Chapitre 1
Why Multiple Models Matter More Than Ever
Have you ever wondered why some of the world's most successful people-from Warren Buffett to Ray Dalio-rely on mental models to navigate complexity? In our increasingly interconnected world, understanding complex systems has become essential. Scott E. Page's "The Model Thinker" offers a revolutionary approach: instead of searching for a single perfect model, we should cultivate a diverse portfolio of models to understand our complex reality. This book has become required reading at hedge funds like Bridgewater Associates and tech giants like Google, where decision-makers face unprecedented complexity daily. Beyond business circles, the book has gained cult status among intellectual heavyweights who recognize that traditional single-model thinking is increasingly inadequate for our interconnected age.
Chapitre 2
The Many-Model Approach: A New Framework for Thinking
In today's data-rich world, you might assume we need models less than before. Counterintuitively, the opposite is true. The unprecedented dimensionality and granularity of modern data-from customer purchases to student performance to agricultural measurements-doesn't automatically translate to understanding. Data alone can be misleading; models help us make sense of information streams.
Consider this paradox: we have more information about the world than ever before, yet that world has become increasingly complex as technology shrinks time and distance while making economic, political, and social actors more responsive. This creates a situation where any single model will likely fail, making many-model thinking essential.
The many-model approach builds on ancient wisdom that multiple perspectives lead to better understanding. This idea traces from Aristotle through the great-books movement to modern thinkers like Maxine Hong Kingston, who advocated making one's mind "large enough for paradoxes." Despite its intuitive appeal, many-model thinking challenges traditional educational approaches that teach a one-to-one relationship between problems and models.
Take income inequality, for example. A single-model thinker might attribute rising inequality solely to technological change. But a many-model thinker would recognize multiple contributing factors: technological disruption, winner-take-all markets, changing social norms around executive compensation, and declining labor bargaining power. Each model provides a piece of the puzzle, and together they create a more complete understanding.
The benefits of many-model thinking extend beyond professional settings to civic engagement and personal decision-making. By applying multiple models as frames, we develop nuanced, deep understandings of complex issues. This approach has practical value for knowledge workers who increasingly work with data, improving their ability to reason, explain, design, communicate, predict, and explore.
Chapitre 3
The Wisdom Hierarchy: From Data to Understanding
To understand why many-model thinking matters, consider the wisdom hierarchy. At the bottom lies raw, uncoded data-events and phenomena without structure, such as temperature readings, stock prices, or population counts. Above that is information, which names and categorizes data into meaningful patterns and relationships. For instance, temperature readings become climate trends, stock prices become market indicators, and population counts become demographic profiles.
Knowledge organizes information into understandings of correlative, causal, and logical relationships, often taking model form. This level involves recognizing patterns, understanding why they occur, and being able to make predictions. For example, knowing that interest rates typically affect housing prices represents knowledge of economic relationships. At the top sits wisdom-the ability to identify and apply relevant knowledge in complex, real-world situations, often requiring judgment about which models to use and when.
Wisdom requires many-model thinking through three primary approaches: model selection, model averaging, and parallel model application. In model selection, we choose the most appropriate framework for a specific situation. Model averaging combines predictions from multiple models to reduce individual biases. Parallel application uses different models simultaneously to gain multiple perspectives on a problem, similar to how doctors use various tests to diagnose an illness.
Consider a physics problem about a falling stuffed cheetah. A novice might automatically apply the simple gravity model (d = 12gt2), but a wise thinker recognizes that air resistance becomes significant for light, fluffy objects. They would instead apply the terminal velocity model, incorporating air resistance and the object's mass and surface area. The difference in predictions could be substantial - the gravity model might predict the cheetah hitting the ground at 120 mph, while the terminal velocity model correctly predicts 30 mph.
Real-world applications demonstrate the power of many-model thinking. During Iceland's 2008 financial crisis, Oracle's treasurer needed to decide whether to withdraw the company's overseas assets. Instead of panicking based on dramatic headlines, he applied multiple economic models. One compared Iceland's GDP ($14 billion) to McDonald's six-month revenue ($23 billion). Another assessed currency exchange risks and banking system stability. By combining these models, he concluded the crisis, while severe for Iceland, posed minimal risk to Oracle's assets.
Becoming a many-model thinker requires developing working knowledge of diverse models - understanding their core principles, assumptions, and applications without necessarily achieving deep expertise in each. For example, a business leader might need working knowledge of game theory, network models, and behavioral economics without being an expert in any single field. Think of models as specialized tools in a cognitive toolkit - like a carpenter who knows when to use a hammer versus a saw, a many-model thinker knows when to apply market models versus behavioral models.
The key is recognizing that no single model captures all aspects of complex reality. Each model offers unique insights while making simplifying assumptions. By maintaining multiple models, we can triangulate better solutions and avoid the pitfalls of single-model thinking.
Chapitre 4
Why Models Matter: The REDCAPE Framework
Models serve as simplifications of reality, analogies, or fictional worlds that generate insights. They help us understand when certain results hold true-just as the Pythagorean theorem only applies to right triangles, models reveal conditions under which diseases spread, markets function, or crowds make accurate predictions.
The REDCAPE framework captures seven key uses of models:
Reason: Models help identify conditions and deduce logical implications that aren't obvious through intuition alone. They uncover impossibilities (like Arrow's theorem showing no perfect voting system exists), reveal paradoxes (such as Simpson's paradox where subgroup trends reverse in the aggregate), and establish mathematical relationships (like the friendship paradox where people's friends have more friends than they do).
Explain: Models provide logical explanations for empirical phenomena across disciplines. While physics models like Boyle's Law work with remarkable accuracy because gases consist of simple, numerous parts following fixed rules, social phenomena are less predictable since people are heterogeneous, interact in small groups, and respond to social influences.
Design: Models provide frameworks for designing complex systems and institutions. When economists designed the FCC spectrum auction in 1993, they faced the challenge of selling regional cellular licenses whose values were interdependent. Using game theory, computer simulations, and statistical models, they created a multiple-round auction that has since raised nearly $60 billion for the government.
Communicate: Models improve communication by creating common, precise representations of concepts. Unlike vague statements like "bigger, faster things generate more power," models like F = MA relate measurable quantities with mathematical precision.
Act: Good actions require good models. The 2008 bailout of AIG with $182 billion exemplifies this-the government's decision wasn't to save AIG specifically, but to prevent the devastating impact its failure would have had on the entire financial system. By contrast, Lehman Brothers was allowed to fail because it didn't occupy a central position in the financial network.
Predict: Models have long been used for prediction across domains-from weather forecasting to criminal behavior. With increasingly granular data, prediction capabilities have grown. Models can predict both general trends and specific events, as demonstrated when probabilistic ocean current models helped locate the Air France flight AF 477 fuselage a year after it crashed in the Atlantic.
Explore: Models let us explore intuitions and possibilities, including unrealistic scenarios that spark creativity. We can ask "what if" questions about policy changes or even impossible scenarios that help reveal the limits of processes like evolution.
Chapitre 5
The Science of Many Models: Mathematical Foundations
The value of many-model thinking isn't just philosophical-it's mathematically proven through several fundamental theorems that demonstrate why multiple perspectives consistently outperform single models. Two key mathematical principles form the foundation of this approach:
The Condorcet jury theorem, developed in the 18th century, provides mathematical proof that collective wisdom can exceed individual judgment. When multiple models each have better-than-random accuracy (>50%) and make independent classifications, their majority vote will be more accurate than any individual model. For example, if three medical diagnostic tools each have 70% accuracy and operate independently, their combined verdict achieves roughly 80% accuracy. As ecologist Richard Levins memorably noted, "our truth is the intersection of independent lies." This principle explains why medical diagnoses often improve when multiple specialists contribute their perspectives.
The diversity prediction theorem, formulated by Scott Page, quantifies how model diversity improves collective accuracy for numerical predictions through the equation: Many-Model Error = Average-Model Error - Diversity of Model Predictions. This mathematical relationship explains why diverse models with offsetting errors produce more accurate averages than any single model, supporting the "wisdom of crowds" phenomenon. For instance, when estimating market prices, combining technical analysis, fundamental analysis, and sentiment indicators often yields better results than relying on any single approach.
However, these theoretical foundations come with important practical limitations. Constructing truly diverse, independent models faces real-world constraints. Research in binary categorization models reveals that the number of relevant attributes fundamentally constrains how many distinct, useful models we can create. With just two attributes creating four categories, we can construct at most seven accurate binary models. This limitation explains why adding more models eventually yields diminishing returns.
Empirical evidence from various fields confirms these theoretical constraints. Google's hiring process research found that interviewer accuracy improves dramatically with the first few interviewers but plateaus after four. Similar patterns emerge in economic forecasting, where combining three to five different methodologies captures most available predictive power, but additional models add minimal value. Weather forecasting shows comparable results - while combining multiple models improves accuracy, the benefit of adding more models decreases significantly after incorporating the first few major prediction systems.
The practical implications are clear: three to five diverse models often capture most available predictive power. This "sweet spot" balances the benefits of multiple perspectives against the complexity and potential redundancy of additional models. Organizations can optimize their decision-making by focusing on maintaining a small set of truly diverse approaches rather than pursuing an ever-expanding collection of similar models.
Chapitre 6
The One-to-Many Principle: Maximizing Learning Efficiency
To maximize learning efficiency, we should master a modest number of flexible models and apply them creatively across domains-a "one-to-many" approach. This interdisciplinary application of models represents a paradigm shift from traditional academic siloing, where models were confined to their original disciplines. The power lies in recognizing universal patterns that manifest across seemingly unrelated fields.
The cross-pollination of models has produced remarkable insights across diverse fields. Paul Samuelson revolutionized economics by adapting thermodynamic equilibrium models from physics to explain market behavior and price mechanisms. Anthony Downs transformed political science by applying Harold Hotelling's ice cream vendor location model to explain how political parties position themselves ideologically. Social scientists have successfully borrowed particle interaction models from physics to illuminate social phenomena like poverty traps, crime hotspots, and segregation patterns.
The power of the one-to-many principle is elegantly demonstrated through the seemingly simple exponential formula X^N. While most familiar as a tool for calculating areas (N=2) or volumes (N=3), this model reveals profound insights across numerous domains when applied creatively. In architecture and engineering, it helps optimize structural designs. In biology, it explains scaling laws across species. In economics, it illuminates everything from compound interest to network effects.
The supertanker example perfectly illustrates how this model unveils counterintuitive business insights. As ships increase in size, their surface area (representing construction and maintenance costs) grows as S^2, while their cargo volume (determining revenue potential) grows as S^3. This creates an inherent economy of scale where each size increase yields proportionally greater benefits, explaining the economic drive toward ever-larger vessels despite technical challenges.
The model's application to human physiology reveals important insights about body scaling. In the context of BMI, the formula exposes why this widely-used metric systematically misclassifies certain body types. Since BMI uses height^2 rather than height^3, it fails to account for the three-dimensional nature of human bodies, leading to systematic bias against tall or muscular individuals.
In biological systems, the X^N relationship explains crucial scaling laws. The surface-area-to-volume ratio, growing at different exponential rates, determines why small animals must maintain faster metabolic rates for thermal regulation. A mouse, with its proportionally larger surface area relative to volume, loses heat roughly seventy-five times faster than an elephant, necessitating a correspondingly higher metabolic rate to maintain body temperature.
Perhaps most provocatively, when applied to career advancement, this model reveals how subtle biases can create dramatic long-term disparities. In corporate hierarchies, a seemingly minor 10% difference in promotion rates between demographic groups, when compounded across multiple promotion opportunities, results in men becoming nearly thirty times more likely to reach CEO positions after fifteen rounds of promotion. This mathematical insight helps explain persistent leadership gaps and underscores the importance of addressing even small systematic biases.
Chapitre 7
Modeling Human Actors: The Challenge of Complexity
Modeling people presents unique challenges because humans defy simple characterization. Unlike physical objects like carbon atoms, people are diverse in preferences and capabilities, socially influenced, error-prone, purposive, capable of learning, and possess agency.
Rather than making ad hoc assumptions for each model, which would limit cross-model thinking, we can model people as either rule-based actors or rational actors. Rule-based actors follow either simple fixed rules or adaptive rules that change based on information, past success, or observation of others.
The rational-actor model assumes people make optimal choices to maximize a utility function. Despite its simplicity, the model can yield useful insights, as demonstrated by a basic housing consumption model showing people spend roughly one-third of income on housing regardless of price or income level.
Though people regularly violate the conditions for rational behavior, the model persists for several reasons: people often act "as if" optimizing through effective heuristics, learning pushes behavior toward optimality over time, high-stakes decisions receive more careful consideration, the model provides unique testable predictions, ensures internal consistency, and serves as a useful benchmark for policy design.
The rational-actor model faces challenges from psychologists, economists, and neuroscientists who have documented numerous cognitive biases. These include status quo bias, ignoring base rates in probability calculations, overvaluing certainty, and loss aversion. In prospect theory, people choose differently when identical scenarios are framed as losses versus gains-preferring certain gains but gambling to avoid certain losses.
When modeling human behavior, we face a tension regarding how intelligent to make our agents. Given these complexities, we should use multiple models rather than seeking a single best approach. Even seemingly unrealistic models like rational-choice or zero-intelligence models provide value-the former reveals incentive effects while the latter shows how much intelligence matters.
Chapitre 8
Understanding Distributions: Normal and Power-Law
Distributions form a core knowledge base for modelers, helping us understand why certain patterns emerge and why they matter. Two particularly important types are normal distributions (bell curves) and power-law distributions (long tails).
Normal distributions characterize phenomena like heights, weights, and test scores. They're symmetric around their means with predictable properties-approximately 68% of outcomes fall within one standard deviation, 95% within two, and 99% within three. They arise through the central limit theorem: the sum of many (>=20) independent random variables will approximate a normal distribution.
Normal distributions help explain why exceptional outcomes occur more frequently in small populations. The standard deviation of a mean decreases with the square root of population size, making extreme values more likely in smaller samples. This explains why both the safest and most dangerous places tend to be small towns, and why counties with highest cancer rates have small populations.
Power-law distributions appear in diverse phenomena including city populations, species extinctions, web links, book sales, academic citations, war casualties, and natural disasters. Unlike normal distributions, they feature "long tails" with significant probabilities of extreme events.
If human heights followed a power law similar to city populations, the United States would have one person as tall as the Empire State Building, 10,000 people taller than giraffes, and 180 million people shorter than 7 inches. These distributions arise from non-independent events with positive feedbacks-when one person buys a popular book, others follow; when one tree catches fire, it spreads to others.
Long-tailed distributions have profound implications for equity, catastrophes, and volatility. They create a few big winners and many losers compared to normal distributions. In markets with positive feedback effects, slight advantages can compound into massive success, as shown in music lab experiments where social influence created bigger winners but not necessarily better ones.
Chapitre 9
From Linear to Complex: Building Sophisticated Models
While linear models represent the simplest functional relationships between variables, most real-world phenomena aren't linear. Nonlinear functions can curve downward (concave) or upward (convex), form S-shapes, kink, jump, and squiggle.
Convex functions have an increasing slope: the function's value increases by larger amounts as a variable increases. The exponential growth model exemplifies convexity, describing how a population or resource grows based on its initial value, growth rate, and time periods. This yields the "rule of 72," which approximates how long it takes for something to double: periods to double ~ 72/R.
Concave functions have decreasing slopes-the opposite of convex functions. They exhibit diminishing returns, where each additional unit provides less value than the previous one. Almost all goods demonstrate this property: leisure, money, ice cream, even time with loved ones. Concavity implies both preference for diversity and risk aversion.
Economic growth models reveal patterns across countries and guide policy decisions like savings rates. The Solow Growth Model shows how long-run equilibrium occurs when investment equals depreciation. Most significantly, equilibrium output increases as the square of technological improvements, creating an "innovation multiplier"-innovations directly increase outputs and indirectly lead to more capital investments.
Beyond these basic models, we can explore more complex structures like networks, which are ubiquitous structures connecting our world-from social relationships and trade to food webs and financial systems. Network models help us understand phenomena like the friendship paradox (people typically have fewer friends than their friends do), six degrees of separation, and network robustness to failure.
Chapitre 10
Applying Many-Model Thinking to Real-World Problems
The true power of many-model thinking emerges when applied to complex real-world problems. Consider two examples: the opioid epidemic and economic inequality.
The opioid epidemic's scale is staggering-in 2015, over 4% of Massachusetts residents had opioid use disorders, with nationwide figures showing 200 million prescriptions, 10-12 million misusers, and 30,000 deaths in 2016.
Four models help explain this crisis: First, a multi-armed bandit model shows how clinical trials demonstrating pain relief led to approval, with initial addiction rates under 1%. Second, a Markov model reveals how seemingly small changes in addiction rates (1% to 2.5%) can increase equilibrium addiction levels five-fold, explaining why longer prescriptions proved catastrophic. Third, systems dynamics models track flows between pain, opioid use, addiction, and heroin use, showing how restricting opioid access can increase heroin use. Finally, network models explain geographical clustering in rural areas beyond what random variation would predict.
Economic inequality demands multi-model analysis because of its complexity and importance. A technology-human capital model shows how automation increased demand for educated workers while immigration expanded low-skilled labor supply. A positive feedback model explains how digital connectivity creates winner-take-all markets. The spatial voting model helps explain why American CEOs now earn 300 times average worker pay compared to just 25 times in 1966, far exceeding international norms.
By applying many models as an ensemble, we can uncover multiple causes behind complex problems, revealing the limitations of any single framework. This approach could be extended to challenges ranging from obesity to climate change.
Chapitre 11
The Wisdom of Model Thinking
The many-model approach applies equally to business decisions-from product launches to supply chain design. For instance, when launching a new product, companies might simultaneously employ market segmentation models, diffusion models for adoption rates, and game theory models for competitive responses. Supply chain optimization might combine inventory models, network flow models, and queuing theory to capture different aspects of the system. While this approach consistently outperforms gut instincts, we must maintain humility about its limitations. Complex systems may resist even our best modeling efforts, as demonstrated by financial market crashes that defy traditional economic models.
The process of building and applying multiple models helps us uncover hidden interdependencies and understand why certain problems are difficult to solve. Consider climate change modeling, which combines atmospheric physics, ocean circulation patterns, and human behavior models. Each model reveals different aspects of the challenge, and together they illuminate why simple solutions often fall short. When examining organizational behavior, combining psychological models of individual motivation with network models of group dynamics and economic models of incentives provides a richer understanding than any single approach.
The acknowledgment that all models are wrong should motivate us to build a diverse portfolio of models capable of producing collective wisdom. Like a jury reaching a verdict, multiple models can triangulate truth more effectively than any individual perspective. This approach has proven particularly valuable in fields like weather forecasting, where ensemble predictions combining multiple models consistently outperform single-model forecasts. Beyond their pragmatic benefits, models offer intellectual beauty and joy-a structured game where we make assumptions, write rules, and explore within logical boundaries, much like mathematics or music composition.
As Einstein famously said, "It can scarcely be denied that the supreme goal of all theory is to make the irreducible basic elements as simple and as few as possible without having to surrender the adequate representation of a single datum of experience." The many-model thinker embraces this wisdom, recognizing that while all models are wrong, some are useful-and together, they can help us navigate an increasingly complex world. This approach has been validated across disciplines, from public health responses to pandemics to environmental conservation efforts, where multiple modeling approaches have helped decision-makers better understand and address complex challenges.
The key lies in maintaining both rigor and flexibility-knowing when to apply which models, understanding their assumptions and limitations, and being willing to revise or discard models when they no longer serve their purpose. Success comes not from finding a perfect model, but from skillfully combining and updating multiple imperfect ones as our understanding evolves.
