What am I looking at?

A map of meaning. Every point is a word or phrase, and where it sits tells you how strongly it matches the two concepts on the axes. The whole idea builds up from a single line.

Start with one line

Pick a concept, say dangerous. Now give anything a score from 0, its opposite, to 100, a perfect match.

Do that for a few animals and each one gets a place on the line. A word has become a number, and the number says something about what the word means.

Six animals on a line from harmless to dangerous: kitten 1, cow 21, horse 47, wasp 77, wolf 94, shark 99. kitten 1 cow 21 horse 47 wasp 77 wolf 94 shark 99 0 · harmless dangerous · 100
These are the scores Jev gave when asked. The wasp outranks the horse: this is about reputation, not size.

Add a second line and you have a map

Ask a second question, big, and stand the first line upright. Each animal now has two numbers, which is enough to pin it to one spot.

Things that get similar answers land close together. The wolf and the shark are neighbours here, not because they look alike, but because both are fairly big and very dangerous.

Swap either question and every point moves. Closeness always depends on what you asked.

The same animals on a map. Across is big, up is dangerous. Kitten sits bottom left, wasp top left, cow and horse on the right, wolf and shark together at the top right. close together = similar answers kitten cow horse wasp wolf shark 0 50 100 0 50 100 big → dangerous →
Same animals, same dangerous scores, now with big running across.

An embedding is the same trick with far more lines

Search engines and AI systems turn every piece of text into a long list of numbers, often more than a thousand. That list is called an embedding. Each number is a position along one line, so text with similar meaning ends up nearby.

That is how a search for “running shoes” can find “sneakers” without sharing a single word.

The catch: a computer worked out those lines for itself, and none of them has a name. You can measure how close two things are, but you can’t read why.

For the longer version, read Text Vectorization ↗, a deeper explanation of how text becomes numbers.

An embedding of “shark”

A thousand or more numbers, none with a name. (Example values.)

“shark” on this map

Two numbers, and you named both.

Here, you choose the lines

You type two concepts. For every phrase you add, Jev, an AI model from TypeSafe AI, reads it and judges where it falls on each concept, from the opposite at 0 to a defining example at 100.

So this map is not a thousand numbers squashed flat. It is two direct judgements, which is why you can read both axes, and why any concept you can put into words can be one.

The scale Jev uses for the concept dangerous, with five reference points from opposite to defining example. Kitten lands at 1, horse at 47, shark at 99. Where does it fall on “dangerous”? kitten 1 horse 47 shark 99 0 25 50 75 100 theopposite leansaway neutral,mixed, n/a leanstoward definingexample
Jev works from five reference points and can land anywhere between them.

Reading the map

A score is an opinion
It reflects common associations as one AI model sees them. It is not a fact or a measurement.
50 often means “doesn’t apply”
The middle is for things that are neutral, mixed, or have nothing to do with the concept.
Confidence is not correctness
It shows how firmly Jev settled on a score. For “shark” it was 97% on dangerous and 60% on big.
Scores can wobble
Ask Jev the same thing twice and the answers can differ slightly, so a point may shift by a point or so over time.

Try it: set X to big and Y to dangerous, add a few animals, and see how close you get to the picture above. Back to the map