Nobody Lives the Average

You were right. It wasn’t. In 1884 Ludwig Boltzmann reached for a new word to describe a strange and convenient property of gases. A system is ergodic when time and the crowd agree — when following one molecule for a very long while tells you the same story as photographing every molecule at once. For a gas this is true, and it is the reason statistics works on gases: the average over the ensemble equals the average over the path. Nearly everything we built statistics on behaves this way. Then we aimed statistics at lives.

A life is not a gas. Ole Peters, the physicist who has spent two decades dragging this fact into economics, offers a coin. Heads multiplies your wealth by 1.5. Tails multiplies it by 0.6. Averaged across everyone who could possibly play, the game grows five percent per flip — a wonderful game, mathematically certified. Now follow one actual player, flip after flip, in order. Their wealth shrinks about five percent per flip. Play long enough and ruin is not a risk; it is the destination.

Both numbers are honest. They measure different worlds. The plus-five belongs to a million simultaneous players pooling their winnings — the lucky few whose heads-streaks grow so astronomically large that they drag the average up for everybody. The minus-five belongs to anyone who plays alone, in sequence, with no pooling. You are not a million of yourself. Nobody pools.

The ensemble average is a Reality without an Actualizer — true of everyone at once and no one in particular.

The Reality Equation puts it in four words: one Reality per Actualizer. A Reality is something an entity has — a person, a company, a nation — along its own path, at an instant. The expected value of Peters’s coin is a number with nobody home. It describes no player who has ever existed or will. It is a lovely, rigorous, fully verified description of a place where nobody lives.

Figure 01

Fifty flips of the same coin

All players, pooled (the ensemble average) ×11.5

The typical player (the median path) ×0.07

The machine speaks ensemble

Now notice what a machine-learning model is. It is an ensemble engine — that is not an accusation, it is the definition. It was trained across millions of recorded paths, and what it hands you is a distilled crowd: the thing that worked on average for people whose situations rhymed with yours. For most questions this is exactly what you want, because most questions repeat. Where to eat, how to phrase the email, which wire goes to ground — small stakes, run a thousand times, effectively ergodic. The crowd’s answer and your answer converge. Take the ensemble’s advice all day.

The danger lives in a small set of questions that do not repeat — the ones with a branch you do not come back from. Broke. Dead. Burned out. Struck off. Mathematicians call these absorbing states, and their defining property is that you stop while the average keeps going. The ensemble does not fear absorbing states, because the ensemble survives them; it always has other members left to average. You do not have other members. Advice that is optimal for the crowd can be ruinous for the path, and the machine, speaking fluent ensemble, will deliver it with perfect confidence to one person at a time, at scale.

The mathematics of taking your own side has existed since 1956, when John L. Kelly Jr., a physicist at Bell Labs, wrote down the rule for betting a single path rather than a crowd: size every bet to maximize the growth of the one bankroll you actually have. Kelly’s rule always bets less than expected value invites — the gap between the two bet sizes is the price of being singular. Retirement planners know a corner of the same law as sequence-of-returns risk: two retirees with identical average returns and identical withdrawals can end in opposite places, one rich and one broke, purely because of the order in which the returns arrived. The averages matched. The lives did not.

What I think happens next

1. “Personal” will come to mean path-aware.

   Assistants split into ensemble engines, which answer with the crowd’s average, and path engines, which know your sequence. The second kind does not need to know your favorite color. It needs to know your reserves, your obligations, and what you cannot come back from — and to quote you your time average, not the crowd’s mean.

2. “Good on average” becomes the start of a question, not the end of one.

   People will learn the reflex: average over whom, pooled how, arriving in what order? Within a few years, a pitch that leads with expected value will be heard roughly the way “trust me” is heard now.

3. Institutions will keep speaking ensemble, and the gap will turn political.

   Records support averages and institutions run on records, so the productivity numbers will be honestly up while particular careers are honestly zeroed — both true, measuring different worlds. The distance between the quoted average and the lived sequence becomes the defining political fact of the AI decade.

4. Ruin-proofing will outsell optimization.

   The premium product of the agent era is not “most growth”; it is “no zeros.” The first personal AI that credibly promises never broke rather than most rich wins the household, for the same reason Kelly bettors outlive braver ones.

5. Order becomes a discipline.

   The same risks taken in a different order are a different life — sequence-of-returns risk generalized to everything. Planning shifts from choosing what to choosing when, and “too early” and “too late” become computed quantities instead of regrets.

One path, with my name on it

My first thought is to delegate the whole outcome — I wrote that three days ago and I meant it. But watch what I delegate to. A rented intelligence is calibrated on everyone, which is another way of saying it is calibrated on no one; that is why I keep insisting the stack be yours. Charleston AI is one company, in one city, with my name on it. There is no ensemble of me to average over, no other members to absorb my absorbing states. When my AI advises me, I want it betting my sequence — because I am the only one running it.

Already true

Peters’s coin and its two growth rates are published record, not metaphor. Kelly’s rule has priced single paths since 1956. Sequence-of-returns risk sits in every planner’s textbook. And every large model is trained across an ensemble of recorded paths — by definition, not by accusation.

What has to happen

Personal AI has to hold enough standing context — reserves, obligations, the handful of irreversibles — to actually compute a path, and people have to route their consequential, non-repeating decisions through it. If personal AI stalls as a chat window with no memory of your sequence, path engines stay a niche for traders and actuaries.

Where I am probably wrong

Most decisions repeat, carry no ruin branch, and are effectively ergodic — for those, the crowd’s answer is simply correct, and this piece changes nothing. And lives are less lonely than the arithmetic: families, insurance, partnerships, and friendship are all pooling devices — ways of lending the ensemble a body. If pooling expands as fast as machine advice does, the average acquires an address, and I have overstated the solitude of the path.

The place on the chart

The advisor’s chart was not wrong. It was a true picture of a real place — the place where all the paths are pooled, every loss is somebody else’s gain to average against, and nothing is ever final. Up and to the right, smooth as glass. You were right not to trust it with your one, ordered, unrepeatable life.

Nobody lives there.