Meaning is a place

A model cannot look up what a word means. So it gives every word a position on a vast map: “dog” sits next to “puppy”, far from “tax return”.

Every token gets a long list of numbers: its coordinates on the map. Words with similar lists sit near each other. The model never learned a definition. It learned the neighbourhood.

Like the seating plan at an enormous dinner. You never met the guests, but who sits next to whom tells you most of what you need.

Close means used together

No one tells the model where to put a word. It reads an ocean of text and nudges every word until words that keep company sit together. Dogs and puppies appear in the same stories, so they drift together. Money words form their own region.

Once meaning is distance, a question about meaning becomes a question about position. Which word fits here? becomes which points are nearest? A machine answers that fast.

king − man + woman ≈ queen

Directions on the map carry meaning. The step from “man” to “woman” is roughly the step from “king” to “queen”. It works often enough to be useful.

The map, drawn down to two

Touch any word. It lights up with its three nearest, and the caption names them. The clusters form on their own: animals with animals, food with food, money and work off in another corner.

Try it Hover, tap, or focus a word.
A lookalike. A real embedding has thousands of dimensions, not two. This is a hand-drawn map of real clusters.

Where you meet it

Every tool that promises to “read your files” runs on this. It turns each document into a point and stores the points. Your question becomes a point too. Then it fetches the nearest ones. No reading happens. It is a nearest-neighbour lookup, thousands of times a second.

Like a librarian who never reads the books, but knows which shelf each one lives on. Ask for a feeling, and you get the shelf.

What the map gets wrong

Two honest limits. The map is not a dictionary: “bank” the money place and “bank” the river share a point. And near is not the same as true: two sentences can sit close together and disagree completely. The map tells you what something sounds like. Never whether it checks out.

Working on this?

If AI is stuck somewhere in your business, tell me where. I read every one of these.