The Lab · Learn

How AI actually works

Small pages that run on this device. Read the plain version, move the sliders, watch what changes, then say what you saw. Tap any word you do not know.

Lesson 8

Embeddings

AI turns words into lists of numbers, so words that mean similar things end up pointing the same way.

It is like: Like signposts at a crossroads. Two signs pointing the same way lead to similar places, however long each sign is.
Why it matters at work: This is how search finds “vacation request” when you typed “time off”. It matches meaning, not spelling.

The exact idea

An embedding turns a word into an arrow of numbers, and cosine similarity compares where two arrows point.

Words on this page

Try this

Pick “cat”. The best matches are other animals, and “car” scores low.
Drag the arrow length slider down. The raw overlap numbers shrink, but the match scores stay the same.
Steer your own arrow between the fruit and the animals and see which side wins.

Change it

Ask for a new word, “scooter”, placed near “bicycle”, then pick it and check its neighbors.

Make the arrow shorter and the raw overlap numbers change, but the match scores do not. Why?

Pick the one that fits what you saw.

Cosine similarity divides by both arrows' lengths, so length cancels and only the angle counts.The match scores are rounded, so small changes do not show.A shorter arrow sits closer to every other arrow, and that evens out.The match scores only update when you pick a new word.

Field notes

Reading the notes.