How a machine turns meaning into numbers

Computers do not understand words; they do arithmetic. So the quiet foundation under modern search, recommendations, and the “memory” in AI systems is a translation step: turn each word, sentence, image or product into a list of numbers — a point in space — arranged so that things with similar meaning sit close together. Those points are called embeddings, and once you see them the rest stops being mysterious.

Picture a map where every word is a pin. On a good map, “Paris” and “London” are near each other, “banana” is off in another region entirely, and “king” and “queen” are neighbours. An embedding is that map, except with hundreds of dimensions instead of two, so it can capture many kinds of similarity at once.

Where do the coordinates come from? Not a dictionary. They are learned from usage, on a simple principle: a word is defined by the company it keeps. Words that show up in similar contexts — “doctor” and “nurse,” “coffee” and “tea” — get pushed toward similar positions, because the model that predicts their surroundings does better when it treats them as related.

Run that over enough text and geometry starts to encode meaning. The famous party trick is that directions become meaningful too: the step from “man” to “woman” is roughly the same step as from “king” to “queen.” It is oversold — it does not work as cleanly as the demos suggest — but it points at something real: relationships live in the layout.

Once meaning is coordinates, hard problems become easy ones. Search becomes “find the nearest points to this query,” which is why you can search by concept rather than exact keyword. Recommendations become “what sits near the things you liked.” And the way an AI assistant recalls a relevant note from thousands is the same move: embed everything, then fetch the neighbours of what you just asked.

The map is only as fair as the text it was built from. If the training data associates certain jobs with certain genders, the geometry inherits that, and the bias rides along into search results and recommendations unless someone catches it. And closeness is similarity, not truth — two sentences can sit near each other because they sound alike while meaning opposite things.

When a system seems to “understand” what you meant — finds the right document from a vague query, recommends the thing you didn’t know to ask for — it is almost always this underneath: meaning turned into coordinates, and a search for what is nearby. Not comprehension. Geometry.