The Address of a Thought: What an Embedding Actually Is

Somewhere in a data center, the sentence you typed this morning — the complaint, the search, the question — was converted into a list of a few hundred numbers. Not compressed. Not encrypted. Located.

Those numbers are an embedding, and if you've hovered around AI at all, you've heard the word deployed like a password to a club you're not in. "We embed the documents." "It's all embeddings under the hood."

In the Open-Book AI series, I introduced the idea in one article: every sentence gets coordinates on a map of meaning, where similar meanings are neighbors. This article is the full tour — because embeddings deserve better than one metaphor in passing. They're the load-bearing idea under semantic search, RAG, recommendations, and half of modern AI. And they can be understood, completely, without one equation.

Coordinates you already trust

Start with a map you believe in. Two numbers — latitude and longitude — pin any spot on Earth. The numbers themselves are meaningless out of context: 23.02, 72.57 tells you nothing until you know the system behind it (that's Ahmedabad). What gives coordinates their power isn't the digits. It's the shared frame — the agreement that near numbers mean near places.

An embedding is exactly this, for meaning. Feed a piece of text into an embedding model and it returns the text's coordinates on a shared map of meaning — typically a few hundred to a few thousand numbers. Each individual number is as meaningless as 23.02. Together, against the shared frame, they say precisely one thing: here is where this thought lives.

The Address of a Thought. That's the whole term. When an engineer says "we embedded the help center," translate freely: we looked up the address of every passage and filed the addresses.

Why hundreds of numbers instead of two

Earth's surface needs two numbers because it varies in two directions. Meaning varies in many directions at once — topic, tone, formality, is-it-a-question, is-it-about-money, urgency, and hundreds more that don't map onto tidy human labels. So the map of meaning needs hundreds of axes to spread thoughts out properly. Squash all of language onto two axes and everything overlaps; give it hundreds, and "refund request" can be near "billing complaint" on the money-axes while staying far from it on the anger-axes.

Nobody can picture a map with hundreds of directions — and here's the liberating secret: nobody needs to. Every operation we care about reduces to near or far, and near-and-far behaves exactly the same in eight hundred dimensions as on a paper map. You've spent your whole life navigating without picturing longitude lines. Same skill.

Three properties fall out of this design, and each one quietly powers something you use:

Everything gets an address — regardless of length. A word, a sentence, a paragraph: each lands somewhere on the same map. That uniformity is why one mechanism can search tweets, contracts, and support tickets alike.

Questions and answers share the map. "How do I get my money back?" and the refund-policy paragraph are different kinds of text — but they're pinned to the same geography, and they land near each other. This single fact is what makes retrieval work: your question is a location, and answers are its neighbors.

Distance is computable. Given two addresses, "how near?" is a mechanical calculation a computer does millions of times per second. Meaning-similarity — the thing the Vocabulary Wall said was impossible — became arithmetic.

The address is not the thought

One honest boundary, because it saves you from the hype later. An embedding captures where a thought lives, not everything the thought says. Two sentences can share an address while differing in ways the map wasn't built to separate — Part 1 of the parent series' negation trap ("can be refunded" / "cannot be refunded") is the famous example, and Part 6 of this series is a full tour of these blind spots.

A useful mental model: the address tells you what the text is about with uncanny precision, and only loosely what it claims. Aboutness is geography. Truth is not. Systems that forget this difference retrieve confidently wrong passages — which is exactly why serious pipelines add a careful second reader downstream (the Open-Book series' "Second Interview").

The catch

The map is only as good as the mapmaker. Every embedding comes from a specific embedding model, and different models draw different maps — different axes, different neighborhoods, incompatible coordinates. An address from one model means nothing on another model's map; mix them in one system and you get gibberish distances. (This bites real teams constantly: re-embedding everything after a model switch isn't optional housekeeping, it's the difference between a working system and quiet nonsense.)

And who taught the mapmaker that "refund" belongs beside "money back"? Nobody did — which is either the most beautiful or the most unsettling part of this whole story, depending on your mood. That story is next.

The one line to remember

An embedding is the address of a thought — a few hundred numbers that say nothing individually and, together, say exactly where an idea lives on a map every sentence shares.

Next — Part 3: "Meaning by Company." Nobody taught the model that 'refund' means 'money back.' How the map drew itself — from nothing but watching how humans use words.