Humans don't discount the future at a constant rate. The rational model assumes exponential discounting — such a person never reverses a decision. Real people discount hyperbolically: near delays are penalized steeply, distant ones barely at all. You fuss over "tomorrow vs. the day after," yet shrug at "one year vs. one year and a day" — even though both are a single day apart.
The shape of this curve produces a non-trivial consequence: preference reversal. From afar you firmly choose the "larger-but-later" reward (health, savings, deep work); but as the "smaller-but-sooner" temptation closes in, the curve spikes and you defect on the spot. Procrastination, addiction, and New Year's resolutions that collapse by June all stem from here — not weak will, but the fact that your present self and your long-term self are different decision-makers. This rhymes strikingly with the Buddhist no-continuous-self: there is no unchanging "I" deciding, but a relay of conflicting selves. In AI terms, treat your future selves as agents adversarial to the present one.
Key corollary: "I'll resist next time" always loses, because every local battle is won by the present self. What works is precommitment — let the farsighted self lock the impulsive self out of the choice, removing the option from the future weakling's hands. Next, reshape the inputs to discounting: pull the distant reward into the present and make it vivid; make the present cost visible.
Ulysses sailing past the Sirens: knowing his "future self" would lunge toward death once the song began, he had the crew bind him to the mast and seal their ears with wax — the canonical precommitment device. The rational, farsighted self stripped the impulsive self of its choice in advance. The behavioral-economics "save more tomorrow" plan is isomorphic: people won't cut today's paycheck to save, yet will commit to automatically saving part of every future raise — because for a distant date, present bias goes quiet.
① AI workflows: if an agent's reward only credits immediate visible output, it will "hyperbolically discount" away long-term goals (sacrificing correctness for a quick reply) — long-horizon tasks need explicit commitment constraints on future states, not "self-discipline." ② Deep work: fighting your phone with willpower fails; locking it in another room (precommitment) works. ③ Parenting: rather than nagging "homework first," co-design with the child a structure that can't be revoked in the moment (fixed slots, device custody). Don't slug it out with the impulsive present self — set the board in advance for the farsighted self.
Each additional unit adds less satisfaction than the last — the utility curve is concave. The first sip of water saves your life; the tenth glass does nothing. This plain curve is the common parent of several key results: ① risk aversion — because the curve is concave, a sure amount is worth more than a gamble of equal expected value (mathematically, E[u(x)] < u(E[x])); ② diversification — spreading resources across several declining curves yields more total utility than pushing one curve to saturation; ③ money's value is logarithmic — the same dollar is worth far more to the poor than to the rich, which is also why redistribution can raise total utility.
The non-trivial point: diminishing returns is the hidden structure behind the "optimal dose" of almost everything — information, practice, features, even a sense of meaning; past the knee, more is waste or even harm. The smart move isn't to maximize an input but to find the curve's knee, stop there, and redeploy the saved resources to another still-steep curve. It rhymes with neuroscience: the brain encodes deltas, not absolute levels (dopamine prediction error, hedonic adaptation), so any stable, unchanging good decays toward zero utility — the "hedonic treadmill." The lever off the treadmill is novelty and variety, which resets the declining curve back to its steep start.
The water-and-diamond paradox: water's total utility is enormous (you'd die without it), but because supply is abundant its marginal utility is near zero, so price is low; diamonds have low total utility yet are scarce, so high marginal utility and a high price. Price always reflects the margin, never the total — a distinction that dissolved the value paradox that plagued early economics, and a reminder never to confuse "how much it matters to me" with "what one more unit is worth."
① Learning allocation: the 5th hour on one topic yields far less than the 1st hour in a new field — cross-disciplinary transfer is essentially arbitraging diminishing returns, shifting attention from a saturated curve to a steep one. ② AI products: the 20th feature can carry negative marginal utility (complexity backlash); the same holds for compute — better to switch curves than pile on parameters. ③ Parenting: the third extracurricular often has negative marginal value; the child's idle, bored time is the steeper curve. Don't maximize a single input — balance at the margin across several declining curves.
Disruptors rarely beat the giant head-on. They enter at the low end or a brand-new market with a product that's worse on the metrics mainstream customers value, but cheaper, simpler, more convenient. The incumbent disdains that thin-margin segment and rationally pours resources into high-end, high-margin customers — which is precisely the trap. As the disruptor improves along its own trajectory, it eventually clears the mainstream's "good enough" threshold, and the high-end performance the incumbent spent years stacking becomes oversupply — suddenly unsellable.
The non-trivial point: what kills the incumbent isn't laziness or stupidity but correct resource allocation — listening to its best customers, chasing the highest margins; every step is "right," yet the sum is suicide. It's a structural trap (the innovator's dilemma), not a competence failure. Always distinguish disruptive from sustaining innovation: the vast majority of improvement is sustaining (making a good product better, favoring incumbents); "disruptive" is badly overused — real disruption always starts "worse, but cheap/simple enough to open a non-consumption market." It's isomorphic to evolution: a new species always invades from the margins of a niche, not the center; the incumbent is a local optimum trapped on a fitness peak, able to see the adjacent peak but unable to cross the valley between.
Practical test: watch two things — on which dimensions are you over-serving customers (performance beyond what they can absorb), and where is a "worse-but-cheaper" thing climbing its trajectory toward you. The first is your exposed belly; the second is the approaching set of teeth.
Steel mini-mills started at the lowest end — rebar: poor quality, thin margins — and integrated mills were happy to exit that nuisance segment and defend high-end sheet steel. The mini-mills used low-end profits to keep upgrading their process, eating angle iron and bar one rung at a time, finally storming the sheet-steel high ground and cornering the giants who had "rationally abandoned" the bottom. Digital cameras vs. film, streaming vs. disc rental follow the same script: every step of the incumbent's retreat is "reasonable," yet they sum to a rout.
① The main theme of the AI wave is disruptive: smaller/cheaper/open models start at the low end and, being "good enough and dirt cheap," devour the many tasks that high-end models over-serve. ② The "AI super-individual" is itself a disruptor — one person using cheap tools to deliver 80% of a consulting/dev output at 1% of the price, hollowing out over-serving firms from below. ③ Reverse self-check: in the value you provide, which part is "high-end performance far beyond what the other side actually needs"? That's the first thing a "good enough" version replaces. Don't ask "can I make it more refined?" — ask "where is a good-enough version closing in at one-tenth the cost?"
A system is ergodic if and only if its time average (one trajectory run all the way through time) equals its ensemble average (countless trajectories averaged at one instant). Physics often assumes ergodicity, but much of life and markets is non-ergodic — especially when the dynamics are multiplicative and there's an irreversible point of ruin. There, using "expected value" to decide for an individual living one path through time will systematically deceive you.
The non-trivial point (the classic coin game): each round, win → +50%, lose → −40%; the ensemble expectation is positive, so it looks worth playing — yet almost every individual trajectory drifts to zero over time, with a negative time-average growth rate. The ensemble mean is positive only because a tiny few who strike it rich drag the average up — and you only get to live one path. The right objective isn't the arithmetic expectation but the geometric (log) growth rate, which penalizes volatility and never compromises with ruin. This is the math behind the Kelly criterion, Taleb's "skin in the game," and every "avoid going to zero" instinct: once an absorbing barrier exists (ruin, death, reputation collapse), it cannot be "averaged away" by someone else's gain. Averaging over a population of parallel worlds you're not in — that's the subtlest error in a non-ergodic world.
Practical test: before deciding, ask — am I averaging over parallel worlds, or over my own time? Whenever there's an absorbing barrier on the path, however tempting the expected value, switch to the time average and put "survive first" above "maximize expectation." Any one-shot, non-repeatable, all-in bet should always be judged by the time logic.
Russian roulette: even if a single round's "expected payoff" is engineered to be positive, you can't keep playing — because "death" is an absorbing barrier; once hit, every positive expectation afterward is irrelevant to you. "On average the survivors made money" means nothing to the one lying on the floor. The multiplicative coin game is the gentle version of the same logic: the ensemble looks profitable on paper while the individual goes to zero over time. Expected value measures "the parallel-world versions of you" — and you live only one.
① Investing: size positions by the geometric growth rate, not the arithmetic expectation — bets must be small enough to "survive and compound to the next round," the core of Kelly and the margin of safety. ② Startups/career: a "positive-EV" bet that puts your whole net worth on the line and, if it fails, hits an absorbing barrier (no next round) should be declined under time logic. ③ AI deployment: a "99% safe" action repeated a thousand times almost certainly fails — risk accumulates over time, it's non-ergodic, and a single-shot expectation can't console you. ④ Parenting: some risks to a child are irreversible absorbing barriers and must be treated differently from ordinary "positive-EV" risks. First ask whether there's a zero point; if there is, survival always trumps maximizing expectation.