signal & secrecy · metaphor 10 of 100

To understand is
to compress.

To understand something is to be able to say it shorter. Information is surprise, understanding is compression — and rhetoric is compression with the losses chosen on purpose.

You have met the student who "knows the material" but can only give it back at full length, recited in the order it was received. That student does not know it yet — knowing it is being able to shrink it and unpack it again on demand. And you have met the opposite: the mentor who listens to your hour of tangled confusion and answers with one sentence, after which the hour was unnecessary. The mentor did not add anything to you. She compressed you.

Meanwhile every summary you have ever read — of a war, a marriage, a quarter's earnings, a week of strikes — was a lossy encoding, produced by an author who decided, on your behalf, what you could afford to lose. Claude Shannon gave all of this a number: a way to price what a message tells you, and therefore a way to see exactly what a shorter message threw away.

a the surprise meter

A small life, one symbol per day: sun or rain. Bias the weather, watch the stream, and see what each day costs to report. Probable days are cheap; improbable days are expensive; the average cost is the entropy.

p(☀) = 0.80 · p(☂) = 0.20
bar height = surprise −log₂ p · in bits
entropy H(p)
description budget
observed avg (window)

H = Σ p(x) · log₂(1/p(x)) Entropy is the average surprise: each outcome's cost in bits, −log₂ p, weighted by how often it happens. A fair coin costs a full bit a day; certainty costs nothing.

surprise as currency

The probable is cheap; the improbable is the news.

Shannon's move was to price a message by how unexpected it is: an event with probability p carries −log₂ p bits. Say "the sun rose" and you have spent breath on a certainty — cost near zero, information near zero. Say "it snowed in June" and you have handed over something expensive. This is why clichés carry nothing: a sentence everyone could finish for you was, in the exact technical sense, already known. And it is why the informative sentence always feels slightly improbable — it had to be, or it wouldn't be telling you anything.

Entropy H is just the average of that price — how much a source honestly needs to say per symbol to be fully described. Slide the bias above toward certainty and watch the budget die: a life of pure routine can be described in advance: it has nothing to report. The fair coin, maximally uncertain, is maximally chatty. Interesting lives are expensive to summarize; that is what interesting means.

b shannon's game — understanding = prediction = compression

Guess the next letter of an English phrase, one guess at a time, until you hit it. Shannon ran this experiment in 1951 to measure how much of English a reader already carries. Every guess you don't need is knowledge you already had.

Keys are ordered by a hand-set English bigram table — the codebook, made visible. You can also type on your keyboard while this area is focused.

letters solved0
avg guesses / letter
your estimate
random letters
4.75 bits/char
you, guessing
english (Shannon '51)
≈ 0.6–1.3 bits

Play a few letters and the gap between the purple bar and yours becomes visible — what you know is what you needn't be told.

honest toy: the key ordering comes from a small hand-set table distilled from standard English letter and bigram frequencies, defined in this page's source. Your entropy estimate is the average of log₂(guesses needed) per letter — a rough classroom proxy for Shannon's bounds, not his exact estimator. 4.75 = log₂ 27, the cost of a 27-symbol alphabet with no knowledge at all.

understanding is compression

The mentor's sentence, priced in bits.

You can guess the next letter because you carry a model of English — spelling, grammar, idiom, the drift of a sentence — and whatever the model predicts, the message doesn't have to say. Prediction and compression are the same act. A source you can predict is a source you can abbreviate; the bits you save are a measurement, in the strictest sense, of how much you understood.

This is what expertise is. The chess master doesn't see thirty-two pieces; she sees three familiar structures — her codebook has short words for what the novice must spell out square by square. The mentor's one sentence works the same way: years of cases compressed into a codebook so good that your hour of confusion was, to her, one known pattern with a short name. And it is why explaining something badly takes longer — the person who hasn't compressed the idea can only transmit it uncompressed, at full recitation length, which is how you can hear, in the sheer duration of an answer, that the speaker doesn't yet understand it.

c the rhetoric bench

One week in the town of Fairview: a strike at the mill, recorded as a ledger of 12 true facts. You must compress it to a headline budget. Three editors offer to do it — each keeps only what its policy values. Every summary below contains only true facts.

keep 3 of 12 facts · 25%

honest toy: the ledger and its tags (kind: number / feeling / scene · leaning: union / company / neutral) are hand-written and fixed in this page's source. Each editor is a transparent sort over those tags — no hidden weighting. The tone reading is the plain average of the kept facts' leanings; the loss bars count what fraction of each aspect's facts survived.

the chosen loss

All summaries lie by omission. The question is which omission.

Set the budget to three facts and read the bench again. The partisan's week, the accountant's week, and the novelist's week are all accurate — not one false statement among them — and they are three different weeks. Nothing here required a lie. At any compression ratio below 100%, something true must go, and which truth goes is the entire editorial act. The question to ask of a summary: what did this encoding decide I could afford to lose?

Seen this way, propaganda rarely needs fabrication. It needs only a compression policy applied consistently: always drop the same aspect, week after week, and the reconstruction in the reader's head drifts arbitrarily far from the ledger while every individual sentence survives a fact-check. The defense is detection of the codebook: read any summary asking what kind of fact this outlet never keeps — the quantities, the feelings, one side's whole week — and you have recovered the policy, which is the only part of the message its author didn't choose to transmit.

what to try

Sixty seconds at each bench.

01

Play Shannon's game

Guess a whole phrase in panel b. Notice where you need one guess (mid-word, after q) and where you need many (the first letter of a new word). Your final bits-per-character sits far below 4.75 — that gap is your knowledge of English, measured.

02

Kill the surprise

In panel a, drag the bias to p(☀) = 0.98 and let the days flow. The bars flatten, H collapses toward zero, and the rare rainy day towers over the rest — routine is cheap, and the exception carries the whole story.

03

Run all three editors

In panel c, hold the budget at 3 and click through the editors, watching the tone pin jump while the ratio never moves. Then raise the budget and find where the three weeks converge back into one.

the mapping

Mathematics ↔ life.

MathematicsLife
probability p(x)What you already expect — the model of the world you bring to the message.
surprise −log₂ pWhat an event actually tells you: the improbable is expensive, the cliché is free.
entropy HHow much description a source honestly needs — the going rate for reporting a life truthfully.
compressionUnderstanding: prediction turned into brevity. What your model foresees, the message may omit.
lossy encodingEvery summary, headline, and memoir — shorter than the events, therefore missing some of them.
the compression policyThe rhetoric: which losses were chosen, and whether they were chosen the same way every time.

where the metaphor tears

Three honest failures.

Bits price description, not significance.

Shannon's measure is surprise relative to a model — it knows nothing of consequence. "The test came back positive" can cost one bit and rearrange a life; a lottery draw costs twenty-five bits and matters to no one. A fact can be cheap in bits and heavy in weight, and no amount of information theory will tell you which facts those are. The metaphor prices the telling, never the mattering.

Compression is one face of understanding.

"To understand is to compress" fits knowing-that beautifully and knowing-how hardly at all. The cyclist cannot shorten balance into a sentence, and no paragraph about grief compresses the acquaintance of it. Some understanding lives in the hands and the history, not the codebook — the identity holds for what can be said, and much that we understand best was never said.

Memory is a lossy channel with an editor you never hired.

The rhetoric bench lets you choose the policy and read the distortion report. Your own memory extends you no such courtesy: it compresses every week you live with a policy you didn't set, can't inspect, and only discover by its artifacts — the flattering emphasis, the vanished quantities, the feelings promoted to facts. The most consequential summary of your life is being written continuously, by the one editor you cannot audit.