Words Travel
Two words on the map. You start at one, you have to reach the other. Every step you pick from four neighboring words, and the only signal telling you which is closer is the geometry of semantic space itself.
What it actually is
The map is a 2D projection of a multilingual embedding model (multilingual-e5-large). Every word is a point in that embedding’s high-dimensional space; UMAP projects to the plane you see. Cosine similarity between embedding vectors decides which neighbors sit close. Twelve color-coded biomes are clusters of meaning the embedding settled into on its own: animals, body, food, time, materials, structures, weapons, weather, abstract states.
Cosine over a learned text embedding isn’t an obvious basis for a game. The argument for it is that what cosine captures, in practice, is something close to associative correlation: which words tend to keep each other’s company across the billions of sentences the model trained on. That’s exactly the thing a person does when they think “what’s near this word in meaning?” Walking from “morning” toward “memory” traces one such path through the geometry.
If embeddings are a lower-dimensional shadow of meaning that survives compression into a vector space, then navigating them is a kind of cave-walk: pick the wall closest to the shape you remember, see where it leads.
How to play
You get a budget of moves and need to spend them reaching the target. Each round drops three optional bonus routes sitting in different biomes; visiting them refunds moves and adds score. Most rounds also surface a wild button: a one-shot operation on the embedding itself. Flip to the spatial mirror of where you are. Pick from the cosine-inverse of your neighbors. Hop to a region you haven’t touched.
Hebrew and English both work, toggleable from the menu. The same embedding handles both languages, so the geometry is shared.