Why the computer knows dog and cat are alike

If I ask you to place "dog", "cat" and "washing machine" on a map where similar things go together, you do it without thinking: dog and cat side by side, the washing machine far away. The surprising part is that a computer can do the same without anyone explaining what an animal is. This article tells the trick, and carries the map inside so you can test it with your own words.

The trick: turning words into lists of numbers. Every word gets assigned a long list of numbers — in the map below, exactly 384. No person picks them: they come from a model having read enormous amounts of text paying attention to a single thing: which words show up in similar company. "Dog" and "cat" share sentences, owners, vets and sofas, so their number lists end up alike. "Washing machine" lives in other sentences, and its list lands far away.

Similar is no longer an opinion: it is a distance. With words turned into lists, comparing meanings becomes geometry: two words are alike if their lists point in similar directions. Suddenly you can sort, group and search by meaning with the same machinery used to compute any distance. It is a change of category: meaning became measurable.

The map above computes those lists in your browser and flattens them to two dimensions so you can see them. Type "wolf" and watch where it lands. Try "Madrid" and "Monday". Try your own name. Sometimes the result will look bizarre — and that is the lesson too: the model does not know what words mean, only where they tend to appear. Its entire world is text.

This holds up more than it seems. When a search engine understands you despite typos, when a shop suggests something that fits, when an assistant finds the exact paragraph that answers your question in a 200-page document: underneath there are lists of numbers and distances. In my document assistant I use exactly this to fetch the right fragments before answering. The map you just touched is the central piece of that machinery.