In 1976, a British statistician named George Box wrote a sentence that should be mounted on the wall of every founder's office, every trading floor, and every research laboratory: "All models are wrong, but some are useful."
The statement is not modesty. It's not a hedge. It's a precise technical claim about the relationship between representations and reality — and it has profound consequences for anyone who makes decisions under uncertainty, which is everyone.
A model is any simplified representation of a complex system. A financial forecast is a model. A business plan is a model. Newton's laws of motion are models. Your mental picture of your customer is a model. Porter's Five Forces is a model. The map on your phone is a model. Every framework, every heuristic, every equation you use to make sense of the world is, by definition, a simplification — and simplifications, by definition, leave things out.
Box's insight wasn't that we should stop using models. He spent his career building them. He developed the Box-Jenkins method for time series analysis, created Box-Behnken designs for response surface methodology, and made foundational contributions to quality control and experimental design. The man who declared all models wrong was one of the most prolific model-builders in the history of statistics. His point was epistemological: the value of a model lies not in its truth — no model is true — but in its utility. The question isn't "is this model right?" The question is "is this model useful enough for the decision I need to make?"
The distinction predates Box by decades. Alfred Korzybski, the Polish-American philosopher, coined the phrase "the map is not the territory" in 1931 — arguing that human knowledge always operates through abstractions, and confusing the abstraction with the reality it represents is the root cause of most intellectual errors. Gregory Bateson extended this in 1972: "The map is not the territory, and the name is not the thing named." The same principle appears in statistics (all regression models omit variables), in physics (Newtonian mechanics is "wrong" but lands rockets), in medicine (diagnostic criteria approximate, not capture, disease), and in military strategy (no plan survives contact with the enemy, as Helmuth von Moltke observed in the 1870s).
What makes this a Tier 1 framework — a model about models, a lens that sharpens every other lens — is its universality. Every other mental model in this collection is, by Box's standard, wrong. First Principles Thinking is wrong — it assumes you can reach bedrock truth, when in practice your "fundamental truths" are themselves models of deeper realities. Inversion is wrong — it assumes you can identify the relevant failure modes, when the most dangerous failures are the ones your model doesn't contain. Network Effects is wrong — it captures one dynamic among dozens that determine a platform's trajectory. Each model is useful. None is true.
The practical consequence: the most dangerous operator in any room is the one who has forgotten that their framework is an approximation. They've reified the model — treated the map as the territory — and they will be blindsided by the territory's features that the map didn't include. The more confident they are in their model, the larger the surprise when it breaks.
This is why crises so often catch sophisticated actors harder than simple ones. The sophisticated actor built an elaborate model, tested it extensively within its boundary conditions, and developed justified confidence. The simple actor held a rough heuristic loosely and was prepared to abandon it when the world shifted. Sophistication without epistemic humility is a trap. Box's sentence is the escape hatch.
Newton's laws governed physics for over two centuries. They put astronauts on the Moon. They are also wrong — they break down at relativistic speeds, at quantum scales, and in strong gravitational fields. Einstein's models corrected Newton's at these extremes. Einstein's models are also wrong — they remain incompatible with quantum mechanics at the Planck scale. The point isn't that Newton or Einstein failed. The point is that useful models have boundary conditions, and catastrophic errors occur when you operate outside those boundaries without knowing you've crossed them.
The 2008 financial crisis was, at its core, a failure to remember that all models are wrong. The Gaussian copula model, developed by David X. Li in 2000, gave banks a mathematical framework for pricing correlations between mortgage defaults. The model assumed that historical default correlations — which were low during the housing boom — would persist. It assumed the inputs were stationary. It assumed the past was a reliable guide to the future. Within those boundary conditions, the model worked. Outside them — when housing prices declined nationally for the first time since the Great Depression — it didn't just fail. It failed catastrophically, because the institutions using it had built trillions of dollars of exposure on the assumption that the model was reality.
There's a gradient here worth understanding. Not all wrongness is equal. A model can be wrong and harmless (your weather app says 72°F, the actual temperature is 71°F), wrong and useful (Newton's laws, which are technically wrong but sufficient for building bridges and landing on the Moon), or wrong and catastrophic (the Gaussian copula, which was technically elegant and structurally lethal). The critical variable isn't the magnitude of the error. It's the relationship between the error and the decision it's supporting. A model that's wrong by 5% in a domain where you have a 30% margin of safety is fine. The same model, wrong by 5% in a domain where you're levered 30-to-1, is fatal.
The non-obvious insight: the danger of a wrong model isn't proportional to how wrong it is. It's proportional to how much confidence you've placed in it. A rough heuristic held loosely does less damage than a precise model held with certainty. The roughness forces you to maintain awareness that you're approximating. The precision seduces you into forgetting.
Section 2
How to See It
The fingerprint of this model is a specific kind of failure: someone confuses a useful simplification for an accurate description of reality, and the simplification's omissions come due. Once you recognise the pattern, you'll see it everywhere — and its conspicuous absence in the decisions that produce the most spectacular collapses.
Investing
You're seeing All Models Are Wrong when a quantitative fund's risk model shows a "25-standard-deviation event" — a statistical impossibility that reveals the model's distributional assumptions were wrong, not that reality misbehaved. Long-Term Capital Management's models in 1998 assumed asset correlations would remain stable during stress. They didn't. Goldman Sachs's CFO David Viniar said in August 2007 that markets were experiencing "25-standard-deviation moves, several days in a row." Under a normal distribution, a 25-sigma event should occur once every 10^135 years — far longer than the age of the universe. The model wasn't experiencing an anomaly. The model was wrong.
Business
You're seeing All Models Are Wrong when a company optimises for a proxy metric and loses sight of the underlying reality it was supposed to measure. When Wells Fargo set aggressive cross-selling targets in the 2000s, the model said: more accounts per customer equals more revenue equals more profit. Employees opened 3.5 million fake accounts between 2002 and 2016 because the model measured accounts, not customer value. The proxy had become the target — Goodhart's Law in action — because the organisation forgot that the metric was a model of reality, not reality itself.
Science
You're seeing All Models Are Wrong when a paradigm shift reveals that an entire field was operating within an approximation's boundary conditions without knowing it. Ptolemaic astronomy — Earth at the centre, planets on epicycles — predicted planetary positions accurately enough to navigate for over a millennium. It was useful. It was also wrong. Copernicus, Kepler, and Newton didn't just offer better predictions. They revealed that Ptolemy's model was a local approximation that worked for navigators but collapsed as a description of actual celestial mechanics. The model served its users well for centuries. Its wrongness only mattered when the questions changed.
Technology
You're seeing All Models Are Wrong when a startup's customer persona, validated through interviews and surveys, fails to predict actual purchasing behaviour at scale. The persona was a model — a useful compression of heterogeneous customer data into an archetype. At small scale, the archetype matched the early adopters well enough. At scale, the variance in actual customer behaviour exceeded what the persona could capture, and the product decisions built on that persona started misfiring. The model wasn't updated because the team had stopped treating it as a model and started treating it as a fact.
Section 3
How to Use It
Decision filter
"What assumptions is my current framework making? Where are its boundary conditions? What would reality have to look like for this model to fail — and would I notice before the failure became catastrophic?"
As a founder
Every strategic framework you use — TAM calculations, competitive moats, customer segmentation, unit economics projections — is a model. Treat each one as a hypothesis, not a conclusion. Build explicit review points — quarterly at minimum — where you ask: has the territory changed since we drew this map? Which assumptions have been validated by evidence, and which are still untested?
Jeff Bezos embedded this principle into Amazon's operating culture with his "Day 1" philosophy. In his 2016 letter to shareholders, he warned specifically against using process as a proxy for outcome: "The process becomes the thing. You stop looking at outcomes and just make sure you're doing the process right." That's the reification trap — the model (process) replacing the reality (customer value) it was designed to approximate. Bezos's antidote was structural: keep asking whether the proxy still correlates with the thing it's supposed to measure.
As an investor
When your thesis rests on a valuation model — DCF, comparable multiples, sum-of-the-parts — remember that the model's output is a function of its assumptions, not a description of the company's actual worth. Change the discount rate by 200 basis points and the terminal value shifts by 30–40%. The model isn't telling you what the company is worth. It's telling you what the company would be worth if your assumptions hold. The assumptions won't hold perfectly. The question is how much imprecision you can tolerate.
Howard Marks captures this precisely: "There are two kinds of forecasters: those who don't know, and those who don't know they don't know." The first group is using wrong models consciously. The second has forgotten the models are wrong. Bet on the first group. They're more likely to update when reality diverges from the forecast, because they never expected the forecast to be perfectly right in the first place.
As a decision-maker
When your team presents data-driven recommendations, ask one question before anything else: what did the model leave out? Every dataset is a sample, every analysis uses assumptions, and every recommendation is contingent on those assumptions holding. The recommendation may still be correct. But the confidence with which it's presented should be calibrated to the model's known limitations, not to the precision of its output.
A spreadsheet that calculates to four decimal places creates an illusion of precision that the underlying assumptions don't support. Require every analytical recommendation to include a section on "what this analysis assumes" and "conditions under which this conclusion would reverse." The discipline of articulating the model's assumptions is itself valuable — it keeps the team aware that they're navigating by map, not by territory.
Common misapplication: The trap is using "all models are wrong" as a licence for intellectual nihilism — rejecting all frameworks because none are perfect. That inverts Box's point entirely. He said models are wrong and useful. The operative conjunction is "and." Refusing to use models because they're imperfect is like refusing to use a map because it doesn't show every pothole.
A second misapplication: using the principle selectively, invoking "all models are wrong" only when a model produces conclusions you don't like, while treating models that support your position as truth. This is confirmation bias wearing epistemological clothing. The principle applies uniformly or not at all.
The founders who cite this principle to justify ignoring data are more dangerous than the ones who've never heard of it.
Section 4
The Mechanism
Section 5
Founders & Leaders in Action
The operators who internalise "all models are wrong" don't abandon frameworks. They hold them differently — with a grip firm enough to act on but loose enough to update when the territory diverges from the map. The pattern across these cases is consistent: the advantage comes not from having better models, but from updating them faster than the competition.
What's striking is how varied the domains are — statistics, reflexive markets, quantitative trading, theoretical physics, corporate strategy — and how uniform the underlying discipline is. Each person treats their own frameworks as provisional. Each builds systems to detect model decay. Each has been rewarded not for the quality of their initial model but for the speed and honesty of their revisions.
Munger's entire intellectual framework — the "latticework of mental models" — is explicitly built on the premise that every individual model is wrong. His argument, articulated across decades of speeches and shareholder meetings, is that no single discipline's framework captures enough of reality to make reliable decisions. Economics misses psychology. Psychology misses incentive structures. Finance misses competitive dynamics. The lattice compensates for each model's blind spots by layering multiple models from multiple disciplines.
At the 1994 USC Business School talk, Munger described the approach: "You've got to have models in your head. And you've got to array your experience — both vicarious and direct — on this latticework of models." The key word is "latticework" — a structure where each model supports and constrains the others. A single model gives you a single lens. A lattice gives you stereoscopic vision.
Munger never relied on one framework to evaluate an investment. A company that looked cheap on a DCF model but failed the "durable competitive advantage" model didn't get a pass. The models had to converge. When they didn't, the disagreement itself was information — a signal that at least one model was wrong in a way that mattered.
The practical consequence was Berkshire's avoidance of nearly every major investment catastrophe of the past four decades. When dot-com companies in 1999 looked attractive through growth-rate models but failed through cash-flow models, the disagreement was the signal. When structured credit products in 2006 looked safe through ratings-agency models but failed through basic leverage analysis, the disagreement was the signal. Munger's lattice didn't predict these crises. It detected the model disagreements that preceded them — which is a more reliable process than prediction, because it doesn't require knowing which model is right. It only requires noticing when the models disagree.
Soros built his theory of reflexivity on an explicit acknowledgment that all market models are wrong — and that the wrongness itself is the opportunity. Classical economic models assume markets tend toward equilibrium because participants act on accurate information. Soros observed the opposite: market participants act on models of reality that are inevitably distorted, and their actions based on those distorted models change reality, which further distorts the models.
The mechanism is circular: wrong model → action based on wrong model → reality changes → model becomes more wrong → action intensifies → reality changes further. Soros saw this pattern in the 1992 pound crisis, in the 1997 Asian financial crisis, and in the 2008 subprime collapse.
His edge wasn't having better models. It was recognising faster than other participants that the prevailing model had crossed its boundary conditions. The 1992 pound trade — $10 billion against sterling, yielding over $1 billion in profit — was a bet that the Bank of England's model of sustainable exchange rates was wrong and that reality would force an update. The 1997 Asian crisis repeated the pattern: Thailand, Indonesia, and South Korea had pegged their currencies to the dollar based on a model of stable capital flows and export-driven growth. When capital flows reversed, the model collapsed, and Soros's fund was positioned for the correction.
The deeper lesson: in reflexive systems, the model doesn't just describe reality. It distorts reality. And when the distortion becomes unsustainable, the correction is violent. Most market participants are unconscious model users — they act on frameworks they've never examined. Soros was a conscious model critic — he examined everyone else's frameworks, identified where they were wrong, and positioned for the correction.
The distinction between unconscious model use and conscious model criticism is the practical divide this principle creates. Both groups use wrong models. Only the second group knows it.
Renaissance Technologies, the most successful quantitative hedge fund in history, was built on a mathematical framework that treats every model as provisional. Simons, a former mathematician who broke codes for the NSA and proved foundational results in differential geometry, approached financial markets with a scientist's epistemology: models are hypotheses to be tested, not truths to be trusted.
The Medallion Fund reportedly generated average annual returns of 66% before fees from 1988 to 2018 — a track record unmatched in the history of finance. The fund ran thousands of statistical models simultaneously, each capturing a small, temporary pattern in market data. No individual model was expected to remain useful indefinitely. Patterns decay. Markets adapt. Other participants discover and arbitrage away the same signals.
Renaissance's advantage wasn't in building perfect models — it was in building infrastructure to detect when models stopped working and replace them before the decay eroded returns. The average holding period was reportedly under two days. Every model was disposable. The fund treated model mortality as an engineering parameter, not a failure.
Simons described the approach in rare public comments: the fund operated like a scientific laboratory, constantly generating hypotheses, testing them against data, and discarding them when they failed. Every model was disposable. What endured was the meta-process — the system for finding, validating, and retiring models faster than the competition. Renaissance internalised "all models are wrong" not as a philosophical caveat but as an engineering constraint that shaped every aspect of the firm's operations.
Feynman's 1974 Caltech commencement address, "Cargo Cult Science," is the most precise articulation of this principle applied to scientific practice. Feynman described researchers in the South Pacific who, after World War II, built bamboo control towers and wooden headsets to attract the cargo planes that had stopped coming. They had a model — the form of an airfield — but the model was disconnected from the underlying mechanism.
His point wasn't about the South Pacific. It was about scientists who follow the form of scientific method without maintaining genuine intellectual honesty about their models' limitations. "The first principle is that you must not fool yourself — and you are the easiest person to fool."
Feynman's insistence on re-deriving established results rather than citing them was a direct application of this principle. When he encountered a claim in physics, he didn't ask "who published this?" He asked "can I derive this from what I know to be true?" The re-derivation wasn't wasted effort. It was the mechanism by which he mapped each model's assumptions and boundary conditions — turning implicit wrongness into explicit, documented wrongness.
When a model's derivation revealed hidden assumptions — as his investigation of the Challenger O-ring failure did in 1986 — those assumptions became the focus. NASA's engineering models predicted the shuttle was safe. Feynman asked what those models assumed about temperature and material properties, found the assumption was untested below 53°F, and demonstrated its failure with a glass of ice water. He treated every model, including his own, as a provisional tool — to be used until it broke, examined when it broke, and replaced when something better was available.
Bezos's 2016 letter to shareholders contains the most practical business application of this principle in recent corporate history. He identified a specific failure mode: organisations develop processes (models) to serve customers, then gradually begin serving the process instead of the customer. The process becomes the territory. The metric becomes the reality.
His term for this was "managing to proxies." A customer satisfaction score is a proxy for actual satisfaction. A shipping speed target is a proxy for customer convenience. Each proxy is a model — a simplified, measurable representation of something complex and qualitative. Bezos warned that as organisations scale, leaders stop asking "is the customer happy?" and start asking "is the metric green?" — and the gap between those questions is where value destruction occurs.
The Wells Fargo cross-selling scandal, which erupted the same year Bezos wrote his letter, illustrated the failure mode precisely. The bank's model said accounts-per-customer correlated with profitability. The model was useful — until it became the target, at which point employees optimised for the model (open accounts) rather than the reality (serve customers). Three and a half million fake accounts later, the model's wrongness had become the company's defining crisis.
Bezos's antidote was structural: he insisted that Amazon leaders maintain direct contact with the underlying reality through mechanisms like reading customer complaint emails, requiring six-page narrative memos instead of PowerPoint slides (which compress nuance into bullet points — a model that loses critical information), and personally reviewing product pages. Each practice exists to prevent the proxy from replacing the thing it proxies. The discipline isn't intellectual. It's architectural — building systems that force regular contact between the model and the territory it represents.
Section 6
Visual Explanation
All Models Are Wrong — Every model simplifies reality. The danger isn't using wrong models; it's forgetting they're wrong. [Utility](/mental-models/utility) lives in the overlap. Risk lives in the gap.
Section 7
Connected Models
No model works in isolation, and a meta-model — a model about models — connects to every other framework in the collection. "All models are wrong" doesn't compete with other mental models. It comments on them. That positioning creates distinctive relationships — some reinforcing, some tensioned, some sequential — that sharpen both the meta-model and the frameworks it touches.
Here's how it connects to the broader lattice:
Reinforces
[Circle of Competence](/mental-models/circle-of-competence)
Circle of Competence says: know the boundaries of what you understand. "All models are wrong" explains why those boundaries exist — every model you've internalised has limits, and your competence extends only as far as your awareness of those limits. Munger's insistence on staying within his circle is operationally identical to Box's warning: the danger isn't in what you don't know, it's in what you think you know that isn't so.
The reinforcement is bidirectional. Circle of Competence tells you where to be cautious. "All models are wrong" tells you why caution is always warranted — even inside the circle. Warren Buffett's famous admission during the dot-com era — that he didn't invest in technology because he couldn't predict which companies would win — was simultaneously a Circle of Competence judgment and an "all models are wrong" judgment. His models for evaluating technology businesses were too wrong to be useful. Rather than pretend otherwise, he stayed out. The most sophisticated investors still get surprised, because even well-understood models have boundary conditions that shift.
Reinforces
Bayes' Theorem
Bayesian reasoning is the mathematical formalisation of "all models are wrong, update them." A Bayesian starts with a prior (a model of how likely something is), encounters new evidence, and updates the prior accordingly. The entire framework assumes your starting model is wrong and provides a rigorous method for making it less wrong over time.
Jim Simons's Renaissance Technologies runs on this logic: every model is a prior, every trade is an experiment, and every result updates the model. The fund's extraordinary returns come not from having correct priors but from updating faster and more systematically than the competition.
The power of the combination: "all models are wrong" tells you that your prior is wrong. Bayes' Theorem tells you how to make it less wrong with each new observation. Together, they turn a philosophical observation into an operational protocol — a systematic process for navigating a world where certainty is impossible but improvement is continuous.
Section 8
One Key Quote
"Essentially, all models are wrong, but some are useful."
— George E. P. Box, Empirical Model-Building and Response Surfaces (1987)
Section 9
Analyst's Take
Faster Than Normal — Editorial View
This is the model I wish every founder internalised before they internalised any other model. Not because it's the most actionable — it isn't — but because it changes the relationship you have with every other framework you'll ever use.
The failure mode I see most often is what I'd call model worship — a founder or investor finds a framework that explains their past success, and they promote it from "useful tool" to "fundamental truth." Once a model becomes identity, updating it feels like admitting failure rather than doing good epistemics. I've watched founders ride a go-to-market model into the ground because the model worked for their first product and they couldn't separate the model from their self-concept as a strategic thinker.
The flip side is equally dangerous. Model nihilism — "nothing is knowable, all frameworks are flawed, so I'll just go with my gut" — is the lazy misreading of Box's principle. The whole sentence matters: wrong and useful. The founders who dismiss frameworks entirely are no better calibrated than the ones who worship them. They're just less structured in their errors.
The correct operating posture is model pluralism with active decay detection. Use multiple models. Expect each one to be wrong. Build mechanisms to detect when a model's wrongness has crossed from tolerable approximation to dangerous distortion. Munger's latticework is model pluralism. Simons's systematic model rotation is active decay detection. Bezos's "Day 1" vigilance against proxy metrics is both.
There's a specific tell I watch for in founders. When someone describes their competitive advantage in terms of a single framework — "we have network effects" or "our moat is switching costs" — I ask: what does that framework miss? The answer reveals their relationship with the model. Founders who can list the limitations are using the framework as a tool. Founders who can't — or who bristle at the question — have fused with the framework. The fusion is the risk. Not because the framework is wrong (it may be quite useful), but because the founder has lost the ability to detect when it stops being useful. Markets shift. Customer behaviour evolves. Competitive dynamics mutate. The model stays still. The gap between a static model and a moving reality is where value destruction lives.
One pattern across the best operators: they hold strong opinions loosely. They'll commit fully to a model-based decision — invest the capital, ship the product, hire the team — while maintaining a background process that monitors the model's performance against reality. The commitment is operational. The looseness is epistemic. They act as if the model is right because indecision is costly, but they watch the model like it's wrong because overcommitment is fatal.
Section 10
Test Yourself
The gap between understanding this principle intellectually and applying it in practice is wider than most people realise. Everyone nods when they hear "all models are wrong." Almost nobody builds the systems to act on it.
These scenarios test whether you can distinguish between healthy model use, model worship, and model nihilism — and whether you can spot the moment a useful approximation becomes a dangerous one.
Is this mental model at work here?
Scenario 1
A quantitative trader builds a momentum model that generates consistent returns for three years. When the model produces a 15% drawdown in Q4, she investigates the market regime, identifies that interest rate volatility has changed the correlation structure her model assumed, and adjusts the parameters before re-deploying. She doesn't abandon the model or ignore the drawdown.
Scenario 2
A venture capitalist uses a standard framework: invest in companies with 3x year-over-year revenue growth, 120%+ net dollar retention, and gross margins above 70%. A founder presents a company with 1.5x growth and 95% retention, but a revolutionary supply chain technology. The VC passes because 'the numbers don't fit our model.'
Scenario 3
A CEO reads about mental models and concludes that since 'all models are wrong,' strategic planning is a waste of time. She cancels the annual strategy offsite, eliminates the competitive analysis function, and tells her team to 'just focus on execution and iterate.' Revenue declines 20% over the next year as competitors anticipate market shifts that her team missed.
Scenario 4
During the 2008 financial crisis, a risk manager at a major bank presents a report showing that the bank's Value-at-Risk model indicates maximum daily losses of $40 million with 99% confidence. The bank's actual daily losses exceed $200 million on three separate days that month. The risk manager defends the model, noting it was 'within specification for 99% of days.'
Section 11
Top Resources
The best resources on this principle come from practitioners who built models, watched them fail, and extracted the lessons — not from commentators who theorised about wrongness from a safe distance. Start with Box for the philosophical origin, then build depth with the practitioners who turned the principle into operational advantage.
The original source. Box's paper in the Journal of the American Statistical Association lays out the principle in its statistical context, with particular attention to the problem of overfitting. Short, technical, and foundational. The key passage — on the futility of seeking a "correct" model through excessive elaboration — has implications far beyond statistics.
Marks doesn't cite Box directly, but every chapter is an applied demonstration of the principle. His treatment of risk — particularly the argument that risk models based on past volatility are dangerously wrong because they confuse measurable variance with genuine uncertainty — is the best practical translation of "all models are wrong" for investors.
Zuckerman's biography of Jim Simons and Renaissance Technologies documents what happens when "all models are wrong" is treated as an engineering constraint rather than a philosophical observation. The fund's systematic approach to model rotation — building, testing, deploying, and retiring models in continuous cycles — is the most sophisticated institutional application of the principle in finance. The sections on the Medallion Fund's infrastructure for detecting model decay are particularly valuable.
Kuhn's argument that scientific progress occurs through paradigm shifts — where one imperfect model replaces another imperfect model that has accumulated too many anomalies — is the intellectual framework behind Box's principle. Understanding how models are adopted, defended, and eventually overthrown in science illuminates how the same process operates in business, investing, and strategy. The concept of "normal science" — puzzle-solving within a paradigm — explains why model worship is the default and model updating is the exception.
Feynman's Caltech commencement address is the most penetrating warning against model worship in the scientific literature. His distinction between genuine science (which maintains rigorous honesty about model limitations) and "cargo cult science" (which follows the form of scientific method without the substance) applies with equal force to business strategy, financial modelling, and any domain where the appearance of rigour substitutes for actual understanding. Twelve pages that will permanently change how you evaluate any model's claims to validity.
Tension
First Principles Thinking
First Principles Thinking seeks foundational truths — irreducible facts from which valid conclusions can be built upward. "All models are wrong" cautions that even those foundations are approximations. The tension is real: Musk's first principles decomposition of rocket costs assumed that raw material prices on commodity exchanges were "true" — but commodity prices are themselves models of supply-demand dynamics, subject to cartel manipulation, geopolitical disruption, and market structure changes.
The productive resolution: use first principles to build the best available model, then hold it with appropriate epistemic humility. The decomposition is valuable precisely because it produces a less wrong model than reasoning by analogy. But "less wrong" is not "right," and the founders who confuse the two end up overcommitting to a framework that will eventually need revision. The best thinkers use first principles to build and "all models are wrong" to hold — a combination that produces both conviction and adaptability.
Confirmation Bias is the cognitive force that makes "all models are wrong" so hard to apply. Once you've adopted a model, the brain selectively seeks evidence that confirms it and ignores evidence that contradicts it. The model ossifies into belief, and belief resists updating.
The tension operates in one direction: "all models are wrong" says update when evidence demands it. Confirmation bias says find evidence that doesn't demand it. Every model failure in the history of investing, science, and strategy involved this tension — someone had a useful model, confirmation bias prevented them from seeing its expiration date, and reality delivered the correction the model's holder refused to make voluntarily.
The antidote isn't willpower — willpower loses to cognitive bias almost every time. It's building systems — pre-mortems, red teams, devil's advocates, explicit falsification criteria — that institutionalise the challenge to prevailing models. Ray Dalio's "radical transparency" at Bridgewater, whatever its interpersonal costs, was designed to solve exactly this problem: make it structurally impossible for any model to go unchallenged.
Leads-to
[Margin of Safety](/mental-models/margin-of-safety)
If all models are wrong, every decision based on a model carries inherent uncertainty. Margin of Safety is the direct operational response: build enough buffer that your model's wrongness doesn't kill you.
Benjamin Graham's principle — buy at a price sufficiently below intrinsic value that even significant analytical errors leave you protected — is "all models are wrong" translated into portfolio construction. The margin isn't pessimism. It's engineering.
A bridge designed to hold exactly its expected load will collapse when reality exceeds the model. A bridge designed with a 2x safety factor will survive the model's inevitable wrongness. The same logic applies to cash reserves, hiring plans, delivery timelines, and every other domain where decisions rest on models of an uncertain future. The size of the margin should be proportional to the model's uncertainty — wider margins for models with more assumptions, narrower margins for models grounded in well-tested physical laws.
Leads-to
Inversion
Knowing your model is wrong naturally leads to the inverted question: how is it wrong? What are the specific assumptions that could fail? Which inputs are most uncertain? Where are the boundary conditions?
This is inversion applied reflexively — not to external problems but to your own analytical framework. Soros's reflexivity theory is essentially this move: he inverted standard market models by asking "what if the model itself changes the reality it's modelling?" The question only arises once you've accepted that the model is wrong.
From there, the path to inversion is direct: enumerate the wrongness, identify the highest-consequence errors, and either hedge against them or exploit the gap when others haven't noticed. The combination of "all models are wrong" (the diagnosis) and inversion (the investigation) produces a systematic process for turning model limitations from hidden vulnerabilities into visible, manageable risks.
The practical test: ask a founder "under what conditions would you abandon your current strategy?" If they can't answer specifically and immediately, they've reified their model. The strategy has become identity, and identity doesn't update gracefully. The best founders I work with can describe the falsification conditions for their own thesis in thirty seconds — not because they lack conviction, but because they've thought carefully about where their map might diverge from the territory.
This model gets more valuable the more successful you become. Early-stage founders hold models loosely because they have to — the feedback loops are fast and the data is scarce. The danger zone is post-product-market-fit, when a model that's been working starts feeling like truth. The companies that stagnate are almost always companies where a once-useful model hardened into doctrine. Nokia's smartphone strategy. Kodak's film business model. Blockbuster's retail footprint. In each case, the model worked — until it didn't — and the organisation couldn't update because the model had become the culture.
The actionable takeaway: for every model you rely on, maintain a written record of its assumptions and the conditions under which you'd abandon it. Not because you expect it to fail — you might not — but because the act of articulating the failure conditions keeps the model's provisional nature visible. The founders who do this navigate transitions cleanly. The founders who don't are the ones who wake up one morning in a market that their map no longer describes, wondering why the territory moved.
