game · 1 min
The Ranker
Drag eight cities into the order an embedding model would rank them. Then see the true cosine ranking. If you score well, you have the same priors as a 256-d vector trained on OpenStreetMap features. Suspect.
You against a 256-dimensional vector. The vector knows what “like MIT” means in cosine terms; you know what your gut tells you. Drag the cards into the order you think the embedding would output, hit submit, and see how close you got.
Four anchors rotate: MIT (university towns), Times Square (dense urban), Amsterdam centrum (walkable historic core), Boulder (compact mountain-flanked towns). The candidate scores are hand-baked but mirror what the production system actually returns.
If you nail it, this site has more to teach you than I do.