A machine lays out words as places on a map. Watch which words count as nearest change as you turn the view.
move your cursor left and right — the view turns, and the nearest list re-answersdrag the line below — the view turns, and the nearest list re-answers
The word bridge never moves. But which words sit nearest to it keeps changing as the view turns. That single fact is the whole lab.
Why it matters: whenever a tool tells you two things are “similar” — a search result, a suggested contact, a “related” article, a CV matched to a job — it measured that closeness along one direction it picked. Change the direction and the answer changes. This page lets you feel that happen.
The geometry of meaning
How a machine turns words into places on a map — and why nearest depends on which way you look.
Narrated by a synthetic voice — not a recording · ~0.5 MB
A machine can't read a word the way you do. So it does something stranger: it gives every word it knows a fixed spot on a map. Words it has seen used in similar ways — river and stream, home and comfort — get placed close together. Words used differently sit far apart. That is the whole trick of this layer of an AI: meaning turns into a location, and “similar” turns into “close by.” Engineers call one of these word-maps an embedding; on this page I'll just call it a word-map.
So “nearest on the map” is really the machine's guess at “most alike.” I wanted to see one of these maps, so I dropped a few dozen ordinary words into one. But here is the catch, and it's the whole reason for this page: a real word-map isn't flat. It sorts words along hundreds of directions at once. A screen has only two. So to show you anything, I flatten the map down to two directions — and one of them is yours to choose. The moment you choose, you've decided which kind of “similar” the picture will show.
What is illustrative. Every spot and every direction on this page is hand-drawn to teach the shape; none was measured from a real AI tool (the part that learned where words go from many examples). Real word-maps are specific to each tool, live in hundreds or thousands of directions, and any flat picture like this throws most of them away. (And a real tool actually places tokens — the sub-word pieces from the last lab — not tidy whole words. I'm using whole words here so the map stays readable.)
What the picture leaves out. A real AI tool does not stop at this fixed map. Once a word lands in a sentence, the tool builds it a second, moving position — shaped by the words around it, different every time the word appears. This page freezes that motion. It shows you the fixed map underneath, not the live thing standing on top of it.
You're reading the pocket version — drag the line to turn the view, tap any word to pin it and see its nearest words.
The full map is next. Start by just watching it.
Can a word change its neighbours without moving?
You'd expect a word to have one set of nearest words, settled for good. Watch what happens when only the view turns. One word is pinned below: bridge. Hit play, and the map will re-sort itself — from grouping words by made vs. living to grouping them by divides vs. connects. The orange lines reach for whichever words are nearest now.
Sorted by made things, bridge sits near tower, road, key. Sorted by connectors, it sits near link, between, translate. Same word, same map — bridge never moved. Only the direction you measured along did.
Next: what a spot on this map actually means ↓What does a spot on the map actually mean?
Here are three dozen everyday words, dropped onto one shared map. Left-to-right is the direction you get to choose; up-and-down is fixed — warm words float, cold words sink.
None of these spots is a dictionary definition. Nobody typed them in. As the tool learned from lots of text, every time two words got used in similar ways it nudged them a little closer — so a word's spot is nothing more than a memory of the company it has kept.
Look at the gaps, not the words — distance is the only thing this map knows.Takeaway: each word here has one fixed spot — really a short list of numbers, its address on the map. The distance between two addresses is all this layer means by “related.” Next: change which direction the screen draws left to right.
Can a new view change what looks near?
Drag the line along the bottom — the control that turns the view. Right now the map is sorted by made vs. living: factories on one side, forests on the other. Drag it across and the same words re-sort by whether they divide or connect.
Each word's spot on the map never moves. All that changes is which direction the screen draws left to right — you're turning the map to look along a different quality, not editing the words. The page just re-measures which words are nearest inside the view you've chosen.
Try it — drag the line end to end; one cluster dissolves, another forms.Done — you turned the view, not the words. Their spots stayed put; only the direction you're looking along changed. Next: pin one word and read off its nearest words in this view.
So which nearest words are the real ones?
Pin home and turn the view. In one direction its nearest words are warmth, comfort, promise; in another they're coin, key, gate. Same word, opposite company — decided entirely by the direction you chose.
The orange lines mark the nearest words measured by straight-line distance in the view you're looking at. (A label may be nudged aside so you can read it; its faint thread points back to the exact spot being measured.)
Try it — pin any word, then turn the view. Try to pin down its “true” nearest words.Every view gave a different answer. None is wrong — and none is the whole story. The full answer needs every direction at once, more than any flat picture can hold.
There's one more catch ↓What does every flat view leave out?
This is why nearest is slippery. On the fixed map this page imitates, a word really does have one settled set of neighbours — but that set lives across hundreds of directions at once. The instant you flatten it down to a picture you can see, you've picked which of those directions to keep and which to squash away. A different pick, a different set of “nearest” words.
And this fixed map is only half the real story. In a live AI tool, once a word lands in a sentence its position shifts again to fit the words around it — so its true company keeps moving even after you've chosen a view. This page can only be honest about the fixed map, and about the one direction you're looking along. That direction is yours now — turn it, pin anything.
Takeaway: a “nearest words” list is never a plain fact about a word — it's the answer to one question (which direction did you measure along?). Carry that question into any similarity map you meet.
That's the whole map. One thought to leave with ↓