WikiPath

The map below represents a selection of ~250,000 Wikipedia articles, laid out so that articles with similar meaning sit closer to each other. By converting each article into a long array of numbers — a vector — using a language model trained for exactly this, the position of each article in a very high-dimensional space (hundreds of dimensions) actually conveys meaning. The map you see is a 2-D projection of that space, so neighbors on screen are usually neighbors in meaning, give or take some unavoidable squashing.

Even more spectacularly, this lets us do arithmetic on concepts instead of just numbers. The famous example: in a space like this, taking the vector for “King”, subtracting the vector for “Man”, and adding the vector for “Woman” lands you approximately at the vector for “Queen” — the gendered-royalty relationship is a literal direction you can travel.

Each axis below the map is one of those directions, hand-built the same way: average a bundle of phrases describing one pole, subtract the average of phrases describing the opposite pole. Stepping along an axis adds that direction to your current article's vector and snaps to the real article that best matches the result — moving you from where you are to the article that best fits “what you started with, plus the concept on that axis”.

Loading…

Show all axes