What Is an Embedding?

A computer cannot compare two sentences the way you can. It has no notion that "the film was wonderful" and "I loved the movie" mean nearly the same thing. Embeddings are how that gap gets closed.

The core idea#

An embedding turns something — a word, a sentence, an image — into a list of numbers. That list is a position in space. Things with similar meaning are placed near each other; things with unrelated meaning are placed far apart.

Once meaning has become geometry, "how similar are these two ideas?" becomes a question you can answer with arithmetic rather than understanding.

Why this matters so much#

Almost everything built on top of modern AI depends on this one move:

• Search finds documents whose embeddings sit near your query's embedding
• Recommendation finds items near the things you already liked
• Clustering finds groups of points that sit together
• Language models operate on embeddings at every internal layer

The representation is not a preprocessing detail. It is where most of the capability actually lives.

The part that is easy to miss#

An embedding only knows what it was trained on. If the training data treated two ideas as related, the geometry will treat them as related too — including when that relationship is an accident, a bias, or an artefact of the data rather than a fact about the world.

Embeddings do not encode meaning. They encode the patterns of whatever they were shown.