scale & measure · metaphor 100 of 100
the last of one hundred

Roughly right

Faced with a question too big to know — how many, how much, how likely — most people freeze into "there's no way to tell." Fermi's gift is the confidence to be roughly right rather than precisely ignorant: decompose the unknowable into knowable pieces, and errors in opposite directions cancel into an answer good enough to act on.

How many piano tuners are there in Chicago? It is the classic that terrifies the precise and delights the resourceful. You cannot know it — there is no register in your head, no fact to recall. And yet you can build it. Roughly how many people live in Chicago; roughly what fraction own a piano; roughly how often a piano gets tuned; roughly how many tunings fill a tuner's year. Each factor is a guess you would stake only a little on. Multiply them and out falls a number — and that number lands within a factor of a few of the truth.

This is the core competence of acting under uncertainty, and its deepest lesson is emotional as much as mathematical. The paralysis of "I can't possibly know" is almost never true; it is a failure to decompose. The moment you break the unanswerable question into a handful of things you can put honest brackets around, the fog lifts, and you are left holding not certainty but something far more useful: a defensible order of magnitude, with its humility built into its error bars.

01 · the instrument

The estimation builder

Pick a question. Give each factor a low and a high guess — a range, not a false point. The builder samples every factor honestly (log-uniform between your brackets), multiplies thousands of times over, and hands back not one number but a distribution: a median, an honest range, and an order of magnitude. Where the true answer is known, it shows how close you landed.

choose a question

the factors — low & high
order of magnitude
your estimate known truth bars: 6,000 Monte-Carlo runs, log scale
where the uncertainty goes — each factor's spread, and how they combine
What you're watching
Each factor's bracket is a confession of ignorance. Yet the independent total spread is visibly narrower than the reinforced one — because when guesses are honest and independent, some run high while others run low, and in log space those overshoots and undershoots partly cancel. Flip correlated errors on and the cancellation dies: every guess now leans the same way, the distribution balloons, and the estimate degrades. Honest guessing beats motivated guessing, and the instrument proves it.
02 · the method

Decompose the unknowable

A whole-cloth guess at "piano tuners in Chicago" is worthless — you have no calibration for the answer directly, so your number could be off by a factor of a thousand and you would never feel it. But you do have calibration for the pieces. You know Chicago is a few million, not a few thousand or a few hundred million. You know a piano gets tuned about once a year, not once a day or once a decade. Each factor is a question your intuition can actually answer to within a factor of two or three — and that is the whole game.

The piano-tuner method is general purpose. It works on anything you can write as a product of estimable pieces: the market for a product, the water in a reservoir, the time a project will take, the plausibility of a claim someone just made at dinner. The refusal to guess feels like intellectual honesty; it is usually its opposite. "I can't know" is a way of avoiding the discipline of writing down what you actually believe, factor by factor, where each belief can be checked and argued with. A number you can decompose is a number you can defend, revise, and be usefully wrong about.

03 · what to try

Three things to do with it

  1. 1Build the piano tuners and check it. Leave the presets on piano tuners and read the answer against the known truth. You will land in the hundreds — the right order of magnitude — from five guesses you made in ten seconds. Now widen one bracket and watch how little the answer moves.
  2. 2Toggle correlated errors and watch it degrade. Turn on correlated errors. Nothing about your brackets changed — only whether the guesses lean together — yet the distribution swells and the honest range blows out. That gap is the entire argument for guessing straight instead of slanted.
  3. 3Run a question about your own life. Switch to hours on your phone this year or heartbeats in a life. The unknowable-feeling personal number resolves into two or three brackets you can actually set — and suddenly you have an answer you can act on.
04 · the real magic

Why the errors cancel

A product becomes a sum in log space: the log of the answer is the sum of the logs of the factors. Each factor carries some error — call it a few tenths of a decade high or low. When the factors are independent, those errors are independent draws that add in quadrature: the total spread is not the sum of the individual spreads but the square root of the sum of their squares. Five factors each uncertain by ±0.3 decades do not compound to ±1.5; they combine to about ±0.7. The aggregate is dramatically tighter than any naive worst case, because overestimates and underestimates partly annihilate.

This is the deep reason decomposing beats guessing whole. A single blind guess at the final number has nothing to cancel against — its error stands alone at full size. Split the problem into five honest pieces and the errors, pointing every which way, interfere destructively. But the cancellation is a gift you can throw away. It relies entirely on the guesses being independent and unbiased. If you are systematically optimistic — every bracket nudged the way you want the answer to go — the errors stop cancelling and start reinforcing, adding linearly instead of in quadrature. That is what the correlated errors toggle shows: a motivated Fermi estimate is worse than no estimate, because it wears the costume of rigor while having quietly discarded the one property that made rigor possible. The honesty is the mechanism.

05 · the point of it all

Roughly right beats precisely ignorant

Strip everything else away and Fermi estimation is really answering one question: is it thousands, millions, or billions? Getting the exponent right is usually enough to act. You rarely need to know that a market is worth $47 million rather than $52 million; you need to know it is tens of millions and not tens of thousands, because those two answers call for completely different lives. The order of magnitude is the answer's real resolution, and chasing more precision than the decision requires is a way of staying busy instead of deciding.

order of magnitude is enough
Is this worth doing at all?

Whether a market, an audience, or a risk is in the thousands or the millions decides the answer. A factor-of-three estimate settles it. Precision here is wasted motion.

a factor of two is enough
Which of these two options?

You need only to know which is bigger, and by roughly how much. Two Fermi estimates, compared, almost always separate the choices cleanly.

rough, then sanity-check
Roughly how much should I budget?

Start with the order of magnitude to catch gross errors, then refine only the factors the decision is actually sensitive to. Most of them, it isn't.

precision earns its cost — rarely
Will the bridge hold the load?

Some questions have hard thresholds where being off by 20% kills people. There, precision is worth its price. These are the exception, not the rule — and you can tell them apart because you started with the Fermi estimate.

"I can't possibly know" is almost always false. You can nearly always bound it — put a floor and a ceiling on it, decompose it, and act on the range. The confidence Fermi teaches is the willingness to commit to an answer while holding its uncertainty openly, in the width of the bars, where anyone can inspect it.

metaphor 100 of 100 — a closing note

This is the last of a hundred, and it is the one the other ninety-nine were quietly serving. Every thinking tool in this collection — every distribution, attractor, and paradox — exists for the same reason: to help you act better under uncertainty, not to achieve a false precision that certainty never grants anyway. The point of a lens is never the lens.

Fermi's is the meta-skill beneath all of them: the move that turns the overwhelming into the decomposable, the unknowable into the roughly-knowable. If a single habit survives from these hundred pages, let it be this one — when a question feels too big to answer, break it into pieces you can bracket, and begin. That is what all hundred were for.

06 · the mapping back

The correspondence

MathematicsLife
the big questionthe overwhelming unknown that induces paralysis — the number you feel you can't possibly know
the factorsthe knowable pieces you decompose it into — each one your intuition can actually bracket
the rangeshonest uncertainty per piece — a floor and a ceiling, not a false point you'll defend too hard
multiplication in log spacewhy independent errors cancel — over- and under-guesses interfere and shrink the total spread
the order of magnitudethe answer's real resolution — thousands vs. millions vs. billions, usually all the decision needs
"roughly right"the confidence to act without certainty — commitment with the humility kept in the error bars

You can almost always be roughly right. "I can't know" is usually not a fact about the world — it is a failure to decompose.

07 · where the metaphor tears

Three honest rips

Cancellation needs honesty
The whole magic rests on independent, unbiased guesses. Systematically optimistic or pessimistic estimates don't cancel — they compound, adding linearly instead of in quadrature. That is the correlated-errors toggle made real: a motivated Fermi estimate is worse than useless, because it borrows the authority of a calculation while having thrown away the property that made the calculation trustworthy. Guess straight, or don't guess.
Some quantities are heavy-tailed
Fermi reasoning assumes the answer sits somewhere sensible on a smooth log scale. But some quantities are heavy-tailed or have hard discontinuities, where order-of-magnitude reasoning fails badly. A Fermi estimate of "how much will this startup make" is nearly meaningless: the outcome is drawn from a distribution with no typical value, where the mean is dominated by a tail you cannot bracket. Multiplying medians is no guide to a world ruled by its extremes.
Confidence can curdle
Fermi confidence can rot into glib overconfidence — the fluent estimator who impresses a room with a number and forgets it was ever a guess. The humility lives in the error bars. An estimate presented without its range has thrown away the honesty that made it worth anything: it has become a false point masquerading as knowledge, which is exactly the precise ignorance the whole method was meant to escape.