Embeddings — Explain Like I'm 5

The Secret Map Inside Every AI

Here’s a puzzle: your phone can finish your sentence. You type “I’m really hungry, let’s get some—” and it suggests “pizza” or “food” or “tacos.” Not “Thursday.” Not “purple.” How does it know those words make sense there, and “Thursday” doesn’t?

It’s not reading a dictionary. It’s using a map.

Every Word Gets GPS Coordinates

Imagine you made a giant map of all the words in English. Similar words live close together on the map. “Dog” and “puppy” are neighbors. “Happy” and “joyful” are basically roommates. “Car” and “vehicle” share a street. But “dog” and “justice”? They’re on opposite sides of town.

That’s basically what an embedding is. Each word gets a set of coordinates — not just two numbers like a map, but hundreds of numbers. That’s why it’s called a vector. But the idea is the same: things that mean similar things end up near each other.

Why Not Just Use a Dictionary?

A dictionary says “cat: a small domesticated carnivorous mammal.” That doesn’t tell you that cats are related to dogs, tigers, pets, fur, meowing, or the internet.

Embeddings figure out relationships from how words actually get used, not from definitions. Computers read billions of sentences and notice: “dog” and “cat” show up in similar contexts constantly. “I love my dog.” “She adopted a stray cat.” After enough of these, the computer learns that dog and cat belong in the same neighborhood — without anyone defining what those words mean.

The Really Wild Part

Because words are coordinates, you can do math on them. And the math works.

If you take the coordinates for “king,” subtract “man,” and add “woman” — you get coordinates that land almost exactly on “queen.”

Nobody programmed that. Nobody typed in “king minus man equals queen.” The computer discovered that relationship just from reading enough text.

Same trick works for: Paris minus France plus Italy ≈ Rome. Or: walk minus walking plus swimming ≈ swim.

The map has real geometry to it.

Not Just Words Anymore

Today, embeddings work on images, videos, music, even code. Spotify turns songs into coordinates and puts similar-sounding songs nearby — that’s how “if you like this, you’ll like that” recommendations work. Pinterest does the same with images. When you do a Google Image search, your photo gets turned into coordinates and the engine finds nearby coordinates.

Your face on your iPhone? It’s a set of numbers — a point in face-space. Every time you unlock it, your phone checks whether today’s face-coordinates are close enough to the ones it saved when you set it up.

One Thing to Remember

Embeddings are a way of turning meaning into location — words, images, and sounds become points on a map, and things that mean similar things end up as neighbors. Once you have that map, you can do math on meaning itself.