascades).
“I talk about three models in the book one is something called the preferential attachment model where imagine things kind of like what you imagined it like there's a set of cities or there's a set of books and the probability I moved to a city where the probably a by book is proportional to the number of other people living in that city or by Manette book we can see right away there's positive feedbacks”
A liberal arts education builds perspective-taking ability by exposing students to different views of the same historical events, enabling them to see the world through multiple eyes and understand how different groups experience the same situation differently.
“the reason you want to read literature from a whole bunch of different vantage points like so the reason you don't want to just reads are the great man view of you know Canadian history or US economic history or something like that is because there's all these other people to experience that same thing and saw it from a very different perspective”
Linear thinking (projecting trends as if they will continue indefinitely) is dangerous because many systems exhibit concavity (diminishing returns as you scale), such as industrialization growth, population growth, or return on investment, which means projections based on linear extrapolation will be wildly optimistic.
“in the 1970s Japan had this really fast growth there all these articles saying Japan's gonna overtake the United States in eight years but the thing is if you construct the model you realize that as you sort of industrialized you know pretty fast that there's gives me diminishing returns to that industrialization it seems to have China right so if you do a linear projection of China you know five years ago you either said oh my gosh you know by 2040 China's economy is gonna be used enormous but the reality is growths gonna fall off because what the model shows in order to maintain anything even close to linear growth you have to innovate like crazy”
Business decisions have become vastly more complex than in the past—they now must account for environmental impact, talent attraction, strategic positioning, innovation capacity, and brand effects, not just cost-benefit analysis, requiring multiple disciplinary perspectives.
“when you make a business decision there's a recognition that there's environmental impact there's an understanding that's it's gonna affect your ability to attract talent right because it's gonna be an interesting problem there's a question of how does that position you strategically for the long run there's a question of what it does for your capacity there's a question of what it does for your brand”
Different disciplines—biologists, economists, statisticians, philosophers—have independently converged on collective intelligence principles because it's a fundamental solution to complexity that appears across domains.
“there's a group of people who were you know some philosophers some economists some statisticians some biologists kind of playing in this space of collective intelligence... a biologist in the space but if you think of ants each individual ant has um - models it's a map of the terrain of where the food sources are and they can sort of aggregate that collectively”
Charlie Munger's concept of a 'latticework of models' represents the ideal of accumulated mental models arrayed in your mind, available for deployment in decision-making, which requires building knowledge breadth while also identifying which models work best in which contexts.
“Munger has this wonderful quote about you want to sort of array your experiences on a latticework of models”
Atul Gawande's career exemplifies filling a 'structural hole' by combining deep expertise in medicine with knowledge of political philosophy, literature, and public policy, enabling him to connect domains that specialists do not bridge.
“your interview is Atul Gawande he made this fabulous point about his method of making a contribution to the world was sort of being able to communicate across different types of people in different areas right so he brought sort of a he'd been trained by doctors so he had his parents were doctors and so he he'd sort of absorbed what the medical profession was all in love with that but at the same time you had these his deep interest in science in this deep interest in sort of political philosophy and literature and sort of public policy”
The traditional hierarchical division of academic disciplines (physics, economics, biology) is outdated and economists should embrace the fact that their field is better understood as a teaching domain—teaching others to think clearly about cause-and-effect, incentives, and trade-offs—than as a unified body of knowledge about 'the economy'.
“a complex world your ability to contribute...requires filling a niche...but it's gonna be filling a niche...and there's a set of people who were you know some philosophers some economists some statisticians some biologists kind of playing in this space of collective intelligence”
The current American institutional landscape (gerrymandering, Electoral College giving small states disproportionate power, separate Federal Reserve quasi-government structure vs. integrated NASA/NIH) reflects decisions made when states and information flows were very different, and mechanism design suggests institutions should be revisited based on modern conditions rather than historical inertia.
“do you look at the American government at the moment it's kind of a mess everything from like sort of gerrymandering to the fact you know we had this electoral college that made a lot of sense when states are all equal size to roughly equal sized you know now some states that are tiny and still have the same number of senators those states that have 50 times as many people but even how we vote on things what what law but is under the purview of Congress like why do we have a separate in some sense like a financial system you think of like the Federal Open Market Committee in the Federal Reserve System that's quasi-governmental the FDIC is quasi governmental but NASA and the NIH are not quite as quasi yeah you know you think you'd like there's a deep question about um what institutions we use where that is underappreciated”
There is a hierarchy from data to information to knowledge to wisdom: data is raw experience and observations; information is structured data organized into categories and variables; knowledge is understanding correlative or causal relationships between pieces of information; wisdom is discerning which knowledge is relevant to apply to a particular problem.
“the wisdom hierarchy... the bottom out there's all this data right... on top of the data is information information is is that this isn't some we structure the world... then what you do on top of information is knowledge and what knowledge is is understanding either correlative or causal relationships between those pieces of information... what wisdom ins is wisdom is understanding which knowledge is to bring to bear on a particular problem”
Success and failure outcomes in competitions are shaped partly by luck (through mechanisms like preferential attachment and random variation) and partly by skill/ability, so successful people are right to think they're capable, but wrong to attribute their success entirely to skill—they benefited from fortunate positioning.
“they typically are and they tend to think they're there because they've had a lot of ability they have a lot of ability which means that they've got flexibility in terms of you know what tools they required but the point is getting them done recognizes for the group to be better right you want people in the daughter we they're tools right so it's tricky because these people think that you're people in successful singing that's so because they they've won you know because they're good when in fact you know maybe they've won because they just happen to have the right combinations of talents at the I kind of think of that in an evolutionary science right where we have consider a gene mutation today that might be selected as valuable but a million years ago the same gene mutation might have been you know negatively selected or filtered if you will because the environment has changed”
When deciding which models to use for a problem, one should first ask what class of problem it is: Is it a decision by a single isolated actor? A strategic situation involving multiple actors? Or an embedded choice within a larger social/ecological system? This question determines which models are appropriate.
“I think one of the first things you want to ask is who are the relevant actors right so is this a single actor who's sort of just making a decision or is this a strategic situation where someone is taking an action and they've really got to take into account what the action is of someone else”
Linear functions have constant derivatives (constant slope), concave functions have diminishing returns (decreasing slope), and convex functions have increasing returns, with convexity often hidden in projections like Japan's 1970s growth or China's recent expansion.
“Linearity register something has the same slope always so the next doll you know the next thing is with just as much looking... the fundamental as so many models throughout the book is some assumptions of either concavity which is sort of diminishing returns or convexity which is increasing Richards”
MIT's new data science school seeks to train people who are 'bilingual'—able to communicate between sophisticated AI models and the real world—because the critical gap is explaining why a model makes a prediction, not just using the model as a black box.
“MIT just started this this new school right this this new sort of data science school right there they get there first because the first one they're starting like 30 40 years together raised a billion dollars to this and one of the things they want people who are bilingual who can communicate between these sort of really sophisticated artificial intelligence models and the real world because the thing is people are afraid of sort of just throwing all this information into this giant a I am on like spit something out if you're using relatively simple models you actually can it is easy to be buys like well it's easy to sort of look deeply at this you know whatever model you doesn't say why is the model saying that's okay I like that a lot”
The career path you should consider depends on your cognitive style: you can specialize deeply in one or two domains (T-shaped human capital), generalize across many areas with some depth in a few (pi-shaped), or maintain broad awareness across many models with deeper expertise in one area.
“there's even people who I describe themselves as having the third human capitals in the shape of a tee right in the sense that like there's a lot of there's a whole bunch of things they no decent amount about and then one thing they know deep where other people describe themselves is like a symbol for pie right well there's two things they know pretty deep not as steep as the tea person and then a range of things that sort of connect those two areas”
The shift from physical beads to digital bits (information exchange) and the emergence of algorithms as a coordination mechanism alongside markets, hierarchies, and democracies means that the question of which institutional form to use for which problem should be reconsidered, as algorithms can now solve problems that previously required markets or hierarchies.
“we just got to let it go you decide to go somewhere I decide to go somewhere and then it's a total mess for the most part right but when we made these decisions about where we have markets hierarchies and democracies that was made in a world where there was no data no information technology where we're exchanging beads as opposed to setting bits through the mail but now there's this fifth bang right there's these algorithms and a lot of stuff a lot of things can be done by algorithms as opposed to markets hierarchies and democracies and there's a question because the sort of the cost of change for these institutions should we allocating problems right across these different institutional forms”
When a linear regression model and a group of people make divergent predictions about something, it is often valuable to investigate the discrepancy by talking to the people to understand what variables they are using that the linear model omitted, rather than simply averaging the two predictions or trusting one over the other.
“what you should do instead is if if the linear model and the people are close you know the predictions you probably should go to the linear model because it's really well calibrated right let's just probably gonna you know be better but if they're far apart if the linear model and the humans are giving very new predictions then you want to go talk to them I mean talk to the people and talk to the linear model now you can say how do you talk to them in your mind while you're looking to say what variables are in there what variables are the people using that the linear models not”
The design of models matters significantly because when leaders (CEOs, policy makers, traders) construct and deploy models, they are defining the reality and game rules that actors respond to; models are not just descriptive but prescriptive and generative of behavior.
“one reason people construct models is to build things right to the buildings to the policies to build strategies when you do that you're defining in some sense this you know the state space you're defining reality so if you tell your traders we're looking at these ratios you're defining the game for them right and so I think that it is I think the design aspect of models is often underlies overlooked underappreciated”
The simple models taught in school (force = mass × acceleration, supply and demand, etc.) were designed for an era when success depended mainly on individual mastery of a single domain, but in the modern complex world, success increasingly depends on filling a niche that requires combining skills and models from multiple domains.
“I think we could do a little bit more of sort of meta teaching in the sense that one of the one of the things that people really like about I did an online course called the model thinking which is a MOOC and one of the sort of trusts in there is that when you there's something that's called they borrowed from my colleague Mark Newman when he talks about distributions which is logic structure function”
The meritocratic ideal ('success = intelligence + effort') assumes that individual silos are independent and that hard work and ability determine outcomes, but this model only works in simple, low-dimensional domains where individual effort directly produces outcomes.
“I was putting this fabulous book called the rise of the meritocracy which is an old book like 50 years ago talks about sort of like you know successes intelligence plus effort right and it's it's actually where the word meritocracy came from if you imagine the world a collection of individual silos and the you know instead of this Sun the amount of grain in your silo depends on sort of how intelligent you're on how hard you work then it is all kind of about like your ability work hard get AIDS right in class develop these skills this is a very instrumental view of the world but in a complex world your ability to contribute”
There has been a large shift in academic fields (economics, political science, business) toward empirical research relative to theoretical or model-building research, which is valuable for measuring exact effect sizes but has a cost: empirical work takes the world as it is and cannot explore how different institutional designs or mechanisms might produce better outcomes.
“there's been because there's so much data there's this huge shift toward empirical research... this is you know I plot it the work is much better right there's much more data we can get a causality huge fan of it but I think there's a cost to that because what that a lot of energy record is doing is really nailing down exactly what's the what's the size of this effect what's the slow right of that line what's the size of the coefficient how significant incident right... however that's taking the world as it is and one of the really cool things about models are trained by these people who did mechanism design it's thinking about can we based on our understanding about people tagged redefine the world construct mechanisms institutions that work better”
In organizations, mental models can become self-replicating through coordination and efficiency: once a team adopts a shared mental model, it becomes easier for individuals to adopt that model than maintain their own, leading to homogenization of thinking which reduces diversity and increases predictability.
“I'm then I go work in some organization I'm working in some community of practice and I've got a collection of metal models I'm using it's it just becomes easy for me to start coordinating on other people's mental models right using other people's terminology just work it's it's more efficient and I know you know how to appeal to them how to persuade them how to interact with them how they see the world and then they're predictable”
Teaching people a broad set of mental models (even at a surface level) gives them awareness of multiple frameworks they can apply to problems, which provides robustness (they're less likely to rely on a single flawed model) and can give them a bonus from combining models in novel ways (the ensemble outperforms components).
“so there's a lot of stuff and I think about like where I'm gonna go do my laundry or what coffee shop I'm gonna go to but I probably don't sit around and rash I'm thinking about I just kind of like just follow some sort of routine and maybe I adapt that routine slowly maybe I learn a little bit but for the most part I might just follow rules so you want to ask how are people then you can ask yourself is my logic correct”
Bees and markets are nearly as effective as expert physicists at finding the highest point on a smooth, simple optimization landscape, but all three approaches fail on highly rugged, multi-peaked landscapes, suggesting that the difficulty of the problem determines whether decentralized (markets, swarms) or centralized (physics expertise, analysis) approaches work best.
“the bees did you miss assists right infection another five peaks did its ironically just a tiny bit better than the visitors and so we're talking about this afterwards and someone says well that's because the bees can take a derivative yeah nobody's like what because well no like to solve this you just gotta gotta take derivatives that they can set up find the highest they could find the highest point and then they could take derivatives because they could see who is lagging long right and it was only on the really hard problem physicists did the best”
Many academic researchers concentrate their intellectual effort on narrow, specialized domains and become world experts in those domains, but their work may only be meaningful to a very small circle; there is an obligation to consider whether that specialization is producing value that justifies its opportunity cost.
“one of the things I and fastening but the Academy is that where people would be in small departments and they'll study something and it gets really interested to them and they're the world's expert in that and that's great because we're advancing knowledge but outside of their small circle no one may find that interesting and I think that it's incumbent upon them to sort of think about you know are they using their talents in a way to I think you are they making that interesting to other people or at least intriguing to what they because I don't think you're adding that much value from these thirty people need your work”
Many human behaviors and cultural practices can be understood as solutions to local coordination problems (like how to greet someone, where to store ketchup, whether to use shoes indoors), where what matters is not the absolute choice but consistency with what others in your group do; these coordination equilibria collectively constitute culture.
“do you shake hands do you bow you fist bump right it doesn't matter what you do but what you do the same thing that other people do right so if you go to bowed I go to shake hands I'm gonna poke your eye out but it's not it's not gonna work so what you want is these are you in some sense sure what we call it in games like a pure coordination game”
The approach Page advocates for (pragmatic model pluralism) differs from a pure liberal arts approach: in literature and humanities, all perspectives may be worth considering and engaging with; but in practical decision-making (investments, drug approval, policy), there is an 'end game' (measurable outcomes), so some models should be tried, tested, and discarded if they don't work.
“I think where the difference is is that I'm a you know I think I'm a pragmatist in a way right I mean I can see so many opportunity and so I feel like I'm coming at from a much more sort of pragmatic perspective in terms of going out there and making a difference in the world as opposed to just purely appreciating all these different ways of seeing things and the reason that distinction matters is if in literature it could be that every perspective is worth considering right in engaging in thinking about because there's no there's no end game ironically given the name of the store you snatch it”
There is a strategic value in perspective-taking across different stakeholder groups (shareholders, government, employees, customers, etc.) because each group uses different mental models to understand a situation, and understanding those models allows you to anticipate behavior and communicate more effectively.
“I want to have those models better than mine but it's still worth it for me to hang on to my mental model because it's giving that diversity right so collectively it's worthwhile but there's gonna be so again back to the point you raised early about evolution and this is where the many model thinking before fun is you realize like so I'm then I go work in some organization I'm working in some community of practice and I've got a collection of metal models I'm using it's it just becomes easy for me to start coordinating on other people's mental models”
Successfully contributing to the world requires combining three things: (1) genuine passion and love for practicing the activity, (2) some level of innate ability or natural talent in the domain, and (3) the ability to connect what you do to something meaningful or valuable to others.
“what you want to think about is finding something that combines three things you have to really love it you got it it has to be your passion he's kind of you've got to love the practice of it so if you you know a great basketball player isn't someone of great ability it's someone who loves practicing basketball a great musician is someone yeah he's got some ability there but they love practicing music so you've got to you really got to enjoy the practice of the thing you do second thing is you've got to have some innate ability right”
The logic-structure-function framework says that when you see a structure or pattern in the world, you must ask (1) what logic produces it, (2) why do we see that particular structure and not others, and (3) does it matter functionally.
“when you there's something that's called they borrowed from my colleague Mark Newman when he talks about distributions which is logic structure function so if you see some sort of structure or pattern out there in the world there has to be some logic as to how that came to be and then you also want to ask yourself is there some functionality of that structure does it matter right”