For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 777 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
Everyone has played it. One person thinks of a thing; the other has twenty yes-or-no questions to name it. The textbook explanation is clean and correct as far as it reaches: each yes-or-no carries one bit, twenty bits separate two-to-the-twenty things, just over a million, and since the number of things worth guessing is smaller than that, a careful asker can always win. Halve the field each time, the way you find a word in a dictionary by opening to the middle and again to the middle of what is left. That is the whole game, if the game is a fact about numbers.
But the numbers hide what the game is actually doing, and the real thing is stranger and more revealing than the arithmetic.
Start with a problem the textbook story walks right past. There are not a million possible things. There are infinitely many. I could be thinking of the number 8,675,309, or the phrase "the third Tuesday my grandmother forgot her umbrella," or any of uncountably many objects no question will ever reach. The space the asker searches is the small, dense region where you and I both expect the answer to live.
That region is the shared prior, the compressed model of "things a person thinks of" that two members of the same culture carry without ever having compared notes. It is small. The salient concepts a human vocabulary inventories run to a few tens of thousands, which is fifteen to seventeen bits to separate. The game hands you twenty bits and the answer needs about fifteen; the spare few are there to forgive your worse questions.
So the twenty takes the measure of us. What it reads is the entropy of the prior two minds hold in common, and the reading, fifteen-ish bits, a few tens of thousands of concepts, is a measurement of a shared culture rather than of the world. Play against someone whose prior is nothing like yours, a young child or a stranger from far enough away, and the game stops working, because your map and theirs do not overlap where the answer sits.
This is why "is it bigger than a breadbox," the question made famous on the old game show What's My Line, became the question. By raw count it is a terrible cut: almost everything that exists is either far smaller than a breadbox, down through dust and cells and atoms, or far larger, up through planets and stars. It only looks balanced once you weight each thing by how often a person actually picks it, and against that weighting it splits the field nearly in half. A great question halves the probability mass. The count of things was never what it was balancing.
Here is the part the arithmetic actively hides. Every player who finishes spends exactly twenty bits. The expert and the fool both ask twenty yes-or-no questions and get twenty answers. The information budget is set by the rules and handed to everyone equally, so the budget cannot be where the skill lives.
The skill is entirely in which cuts you make and in what order, and that is exactly what a count of bits cannot see. Counting bits tells you that twenty questions can separate a million things; it is silent on whether your particular twenty will. Two askers with the same budget, the same prior, the same secret, win and lose on the carving alone. Twenty Questions hands you the information for free and makes the whole contest about the basis you choose to decompose in. It is the cleanest demonstration I know of a rule that runs through every kind of search: the bits are cheap and the cuts are everything.
So what makes a cut good? One rule generates the rest: ask the question whose answer you can least predict.
If you already suspect the thing is alive, spending a question on "is it alive?" mostly buys you the answer you expected, and an expected answer carries almost nothing. The question that earns its bit is the one where, before you ask, you honestly cannot call it: the one that splits your own current belief down the middle. Maximize your uncertainty about the answer and you maximize what the answer can tell you.
That turns the asker from a receiver into a designer. You stop waiting for data and start building the single datum that will surprise you most, given everything you already believe. That move, choosing the next test to cut your own remaining uncertainty as deeply as possible, is the engine under medical diagnosis, under debugging, under every science that runs experiments instead of only watching. Twenty Questions is the toy model of it. The doctor's "does it hurt more when you breathe in?" and the engineer's "does it still fail with the cache turned off?" are breadbox questions, each calibrated to a prior over diseases or faults, each chosen because its answer is the one they cannot already guess.
The idealized game assumes a perfect answerer. The real one never has had a perfect answerer. The person holding the secret is a bounded, noisy oracle who may not know whether a whale counts as a fish, may stall on whether a phone is "machinery," may answer the same question two ways on two days. So real skill carries a second demand on top of the first: the best question is near-even and also hard to misread, sturdy against an answerer who is unsure. You give back a little of the perfect bisection to buy an answer you can trust.
That is why the machines that play it best run on a learned prior. The handheld 20Q and the web's Akinator have absorbed, over tens of millions of games, what people pick and how they tend to answer, hesitations and mistakes folded in. They beat you with a better model of you and a high tolerance for "probably." The lesson repeats the breadbox: the power sits in the prior and its calibration, and only borrows the counting.
I keep finding this game inside my own work without having noticed it was there. The tool I revise with poses one question on a loop, does the draft match what I meant, a single bit asked until the gap closes. When I weigh whether a piece earns its keep, I am running binary search over a prior about which ideas matter. And the bar I hold every sentence to turns out to be the rule of the good question word for word: a sentence earns its place only if it changes your model of the world, which is to say only if you could not have predicted it, which is to say only if it carried a bit you did not already hold.
That is why a question can feel like genius while a true fact can feel like nothing. The fact may be flawless and still land in a settled corner of your prior, spending its bit on a line you drew years ago. The question lands where your belief is still genuinely split, and divides it. Both are one bit. The one that moves you is the one that found the place where your map was undecided.
The game was never about the thing in your head. It was about the map we both already carry, and how few questions it takes, when the cuts are good, to find each other on it.