1. Bayesian Thinking
Calibrate your beliefs with new evidence
In Depth
The core mechanism of Bayesian thinking is simple: beliefs aren't 0/1 verdicts but probability distributions. When new evidence arrives, you don't throw out the old judgment — you update it according to posterior ∝ prior × likelihood. The prior is what you believe today; the likelihood asks "if the hypothesis were true, how likely is this evidence?" Decisions stop being "right vs wrong" and become "how likely to be right."
The non-obvious insight: the more extreme your prior, the harder it is to shift with a single data point; the more vague your prior, the more easily noise hijacks it. Real experts aren't "neutral" — they hold priors with reasonable strength and are highly sensitive to disconfirming evidence. Another counterintuitive point: absence of evidence is itself evidence. What you don't see updates your belief just as much as what you do.
How to practice it: before a decision, write down three things explicitly — (1) your prior probability as a percentage, (2) what kinds of evidence should move it up or down, and (3) by how much. After the fact, audit how you actually updated against your intuition, and look for systematic biases.
Classic example: A medical test for a rare disease (base rate 0.1%) returns positive, with 99% test accuracy. Intuition says the patient is 99% likely to have the disease. Bayes says under 10% — because the prior is so low that false positives swamp true positives.
BigCat scenario: You're evaluating an early-stage AI startup. Prior: comparable projects have a 15% five-year survival rate. But the founding team proactively presents counter-data three times ("here are the users who don't come back") — behavior that's extremely unlikely for a low-quality team (huge likelihood ratio). The posterior should jump to 35-40%.
Your kid says "I'm bad at math." The prior is one recent disappointing unit test (70%). But the longer-run evidence — voluntary daily problem sets, clear solution paths — is poorly compatible with "bad at math." Resist the emotional drift of the prior. Recalibrate "I'm bad" into "I'm 80% likely stuck on one specific sub-concept in this chapter" — a solvable problem.
English Summary
Bayesian Thinking treats beliefs as probability distributions, not binary verdicts. Each new piece of evidence — including its absence — multiplies your prior by a likelihood ratio to yield an updated posterior. Strong priors resist single data points; vague priors get hijacked by noise. The discipline: write priors explicitly, specify what evidence would shift them, and reconcile actual updates against intuition.
AI Prompts
English Template
Apply Bayesian reasoning to my decision: [decision].
My current prior is [X%], based on [reasoning].
New evidence: [evidence].
Please: 1) Estimate the likelihood ratio; 2) Compute the updated posterior; 3) Flag disconfirming evidence I may have overlooked; 4) List 3 future observations that would most efficiently update this belief.
2. Black Swan
The "impossible" events are the ones that decide the outcome
In Depth
Black Swan events have three signatures: extreme rarity (outside the known sample), massive impact (they set the trajectory of the system), and retrospective rationalization ("we should have seen it coming"). The mechanism is that the world isn't normally distributed — it's often a power law. In a power-law world, averages lie: a tiny number of extreme events accounts for most of the total.
The counterintuitive insight: risk management shouldn't focus on "what's most likely to happen" but on "if it happens, will I be wiped out?" This is Taleb's ergodicity problem — a positive-EV bet that includes a single ruinous outcome guarantees long-run failure. Another widely missed point: Black Swans cut both ways. Positive Black Swans (a 100x return on one bet) exist too; the question is whether your structure can catch them.
How to practice it: (1) Replace probability forecasts with stress tests — don't ask whether it will happen, ask whether you'd survive if it did. (2) Hold a barbell portfolio: extreme safety on one end, bounded high-risk exposure on the other, nothing in the middle. (3) Maintain a "never" list — no single variable should ever be able to wipe out 80% of your assets, energy, or health.
Classic example: Before the 2008 subprime crisis, every bank model put the probability of a nationwide 20%+ housing decline at near zero — it had never happened in the historical window. But the models used historical data, not structural fragility. When it did happen, the global financial system nearly collapsed.
BigCat scenario: ChatGPT's late-2022 release was a textbook positive Black Swan for knowledge workers. Beforehand, most career models assumed "AI can only handle narrow tasks." Afterward, everyone said "we should have seen the Transformer's potential" — but very few actually positioned for it. Whether you catch a positive Black Swan depends on whether you've built scalable leverage structures in advance: personal content assets, cross-domain learning habits, technical literacy.
When planning your child's education, the traditional path assumes "the career landscape will be roughly stable for the next 15 years." But the pace of AI iteration is itself a Black Swan source. Rather than over-betting on one specific skill, bet on antifragile capabilities: meta-learning, the ability to ask good questions, judgment in collaborating with AI, psychological resilience. This is making unpredictability part of the educational strategy.
English Summary
Black Swans are rare, high-impact, retrospectively-rationalized events that dominate outcomes in power-law systems. Forecasting their probability is futile; engineering survivability is not. Apply barbell strategies — extreme safety paired with bounded high-risk exposure — and maintain a "never" list that prevents any single variable from inflicting irreversible loss. Position structurally for positive Black Swans as well: optionality compounds.
AI Prompts
English Template
Stress-test the following plan for Black Swan exposure: [plan].
Please provide: 1) Three plausible negative tail scenarios with cascading effects; 2) Three positive tail scenarios I could structurally capture; 3) Single points of failure that could cause ruin; 4) A barbell-style restructuring to reduce downside while preserving upside optionality.
3. Antifragility
Not just resilient — actually getting stronger under stress
In Depth
Taleb distinguishes three states: fragile (breaks under stress), robust (unchanged by stress), and antifragile (strengthened by stress). The mathematical signature of an antifragile system is a convex response to volatility — the upside slope exceeds the downside slope. Biological evolution, muscle training, startup ecosystems, immune systems — all are textbook antifragile.
The counterintuitive insight: over-protection manufactures fragility. Sterile environments weaken immunity, zero-failure schooling destroys resilience, suppressed market volatility accumulates systemic risk. "Stability" is often fragility in disguise — it compounds small risks into large ones. A second insight: antifragility isn't about avoiding the hit — it's about turning every hit into information and adaptation.
How to practice it: (1) Cap the downside (stop-losses, insurance, exit routes) while leaving the upside uncapped. (2) Deliberately expose yourself to small, recoverable doses of stress. (3) Replace one large bet with many small, low-correlation bets. (4) Review every "hit" for what new information it revealed — let the system learn from disorder.
Classic example: Muscle training. Microscopic tears (stress) trigger supercompensation, and the muscle returns stronger. No exercise (avoiding stress) leads to atrophy; one-shot overload (unrecoverable damage) leads to injury. The key to antifragility: controlled dose plus full recovery.
BigCat scenario: Design your "AI super-individual" learning system as antifragile — spend 20% of each week deliberately testing new tools, prompts, or workflows that have a 50% failure rate. The 10% that fail is cheap tuition; among the 10% that succeed, one or two will be transformative (convex payoff). Sticking only to tools you already know means your capability atrophies relative to a fast-moving frontier — that's hidden fragility.
When your kid fails a test, the fragile response is anxiety, more tutoring, easier material. The antifragile response is to turn the wrong answers into specific diagnostic information — which underlying concept did this expose? Reframe every "failure" as a free diagnostic scan. Over time, a child who extracts information from volatility builds a steeper capability curve than one whose scores never wobbled.
English Summary
Antifragility is a step beyond robustness: systems that gain from disorder, volatility, and stressors via convex payoffs. Overprotection — sterile environments, failure-free trajectories, suppressed volatility — manufactures hidden fragility. Engineer antifragility by capping downside, exposing yourself to small recoverable stressors, diversifying into many low-correlation bets, and treating every shock as a free diagnostic signal.
AI Prompts
English Template
Apply the antifragility framework to my [system/project/habit]: [description]
Analyze: 1) Is it fragile, robust, or antifragile, and why? 2) Which hidden over-protections are accumulating long-term fragility? 3) What small, recoverable stressors could introduce convexity? 4) Provide a 30-day antifragility retrofit plan with measurable signals.
4. Expected Value Thinking
The geometric center of decision-making
In Depth
The basic formula: EV = Σ (probability × payoff). The core idea is to decouple a single outcome from the quality of the decision behind it — a decision can be correct (positive EV) but the outcome can still be a loss, and a decision can be wrong (negative EV) but still pay off through luck. Only EV compounds in the long run.
The non-obvious insight: real-world EV is asymmetric. Many opportunities have bounded downside (the cost of trying a new tool is a few hours) and potentially enormous upside (it reshapes your whole workflow). These asymmetric setups are the expert's hunting ground. A second counterintuitive point: when sample size is small, EV thinking yields to the ruin boundary (Kelly criterion) — even a positive-EV bet, sized too large in a single shot, can still destroy you.
How to practice it: (1) For every important decision, list 3-5 possible outcomes with their probabilities and values. (2) Separate repeatable decisions (use EV) from one-shot irreversible ones (use maximum tolerable loss). (3) Actively hunt for bounded-downside, asymmetric-upside opportunities. (4) Audit past decisions and separate decision quality from outcome luck.
Classic example: Flip a fair coin: heads wins $200, tails loses $100. EV = 0.5×200 + 0.5×(-100) = +$50. Even if you lose $100 on this round, repeating the bet long-term is profitable. The mistake is abandoning a positive-EV game over one bad outcome.
BigCat scenario: Should you spend 4 hours learning a new AI framework? Outcome distribution: 60% marginally useful (saves 5 hrs/month), 30% useless (lose 4 hrs), 10% extremely useful (reshapes a workflow, saves 50+ hrs/month). EV ≈ 0.6×5 + 0.3×(-4) + 0.1×50 ≈ +6.8 hrs/month — a classic bounded-downside, asymmetric-upside decision.
When choosing your child's extracurriculars, instead of picking only the "most certainly useful" activities, build an EV portfolio: one high-certainty pick (e.g. academic reinforcement) + one mid-certainty pick (e.g. public speaking) + one low-probability, high-payoff pick (e.g. robotics, coding, creative writing). The third category will usually be just an experience — but with a 10% chance of igniting a lifelong passion, whose expected value dwarfs the other two combined.
English Summary
Expected Value Thinking decouples decision quality from single outcomes: EV = Σ (probability × payoff). Hunt for asymmetric opportunities where downside is bounded but upside is uncapped. Use EV for repeatable decisions, but switch to ruin-avoidance (Kelly-style sizing) for irreversible ones. Audit decisions by their EV at the time of choice, not by whether luck happened to land in your favor.
AI Prompts
English Template
Run an expected-value analysis on this decision: [decision]
Please: 1) Enumerate 3-5 possible outcomes with probability and payoff; 2) Compute the EV and flag the most uncertain input; 3) Assess whether the structure is asymmetric (bounded downside, uncapped upside); 4) If the decision is one-shot/irreversible, re-evaluate under a ruin-avoidance lens and recommend a sizing/exposure level.