ONE OF SEVEN WAYS OF SEEING
Borrow powerful explanations from the natural world.
Dual-process theories partition cognitive operations into two families: System 1, which is fast, associative, and intuitive, and System 2, which is slow, rule-governed, and deliberate. These processes are distinguished by speed, controllability, and the content they operate on.
Economies of scale occur when average cost falls as output increases, for example through bulk purchasing or spreading fixed costs across more units. The gains may diminish or reverse as coordination costs and resource constraints grow.
Multiple regression can identify associations, but adding control variables does not by itself establish causation. Nisbett explains how self-selection, unmeasured confounders, and measurement errors can mislead researchers. Well-designed randomized experiments offer stronger evidence for the effects they test; observational studies remain useful when experiments are impractical or unethical.
Empirical evidence comes from observation or experiment. Scientific studies and organizational records can both provide it; their value for a decision depends on how the observations were collected and what conclusions they support.
Entropy provides an analogy for maintenance and irreversible consequences: relationships can weaken through neglect, and some damage cannot be fully repaired. The social analogy does not make relationships obey a thermodynamic law.
Maintaining equipment, service quality, or a working process takes resources. A plan should account for upkeep as well as the effort of building or changing something.
Taleb contrasts economic equilibrium with the physical equilibrium of an inactive system. Living organisms maintain their organization through ongoing exchanges of energy and matter. He uses this contrast to argue that suppressing all variation can harm complex systems.
Ergodicity is the property that over time, randomness is eliminated, and the long-term average of a single system converges to the ensemble average. This concept challenges the reliability of past track records, as survivorship bias can mislead us into thinking past success predicts future performance.
When irreversible ruin is possible, the ensemble average (over many individuals at one time) diverges from the time average (over one individual over time). Without a logarithmic transformation, the time average inevitably leads to ruin, challenging standard decision theory and emphasizing the need for the precautionary principle.
Error-correcting codes allow reliable communication over noisy channels by adding redundancy. The Hamming code, discovered by Richard Hamming, uses a geometric structure based on the Fano plane to correct single-bit errors efficiently.
Limits give a precise meaning to convergent infinite sums and recurring decimals. The historical discussion connects debates about infinitesimals and Zeno’s paradox with the equality 0.999... = 1.
A valid implication makes a conclusion follow when its premise is true. Evidence can change how plausible a claim is without proving it, and selected confirming examples do not establish a universal statement.
Applying the principles of evolution by natural selection—such as faithful replication, positive mutation, and fitness—can help assess the success and longevity of systems you create, whether in business, technology, or personal habits.
Natural selection changes populations when heritable differences affect survival and reproduction. Cultural practices can also change through transmission, variation and differential spread, though cultural copying differs from genetic inheritance.
Natural selection changes the frequency of heritable traits across generations when those traits affect survival and reproduction. Human reward and aversion systems reflect this history, while experience also shapes what people learn to seek or avoid.
Expected value is the sum of possible outcomes weighted by their probabilities. A lottery example shows why a higher chance of winning a prize can still come with an expected financial loss.
Observational comparisons can suggest causes, while well-designed experiments can test them more directly. This chapter explains what random assignment and blinding contribute, where natural experiments fit, and why plausible interventions still need evaluation. Its historical examples show how education and social programs can disappoint or cause harm despite good intentions.
Randomized experiments can separate a treatment’s effects from differences between people who choose different treatments. Nisbett argues for using experiments where feasible; natural experiments can also help when their assignment process supports a credible comparison.
Exploratory data analysis looks for patterns and examines what remains unexplained. Banasiewicz recommends focusing that search on a decision, reviewing earlier findings and checking data quality to reduce the distraction of chance associations.
Exploratory analysis searches for patterns; confirmatory analysis tests hypotheses specified in advance. Both can use statistical models and visualizations. The distinction concerns how a question and its test were chosen, not a fixed division between tools.
Externalities are costs or benefits of an activity that spill over to others and are not reflected in its price. Taxes, subsidies, or other policies can bring private incentives closer to those wider effects.
The Fano plane is the smallest finite projective plane, consisting of seven points and seven lines, each line containing three points. It satisfies the axioms that every pair of points lies on exactly one line and every pair of lines intersects at exactly one point, demonstrating that geometry can be defined purely by axioms rather than physical appearance.
Clear feedback helps people learn from practice. Many decisions provide delayed or incomplete feedback, especially about alternatives that were never tried, so repeated experience may leave important gaps.
Feynman describes a practical sequence: propose a hypothesis, work out its consequences, and compare those predictions with experiment or observation. A mismatch calls for examining the hypothesis and test; agreement provides support for further inquiry. Curiosity and a willingness to accept unwelcome results help keep the process honest.
In fractal or power law distributions, statistical self-similarity means the degree of inequality remains constant across scales, unlike Gaussian distributions where it decreases. The exponent parameter critically determines the concentration of values in the tail, with small changes leading to dramatic differences in top shares.
A power-law tail follows P>x = K x^-α: multiplying the threshold by a given factor changes its exceedance probability by a fixed proportion. This scaling property links power laws to fractals, but identifying a power law in data and deciding where it applies remain separate problems.
Francis Galton discovered that scatterplots of parent and child heights form an elliptical pattern, leading to the concepts of correlation and regression to the mean, which quantify the strength of association between two variables.
For independent coin flips, the distribution of the suitably scaled total can approach a Gaussian bell curve as the number of flips grows. Smaller stakes can hold variance constant while the number of flips increases.
Kurt Gödel proved in 1931 that any consistent logical system rich enough to express arithmetic must contain statements that are true but cannot be proved or disproved within the system, a result that shattered the belief in a complete, logically perfect mathematical universe.
A periodic window is a parameter range in which a system otherwise capable of chaotic behavior settles into a repeating cycle. A period-3 window in the logistic map illustrates how deterministic models can exhibit both regular and chaotic behavior.
Taleb argues that uncertainty about model parameters can raise estimated tail probabilities through nonlinear effects. He uses uncertainty about Gaussian standard deviation to illustrate the sensitivity of rare-event estimates and the danger of treating parameters as known.
Taleb warns that statistical results can create false confidence when models understate extreme-event risks. His criticism centers on applying Gaussian assumptions outside the conditions where they are useful.
Bayesian reasoning provides rules for combining prior beliefs (base rates) with evidence diagnosticity to discipline intuition. It emphasizes that base rates matter and intuitive impressions of evidence are often exaggerated.
This approach compares alternative hypotheses against consistent and inconsistent evidence. It helps identify weaknesses in explanations and rank plausible scenarios, including when the evidence does not settle the question.
Test a decision when a useful experiment is feasible, even if the most senior person is confident about the answer. Randomly assigning alternatives can reveal which performs better, provided the comparison measures the intended outcome and accounts for uncertainty.
Using a crossword puzzle with two valid solutions, the author illustrates the philosophical concept of indeterminacy of translation, where multiple interpretations can fit the same evidence, but in practice, numerous constraints usually yield a single interpretation.
Mathematical induction proves a statement for positive whole numbers by establishing a starting case and a general step to the next case. Empirical induction instead extends observations to unobserved cases without a deductive guarantee. Cheng’s cookie examples illustrate why a plausible step or a chain of small differences is not enough.
Belief inertia is a metaphor for the persistence of established beliefs. It can explain both resistance to new evidence and the determination to continue through setbacks; the question is whether the belief remains responsive to what we learn.
Instant-runoff voting (IRV) can fail to elect the Condorcet winner—the candidate preferred by a majority over every other candidate—as demonstrated in the 2009 Burlington mayoral election. Additionally, IRV can exhibit monotonicity failure, where receiving more first-place votes can cause a candidate to lose.
Inter-Expert Reliability (IER) quantifies the consistency of responses among multiple experts, using a formula and a decision rule to determine if the level of agreement is statistically reliable.
Interdependence analysis measures the co-occurrence of variables, primarily through correlation, which quantifies the direction and strength of a linear association. It is distinct from causation and includes methods like Pearson and Spearman correlation coefficients, as well as dimensionality reduction techniques.
Kinetic theory explains bulk behavior through the motion and interactions of particles. The ideal-gas model treats particles as widely separated and collisions as elastic; liquids and solids require attention to closer packing and interactions.
The Law of Large Numbers states that as the number of repetitions of an experiment increases, the observed frequency of an outcome approaches its theoretical probability. It is a foundational theorem in probability theory, not a force that 'corrects' short-term deviations.
Thermodynamics describes energy transfers, thermal equilibrium, and the limits on converting heat into work. Energy is conserved, and the total entropy of an isolated system cannot decrease. Local order can increase when a system exchanges energy or matter with its surroundings.
Noise in judgments can be decomposed into level noise (systematic differences in average severity across judges) and pattern noise (idiosyncratic reactions to specific cases), both contributing to system noise.
A lever is a simple machine that amplifies force. Leverage uses this effect to lighten heavy loads and make work easier.
In chaotic dynamical systems, small differences in initial conditions can grow rapidly and limit precise long-term forecasts. Taleb uses billiards and Lorenz’s weather model to illustrate this sensitivity; it does not mean every dynamical system is chaotic or every kind of forecast is impossible.
Nothing can grow forever; at some point, the system will fight back and intervene to regulate exponential growth.
A literature review surveys research on a topic; a meta-analysis quantitatively combines results from multiple studies. The cited chess review found that more rigorous studies reported smaller cognitive benefits.
Taleb defines local fragility as sensitivity of harmful tail outcomes to changes in variability. A nonlinear response to stress can transmit that fragility from an exposure to its payoff. The effect of size depends on the response function, the tail distribution, and how exposures are combined.
The logarithm of a number is approximately the number of digits it has (the flogarithm), and it grows very slowly: the flogarithm of 1000 is 4, of 1 million is 7, of 1 billion is 10.
Taleb uses logarithmic transformations to examine strategies that can be sustained through time without ruin. A logarithm makes losses near zero wealth increasingly severe on the transformed scale. Separately, subexponential tail theory describes how rare large values can dominate extreme aggregate outcomes. The transformation alone does not establish ergodicity or guarantee growth.
Logical implication is a fundamental building block of logical arguments, where 'if A then B' means that if A is true, B must be true by logical necessity. This contrasts with everyday uses of 'if...then' that express causation, threats, promises, or personal taste.
When a study has low power for a small effect, results passing p < .05 can disproportionately reflect chance or exaggerated effect estimates. Low power alone does not determine whether a particular finding is false.
Skepticism evaluates a claim through questions, observation, research, and tests, while remaining open to revision. Harrison applies this approach to supernatural and paranormal claims, arguing that popularity and conviction cannot substitute for evidence.
Taleb presents Mandelbrot’s work on fat-tailed price movements as a challenge to Gaussian models of financial returns. The passage argues that extreme events require different modeling assumptions; its sweeping rejection of conventional tools is Taleb’s judgment.
Gaussian bell curve distributions have probabilities that drop rapidly away from the mean, whereas Mandelbrotian (scalable) variations lack this restriction, allowing for extreme events.
The maximum of an average cannot exceed the average of the components’ separate maxima. Because components can peak at different times, an index and its individual stocks can show different declines from their respective highs.
Mean squared error (MSE) is a measure of overall error that averages the squares of individual errors. It provides a way to weigh both bias and noise in forecasts.
Mean Squared Error (MSE) is a measure of overall error that squares individual errors, giving greater weight to large errors. It is the basis of Gauss's least squares method and can be decomposed into the sum of bias squared and noise squared.