Antifragility & risk #

Notes on convexity, randomness, and living in a world of thick tails — where the rare event is not the exception to the story but the whole of it.

Risk is not probability times consequence #

We are taught to measure risk by multiplying probability by consequence. That works while the losses are survivable, and it fails exactly where it matters. A chance of one in ten thousand is small only if you are still there afterwards to collect the average: if the consequence is ruin — an absorbing barrier, the kind you never come back from — then no probability makes the bet acceptable, because there is no afterwards in which the good outcomes can compensate. This is the thick-tailed domain, where events are not independent and the tail carries the meaning.

Think of Russian roulette played with a revolver of a hundred million chambers. Pull once, for a large enough prize, and you would be right to take it. The trap is repetition: play every morning and ruin stops being unlikely and becomes only a matter of time. The arithmetic of the average is fine; it simply describes a man who is no longer alive.

Via negativa #

Concentrate on what a thing is not, rather than on what it is. Not everything can be said or explained, and the definition that works is usually indirect. Language has its limits. Logic has its own, and expecting it to settle every question is naive.

In practice, in a complex system, it is a matter of what not to do rather than what to do. Less is more, and things can always be added later; removing rarely harms a complex system. Remove what is fragile and you are left with the robust.

The deeper reason is asymmetry of surface area. Elimination is final: once a thing is shown to be false it stays false. Confirmation is never final, because we cannot know whether the counterexample is merely one we have not met yet. This is Popper’s asymmetry, and the ancient sceptics had most of it before him.

Skin in the game is a filter #

Skin in the game is a filter. Not because we learn nothing from our mistakes — a burnt hand teaches a child something permanent about stoves — but because what we learn that way is narrow and local. The child acquires a reflex, not a theory of heat, and cannot carry it across to the next domain. So the useful knowledge accumulates somewhere other than in the individual who paid for it.

We get eliminated from the environments we don’t fit. By taking risks we find, by trial and error, where we do fit — like a blind man guided by a white stick. And it is the system that learns, not us: it selects those who did not make certain errors, whether or not any of them understood why. Bad drivers end up in the cemetery, and the roads get safer without a single bad driver having improved.

Which is the argument against centralization. A decision maker who cannot be wiped out by his own mistake removes himself from the filter, and the error stays in the system and compounds. The individual survives the mistake; the civilization pays for it.

When being wrong costs nothing #

Being right or wrong is not what counts. If being wrong costs you nothing, then it doesn’t count — it was a free experiment, a conjecture put up to be refuted. What counts is what the error costs, and to whom.

Diminishing returns #

The law of diminishing returns has something to do with convexity — but what exactly is the link? It may come down to size. Every complex system has a range of sizes at which its organization works, and passing out of that range does not merely make it bigger: it changes what the thing is. A government that grows, a cell colony that becomes a cancer, a slope of snow that becomes an avalanche.

And a complex system is not the sum of its parts but an entity of its own. A human being is not a heap of cells: no quantity of cells piled together, however alive, will produce a work of art. Whatever AI turns out to be, it will be that kind of entity too — not a large calculator, but a thing with properties none of its parts possess.

Why we don’t learn from history #

First, history is written by the victors, and like human memory it is a story continuously overwritten by the present.

Second, we are domain-blind. Humans transfer knowledge from one domain to another very badly; a man can be acute in area A and a fool in area B, and the same man will not notice the change. So it is a mistake to treat narrow expertise as expertise in general — the lesson of the last war is learned by the generals, and applied to the wrong one.

Third, expertise itself has an expiry date. Knowledge advances by conjecture and refutation, which means today’s competence is a set of conjectures that have not yet been refuted. To think of oneself as an expert is to forget that time keeps running.

Black swan as a black mammoth #

What if the black swan were called a black mammoth — a giant black dinosaur that could swallow you like a fly? It would be far easier to grasp the weight of these random events. And a positive one — why not a goose that lays golden eggs?

Understanding the unknown from the known #

How can we come to understand the properties of the unknown (the infinite) from the known (the finite, the past)? That is: how can we predict the future?

Alan Kay’s answer — the best way to predict the future is to invent it — is usually quoted as optimism. Read on this page it is something bleaker and more useful: not a promise that you can foresee the future, but an admission that you cannot, and that acting is the only remaining way to have any purchase on it. You do not get to know what comes. You get to choose what you expose yourself to.

Random, with rules that fit only for now #

It is easier to think of nature as wholly random, with rules that fit only temporarily (the turkey), than to think of hard rules plus the occasional black swan.

The turkey thinks it will live forever #

The hand that feeds you may be the one that wrings your neck — the last words of the turkey.

It believes it will live forever, and every day of its life is evidence for the belief. But nothing lasts forever; such is the order of things. And we get stuck in our habits the same way, like a needle in the groove of a record, mistaking a thousand quiet days for a law of nature. The hard question is not how to get unstuck. It is how to know that one is stuck at all — since from inside the groove the music sounds continuous.

Entropy wants you to be ’normal' #

There is nothing to understand in the phenomenon of the black swan. The world is random and temporarily in order — like plants, animals and humans: something in order for a while, each with an expiry date. Order is the expensive local exception; disorder is what everything relaxes back into the moment the effort stops. That is the sense in which entropy wants you to be ordinary: being a rare and organized arrangement of atoms takes continuous work, and the work always stops eventually.

There are even physicists who suspect the constants themselves drift. And we may be living in a multiverse, or in a simulation — conjectures I have no way to test, and mention only to mark how far down the uncertainty may go.

A dupe’s problem #

The black swan, or the turkey, is a dupe’s problem. The butcher knows what will happen to the turkey; the turkey does not. That is exactly why it is a turkey — because it is a dupe.

Slow to build, quick to destroy #

Positive black swans accumulate slowly; the negative ones arrive all at once. The asymmetry is not psychological but physical: there are vastly more ways for a thing to be broken than to be whole, so chance alone favours destruction. Every organized thing in nature — a cathedral, a living body — is a temporary arrangement held against that current, and it will be handed back atom by atom to the disorder it was borrowed from. To take the organized things for the permanent ones is to read the exception as the rule.