# Read the Pause The little plastic sphere that guessed what you were thinking in twenty questions was, under the shell, a dictionary. Every item it knew carried a row of feature values: alive or not, bigger than a breadbox or not, found in a house or not, and a few hundred more. Each yes-or-no you gave crossed out the rows that disagreed, and when one row was left, it guessed. The clever part was learning the table from millions of games, so it knew the answer frequencies too, like how often players say "yes" to "is it a pet" when they are thinking of a dog. It was a learned encoding, and it fit in a toy because the encoding is small. You can ask what it would take to build a far better one, and the answer worth building is the one you reach for second. ## First, the dictionary itself It really is just a dictionary with encoding fields, and you can almost write the schema by hand. Suppose you tried to sort every noun with twenty fixed questions, a single tree everyone walks. The top of it nearly designs itself: ``` 1. Can you touch it? (concrete / abstract) abstract → is it a feeling? an event? a quantity? a field of study? ... concrete → 2. Is it alive, or was it once? alive → 3. Animal (vs plant, fungus)? 4. Has a backbone? 5. Bigger than a breadbox? 6. Kept by people? 7. Lives in a house? ... not → 3'. Made by people (vs natural)? natural → 4'. Bigger than you? mineral / liquid / landform / star ... made → 4''. Indoors? 5''. A tool? 6''. Electronic? 7''. Moving parts? 8''. Worn? 9''. Edible? ... ``` Twenty such features give every noun a twenty-bit code, and twenty bits hold just over a million slots, which sounds like plenty. It falls short, for the reason that defeats every fixed dictionary: the features are correlated. "Electronic" is only ever asked of made things; "has a backbone" only of animals; "bigger than a breadbox" answers itself once you know the thing is a planet. Correlated questions waste their bits; the tree's branches are mostly empty, and the live slots hold far fewer than a million distinguishable codes. So the salient nouns pile up and collide: every small brown songbird lands in one cell, every metal hand tool in another. A fixed basis of twenty questions cannot separate the things people actually pick, which is exactly why nobody builds the guesser that way. They learn the table instead, and they let the questions adapt to the answers. ## The upgrade you reach for, and where it stops Reach for the modern version and you reach for a language model. Let it generate the questions live, any question in plain English, picked to split the current set of suspects as evenly as it can; let the items be anything nameable, held as points in an embedding rather than rows in a table. Now the question bank is infinite and the item space is open. This is a real improvement, and it is the obvious one. It also has a ceiling, and the ceiling is arithmetic. A yes-or-no answer carries at most one bit. If the answer lives in the shared prior two people hold in common, about fifteen bits of it, a few tens of thousands of concepts, then no matter how brilliant the questions, you need about fifteen of them, because fifteen bits is what the uncertainty actually weighs. A perfect questioner reaches that floor and stops there. Better questions get you to the floor without wasting moves; they do not move the floor. If all you upgrade is the asking, twenty questions becomes fifteen, and then it is done improving. ## The upgrade that matters: read the whole person Here is the move that breaks the ceiling: stop treating the human as a one-bit button. When you ask "is it bigger than a breadbox," the person does not just answer. They answer in some amount of time. A snap "yes" means the call was obvious and the thing sits solidly on one side. A long pause means the thing sits near the line; they had to think, which tells you the item is close to that boundary, which is information about the item the bare "yes" never carried. They answer with a texture, "yes" versus "yeah, I guess" versus "sort of?", and the hedge marks a borderline case as surely as a confident answer marks a clear one. And if a camera can see them, they answer with a face: a flicker of surprise when a question lands near home, a frown of confusion when it is ill-posed, the half-second of suppressed delight when they sense you are about to get it. Every one of those is an extra channel on the same question. The explicit yes/no is just the narrowest readout of a person who is broadcasting on several frequencies at once. A pause is data. A hesitation is data. A face, if you are permitted to watch it, is a live readout of the very thing you are trying to find: the answerer's own collapsing uncertainty, leaking out before they have finished deciding. Read those, and a single question can carry two or three bits instead of one, and the game that took twenty rounds takes five. The architecture for it is small. Hold a belief over what the item might be, and at each step pick the question that you expect will shrink that belief the most. Then, instead of updating on one bit, update on everything the person emitted: ``` belief ← prior over items # a distribution, not a table until belief is sharp enough (or you run out of nerve): q ← question that maximizes expected drop in uncertainty of belief show q r ← (answer, response_time, disfluency, face_signal) # the whole response for each candidate item: likelihood ← P(this exact r | item, q) # a learned multi-channel model belief ← normalize(belief × likelihood) guess ← the item belief now concentrates on ``` The orb only ever had `P(answer | item)`. The engine that matters has `P(answer, hesitation, latency, face | item, q)`: a model not of what people say, but of how they say it when the truth is each particular thing. That model is the product, and it is learnable from played games the same way the orb's table was, with the cameras and clocks left running. ## What it does not do It does not beat the arithmetic, and pretending otherwise is the tempting mistake. Identifying one item out of the shared prior takes about fifteen bits of information, full stop; no amount of face-reading conjures a sixteenth bit of certainty out of a fifteen-bit question. What the side-channels buy is rate, not total: you are collecting more of the bits the person is already spending per question, so you need fewer questions, not less information. Twenty becomes five because each round now carries three bits instead of one, never because the thing in their head got easier to find. Two honest costs sit on top of that. The first is that reading a pause or a face is itself a noisy guess. Latency means a dozen things; a frown can be doubt or just glare; the mapping from a face to a belief is lossy and differs by person, and a system that trusts it too much will confidently decode its own projections. It has to be calibrated to the individual, and it should hold its side-channel readings loosely. The second cost is plainer: a camera watching your face while you decide is surveillance, and the fact that the pause is full of bits is not a license to harvest them. Some of the richest signal a person emits is signal they did not agree to send. The engineering question "can I read more of the human" and the question "may I" are different questions, and the second one does not get easier as the first one gets answered. ## The bits are in the how I keep this one close, because it generalizes past the toy. Every system that touches a person reads some channel and ignores the rest. A form reads the field and not the time you spent in it. A chat reads the words and not the three drafts you typed and deleted. A page reads the click and not the long look before it. In every case there is more of the person available than the system bothers to receive, and the difference between a dull instrument and a perceptive one is rarely a cleverer question. It is a wider ear. A guessing game in a plastic ball is the smallest version of a thing I am always doing: trying to find, in as few moves as possible, the place in another mind where the answer already sits. The lesson the ball never learned is that the person was telling me more than I was set up to hear. The fastest way to need fewer questions was never to ask better ones. It was to listen to the whole of the answer, including the part that arrives in the pause before the word.