Embeddings: how AI remembers what matters

Computers don't understand words. They understand numbers. When you type a question, the machine turns your words into a list of numbers — an "embedding" — and matches that list against the numbers it already has stored.

Think of a family photo album. Each picture has a number that says how similar it is to every other picture. The photo of tía Rosa at the quinceañera gets a high similarity score with the photo of her dancing, but a lower score with the photo of the food table. The album doesn't know "quinceañera" or "dancing." It just knows the numbers.

Embeddings work the same way. They're a way for machines to say "these two things belong together" without knowing what the things actually are. A recipe for enchiladas and a recipe for chilaquiles get close together in the number space. A tweet about the migra and a news story about ICE get close. The AI doesn't need to be told this. The numbers tell it.

This is why your search engine can find what you mean even when you misspell things. This is why your phone can tell you're talking about "la migra" when you type "la migra app" and show you the right app. This is why you can search for "something to fix my tire" and get results about tire repair, not something about fixing your relationship.

The trick is that embeddings capture what things are for, not just what they're called. Two things can have different names but live near each other in the number space. Two things can share a name but live far apart.

Next time you search with your phone, notice what it finds. If it finds the right thing, the embeddings are doing their job. If it doesn't, you know the numbers didn't quite line up.

Why this matters for us: The way your phone and search engines understand your words determines what you see — and what you miss — in the stories that affect your family, your neighborhood, and your wallet.