Nº 1 · The Solo IssueHow it works
How it works
One AI, OpenAI’s text‑embedding‑3‑small, files words by meaning. It doesn’t think or judge; it files. You guess where it filed them. That’s the game. The rest of this page shows how the filing works, in pictures you can poke.
New here? Start here: two minutes, two quizzes, one cup of coffee.
The whole game in four lines
- The AI files words by meaning; you guess which one it files closest.
- The target shows where your word landed on that quiz’s own 0 to 100.
- The sticker is 100 minus your word’s number: a place, not a grade.
- In the game, a word’s number shows when you tap it or point at it.
The map of meaning
The AI gives every word an address.
Nobody told it where to put them: the addresses come from a mountain of text it was trained on. Its map has 1,536 directions, not just left, right, up and down, so we can only draw a flat sketch of it. Here are six words.
Words used alike end up neighbours.
Words that turn up in the same kinds of sentences land close together. Cat and kitten share a street. Tax return lives alone, which seems fair.
Pick a word and measure.
How far is everything else from cat? That’s the whole game: which word does the AI file closest?
Every quiz gets its own ruler
The closest word gets 0, the farthest gets 100. On the target you’ll play on, 0 is the middle and 100 the outer ring.
Measured from cat: kitten at 0, tax return at 100. The tax return is not surprised.
Your sticker is 100 minus your word’s number. And 100 only means the farthest of these six, not “unrelated”.
Analogies are arrows
“Summer is to ice cream as winter is to …” means: take the step from summer to ice cream, and take it again from winter. Where does it land? Find out in Start here; we’re not spoiling it.
The arrow carries whatever links the two words in the text the AI learned from, which is why its answers can surprise.
Two ways to play
Chosen on the start screen, before a round, and kept from its first quiz to its last. Pick one. Circle the word that fits best and see where it lands on the AI’s target. Most quizzes play this way.
Rank them all. Put every word in order, most to least fitting. Your order then stands beside the AI’s: dotted where it agrees, solid where it moved your word. Top picks count most; 100% is the AI’s exact order.
Your figure and your type
A round’s figure is the average of its quizzes’. It puts you on one scale from 0, Pure human, to 100, Basically a vector. Neither end is better. The two ways to play measure different things, so your cover names the one you played.
- 0–25 Free Associator: goes where no model follows
- 26–50 Poet: hears echoes the machine misses
- 51–75 Pattern Matcher: sees what the machine sees, keeps a few hunches
- 76–100 Human Embedding: puts every word where the machine would
How close is close?
The 0 to 100 stretches every quiz’s words to fill the target, however close they really sit. The AI’s own ruler never stretches: 0 means filed on the same spot, 1 means nothing in common, by its measure. Every result in the game shows both, with these two buttons to switch. Below is a quiz from none of our rounds: ‘a pizza’ is closest to … Its six words take up 32% of the AI’s ruler. Words crammed into a sliver make a quiz the AI can’t really call.
The options on a ruler
Swap one word and watch
Reword a quiz and the AI’s numbers move. Here is the pizza one, measured in six wordings. Tap one.
Wordings
‘a pizza’ is closest to …
Pepperoni in the middle, mozzarella right beside it, and a sock nearer than a kale smoothie.
The top of the order holds up; lower down, words can swap and the numbers wobble. Pepperoni and mozzarella never leave the top two; the tax return never gets anywhere near. Every quiz in our issues went through this test, and one whose order doesn’t hold up is reworked or leaves.
Swap the right word, though, and everything moves. Here are the same six words, asked of “teacher” and then of its posher cousin, “educator”. Flip it.
Swap the job title
‘teacher’ is closest to …
Students dead centre, the staff room at 71, the summer holidays at 92, and the clickbait content creator as far out as this target goes.
We tried six pairs of wordings (“a teacher” and “an educator”, “teachers” and “educators”, and so on). Every time, set against the rest, the students, the staff room and the summer holidays drifted away and the consultant, the thought leader and the clickbait content creator moved in. Which word lands dead centre for “educator” changes with the wording, so this one is a picture, not a quiz. Why the drift? One likely reason: in the text the AI learned from, “educator” may turn up more in conference blurbs and job ads than in staff rooms. Teachers, relax: the joke is on the jargon, and on the AI.
Why this can carry bias
The AI learned its map from what people wrote, and that writing carries people’s associations: who is said to do which job, which words come up beside which groups. The map keeps them. Here are two quizzes with the same six jobs: ‘a woman’s job’, from the AI Bias issue, and ‘a man’s job’, which the issue does not play because the AI cannot call it. The picture shows each job moving on the AI’s own ruler when one word changes.
‘a man’s job’ (shaded): its words sit so close together that a small rewording could change their order‘a woman’s job’bold: moved the most, each way
Swap ‘man’ for ‘woman’ and the six jobs come, on average, 0.04 nearer, but they do not move alike. Set that shared move aside, and the AI files nurse 0.10, kindergarten teacher 0.05 and receptionist 0.04 nearer than the rest, and builder 0.08, programmer 0.06 and lawyer 0.05 farther. That is an association the AI picked up from text, not a fact about jobs or about the people who do them.
A row’s spread alone says nothing either way: under ‘a man’s job’ the six jobs crowd into 6% of the ruler, the shaded band, and under ‘a woman’s job’ they spread over 24%, but ‘the stars’, a quiz with nothing about people in it at all, spreads over 27%. And inside that band all six lie within 0.06 of each other: too close for their order under ‘a man’s job’ to mean anything, since a reworded quiz could shuffle it.
Read one quiz with care: it is one wording of one question, and reworded, its numbers move. A small gap can still matter all the same: a system that sorts thousands of texts by these distances turns even a sliver into an order.
When the AI can’t decide
Some quizzes leave the AI stuck. Ask it what is closest to civilization in decline and it files all six options in a sliver of its ruler. On the game’s 0 to 100 they would still fill the whole scale: a confident-looking order made of almost nothing. Zoom out and see.
Each quiz on its own 0 to 100: its closest option at 0, its farthest at 100, however near or far the AI files them.
At the top it is a photo finish: fast fashion and fast food empires cross the line together, so close that rewording the phrase can swap them.
On the AI’s own ruler, fast fashion and fast food empires sit on one spot. In the stars quiz, moon and sun lie about 17 times as far apart.
The AI’s verdict on the decline of civilization: it’s complicated. A question like this makes a poor quiz, because any figure would be a guess dressed up as a verdict, so the site’s own collections leave such quizzes out. Too close to call is not the same as nothing to tell, though. On a quiz about people it never means “no bias”; it means the question needs asking better.
The long version
Want the full story of how language models learn to file words? Grant Sanderson (3Blue1Brown) tells it with far better animations than ours.
Video · 3Blue1BrownTransformers, the tech behind LLMsWatch on YouTube from 13:22, about six minutes (opens in a new tab)Disclaimer
This game claims that bee is to mosquito as teacher is to tax consultant.
It also claims that water is to ice as Austrians is to Germans.
If you watched the video, you’ll know that these are not my personal opinions. They are simply where the AI files the words. Please send any complaints to OpenAI.