Most people assume "positive expected value means bet, and the bigger the EV the more you bet." The Kelly Criterion says: in repeated bets, what you should maximize isn't single-shot arithmetic expectation but the long-run geometric growth rate of wealth — the growth rate of log(wealth). The difference is fatal: full-betting a positive-EV wager, as long as there's any chance of going to zero, leads almost surely to ruin over repetition — because ruin is an absorbing state. In a multiplicative world, one zero eats all prior gains.
Non-trivial: (1) Kelly's optimal fraction = edge ÷ odds — essentially a tax on volatility: the larger your edge, the more you bet; the more dispersed the odds (higher variance), the less. (2) Almost everyone should use half-Kelly: you systematically overestimate your edge, and overbetting is punished asymmetrically — slight underbetting costs a sliver of growth, while slight overbetting wrecks long-run growth and can turn a positive-EV bet into negative growth. Half-Kelly is a margin of safety. (3) Kelly is the same object as Shannon's information rate: the optimal growth rate is exactly your "information rate" relative to the market — your edge is your information. With no real informational advantage, the optimal stake is zero.
Give someone a biased coin landing heads 60% of the time, with free betting over 300 flips. Kelly says bet about 20% of current bankroll each time. In experiments, most educated players either occasionally go all-in (one unlucky flip wipes them out) or bet random small amounts — and badly underperform Kelly, despite knowing the coin favors them, because they don't grasp that bet sizing is the key variable. Right call, wrong size, still a loss.
Betting your entire savings and all spare time on a single AI startup idea — even with positive EV — violates Kelly: ruin risk collapses your life's geometric growth. The correct move is to size by edge: allocate more where you genuinely have an information advantage (AI, distributed systems), less or nothing where you're merely following the crowd, and always keep a buffer that no single failure can zero out. Career, health, parenting are all repeated games; going all-in on any one is anti-Kelly.
The essence of convex betting isn't "predicting right" but arranging a payoff shape: losses have a clear ceiling (capped), gains have none (open-ended). Such a position is convex — by Jensen's inequality, uncertainty itself creates value for you: the wilder the volatility, the higher the convex position's expected payoff. You don't need to know which bet pays off, only that every bet is "affordable to lose, unbounded to win," letting time and volatility do the sorting.
Non-trivial: (1) The opposite of convex is concave — many seemingly robust strategies are actually concave: capped gains, bottomless disaster, like "picking up pennies in front of a steamroller," earning small money steadily until one black swan takes it all. Knowing whether your position is convex or concave matters more than predicting the market. (2) The barbell strategy: split resources into two ends — the vast majority ultra-safe (survival), a small slice on extreme convex bets (chasing the fat-tailed upside), deliberately avoiding the "false-safe" concave middle. (3) Isomorphic to biological evolution: mutation is a flood of convex bets — most vanish harmlessly (loss capped at the individual), occasionally one yields a huge fitness leap (upside open). Complex systems advance precisely through cheap convex trial-and-error.
A venture portfolio: seven or eight of ten investments go to zero (each loss capped at the principal), but a single 100× return wins the whole fund. VC never earns the "average" — it earns the fat tail. Options are a natural convex instrument too: the buyer can lose at most the premium, while the upside scales with the underlying's surge. Convex strategies' returns are carried by a tiny number of extreme positive events, not by the average of the many.
The "AI super-individual" is essentially a portfolio of cheap convex bets: write public essays, open-source small tools, run small experiments, proactively forge weak ties — each with a downside of merely "a few wasted hours" and an upside of "possibly seen by millions, drawing unexpected connections or opportunities." A stable but capped job is concave; what actually amplifies you is a fat-tailed breakout from one of these cheap bets. Parenting is the same: let a child explore broadly and cheaply rather than going all-in early on one "success path" — you're betting on the convexity of exploration, not on any single prediction.
Tail-risk hedging is the mirror image of convex betting: you pay a small, long-run negative-EV premium to buy a convex payout that triggers when a rare but devastating event hits. Its purpose isn't profit but ensuring you're never zeroed out by a single catastrophe — only by staying at the table do you get to enjoy compounding.
Non-trivial: (1) The core is ergodicity: a positive arithmetic group-average return does not mean a single individual survives over time. Because ruin is an irreversible absorbing state, an individual who will inevitably meet the tail over time gets slowly pushed toward certain ruin by a bet whose "group expectation is positive." So paying for tail protection isn't irrational timidity — it's the correct response to irreversibility. (2) You must distinguish recoverable losses from unrecoverable (ruin) losses — buy protection only for the latter; boldly bear volatility on the former. Spending premium on recoverable small swings is burning money. (3) It's psychologically brutal: people's aversion to "steady small premium losses" far exceeds their fear of "rare catastrophe," so they cancel coverage right before the crash. Only those who can stomach small losses through 99% of calm can catch the convex payout in the 1%.
A dedicated tail-hedging strategy: small premium losses year after year in calm times, looking like "constantly wasting money"; yet when a crash like 2008 or 2020 hits, a single payout can reach dozens of times the stake — enough to carry all those years of premiums and still profit hugely. A plainer version is redundancy and backups — you pay storage for copies you'll "almost never need," precisely to insure against the one irreversible disk-failure catastrophe.
Disaster recovery in distributed systems is engineered tail hedging: replicas, circuit breakers, rate limiting, graceful degradation — a "redundant waste" in normal times that prevents the whole site from going to zero during failure. At life scale: a cash buffer, portable cross-domain skills, and health habits are premiums against the irreversible tails of "job loss, industry disruption, illness." The right question order is: first ask "what could take me out of the game entirely," then insure that — rather than fixating on daily small swings. In the AI era, continuous learning is itself a premium against your current skills being disrupted in one stroke.
Insurance thinking collapses the previous three models into one decision rule: the right use of insurance is to transfer risks that "could take you out," not to transfer small risks you can "calculate and afford." Most people buy it exactly backwards — paying premiums on phone-screen cracks and appliance warranties (usually negative-EV and affordable, which should be self-insured) while going bare on serious illness, liability, or income loss (low-probability, ruinous, which should be transferred).
Non-trivial: (1) Insurance is "negative-EV but positive-utility" — because utility is concave in wealth: losing a fortune hurts far more than gaining one delights, so spending a little to shave off the huge pain in the left tail is rational. This is the same object as Kelly: both protect geometric growth and dodge the absorbing state. (2) Self-insuring requires a buffer: your reserves are your own insurance company. Without a buffer you can only transfer the unbearable risk out. (3) The reverse trap is moral hazard — once insured, people unconsciously take more risk; the hedge itself changes the system, so you need "skin in the game" to realign incentives. (4) It echoes the Buddhist notion of impermanence: buying insurance is acknowledging the uncontrollable and reserving slack for the unexpected, not fantasizing about total control — a form of humility, not fear.
A household buys every phone plan, screen protection, and appliance extended warranty, bleeding small money yearly — yet lacks adequate critical-illness and family liability coverage. This is the textbook "pay for small certainties, go bare on large uncertainties." A correct insurance budget should flow almost entirely toward the few low-frequency catastrophes that could actually break the family's finances, leaving small recoverable losses to be absorbed by your own reserves. Spend premiums on the cutting edge (the left tail), not on peace of mind.
Draw a risk quadrant for yourself and your family: place each possible loss by "frequency × severity." Top-right (low-frequency, family-ending losses: serious illness, prolonged income loss, major liability) → insure; bottom-left (high-frequency but affordable annoyances: gadget wear) → self-insure. Technical people especially tend to get this backwards — over-insuring small risks while being utterly unprepared for the real tails. First build an "own insurance company" cash buffer, then transfer out the few tails that the buffer can't absorb.