Meaning by Company: How the Map Drew Itself

Here's the fact that should bother you more than it does.

Somewhere on the map of meaning, "refund" sits next to "money back." Nobody put it there. No linguist drew the neighborhood, no dictionary was consulted, no engineer wrote a rule connecting the words. And yet the placement is right — righter, in messy real-world use, than the synonym dictionaries entire teams spent careers curating.

Parts 1 and 2 established what the map does and what an address is. Today: the mapmaker. Because the answer to "who drew the map?" is no one — and understanding how a map draws itself is the closest thing modern AI has to a magic trick explained.

The oldest idea in the newest technology

The principle is older than computers. Linguists compressed it into one line decades ago: you shall know a word by the company it keeps.

Think about how you actually learned most of your vocabulary. Not from definitions — from encounters. Nobody defined "grabbed a coffee" for you; you heard it in enough sentences that its meaning assembled itself. If I write "the borrower must remit the outstanding principal," you might not know "remit" — but from the company it keeps, you know it's something like "pay." You've been doing meaning-from-context your whole life.

Embedding models do exactly this, with inhuman patience: read colossal amounts of text — billions of sentences — and track which words and phrases keep appearing in interchangeable company. "Refund," "reimbursement," and "money back" show up in the same kinds of sentences, surrounded by the same kinds of neighbors ("customer," "policy," "14 days," "request"), doing the same grammatical jobs. So the training process pulls them together on the map. "Refund" and "giraffe" never keep the same company. They drift apart.

Repeat billions of times, and neighborhoods precipitate out of usage — a returns district, a shipping district, a legal quarter — none of them designed, all of them earned.

Meaning by Company, not by definition

Notice what this mapmaker never does: it never asks what a word really means. There's no definition inside the model, no lookup table, no essence. There is only position — earned from company kept. I call the principle Meaning by Company, and it has consequences that explain both the magic and the weirdness of modern AI search.

It handles words no dictionary has met. Slang, product names, typos-that-became-terms — if people use them consistently, they earn correct addresses. The mapmaker doesn't care whether a word is official. It cares how it's used. Language as it's actually spoken, mapped at a scale no lexicographer could touch.

It resolves context for free. "Apple announced record earnings" keeps the company of "quarterly," "shares," and "CEO." "Apple crumble with custard" keeps very different company. Modern embedding models address the passage, not the isolated word — so the two apples land in different districts without anyone enumerating the cases. The synonym-dictionary problem from Part 1 doesn't get solved; it evaporates.

Even languages align. Train the mapmaker across many languages and "refund," "remboursement," and "रिफंड" — words sharing zero letters — land in the same neighborhood, because they keep the same company in their own languages. Multilingual search stops being translation and becomes plain navigation. If any single result convinces you that meaning has a shape, it's this one.

Beauty, meet bias

Now the flip side, stated plainly because this series doesn't do hype.

A map drawn from usage inherits everything in the usage. If the training text discusses certain jobs mostly alongside certain genders, those associations become geography too — real coordinates, faithfully learned, encoding patterns you might not want your search system to act on. Model builders measure and mitigate this, with genuine but partial success. The point for you: the map is not neutral ground truth. It's a portrait of how people wrote — with everything flattering and unflattering that implies.

And Meaning by Company has a built-in blind spot we've met before: words that keep identical company while meaning opposites. "Can be refunded" and "cannot be refunded" travel in exactly the same circles. Company can't tell them apart, so the map can't either. Remember that; Part 6 is where it pays off.

The catch

The mapmaker only knows the company it saw. Usage it barely witnessed — your internal codenames, your industry's acronyms, the term that means something different inside your walls — gets parked on vague guesswork. The map of the public internet is spectacular for the public internet. Whether it's any good for your company's language is a genuinely open question, and it's the entire subject of Part 5.

The one line to remember

Nobody drew the map of meaning. It precipitated out of how humans actually use words — which is why it's brilliant, and why it inherits our blind spots.

Next — Part 4: "The Ruler Problem." Your search returned a similarity score of 0.83. Is that good? What "near" actually means on the map — and why scores only make sense compared to their neighbors.