Embedding Maze
How Embedding Maze Works
You are a single token dropped into a grid of word-rooms. Doors between rooms aren't physical — they're semantic. You may step from one room into an adjacent one only if the two words are related in the bundled association graph. "cat" opens the door to "whiskers", never to "mortgage". The rest is up to your understanding of how concepts cluster in latent space.
- Your token starts on a word and carries that word.
- Tap (or arrow into) an adjacent room. If its word is related to your carried word, you walk in and now carry the new word.
- If the words aren't related, the door stays shut and the room shakes. No teleporting through latent space.
- At a synonym station (the rooms marked swap), tap your own room to trade your carried word for a related one — opening doors that were sealed.
- Reach the highlighted goal word. Clear all ten mazes, each bigger than the last.
Why Is This Hard?
The shortest semantic path is rarely a straight line. You'll dead-end into tantalizingly close-but-unrelated neighbors, backtrack, and learn to think in cosine similarity. Fewer steps is a better embedding of the route. The graph never hallucinates an edge to bail you out.
Slop Fact: Real word embeddings famously claim king − man + woman ≈ queen. This maze claims cat → fur → bear → honey → bee, which is either profound emergent structure or just vibes we hard-coded at 2am. We'll let the interpretability team decide.