Latent Space
Drag the crosshair on the map · Arrow keys nudge · Space locks in
How Latent Space Works
This is a toy embedding you can walk around in. About 200 common words are scattered across a 2D map, hand-placed so that semantically similar words sit near each other — animals huddle in one corner, foods in another, ML jargon doing its own thing in the southeast. Each round the game secretly picks one word as the target.
- Pick Freeplay (random target) or Daily (everyone gets the same word, derived from today's date)
- Drag the crosshair around the map — or nudge it with arrow keys
- Watch the big WARMTH readout. It climbs as you near the hidden target and screams BURNING when you're adjacent
- Tap Lock In Guess when you think you're on it. You get five guesses per round
- Closer = more points; locking in with guesses to spare earns a bonus
Why Is This Hard?
You have no labels, no minimap of the target, nothing — just a scalar heat signal, which is roughly all an embedding model gets too. You're doing gradient ascent by hand on a loss surface you can't see. Cold means out of distribution; warm means you found the right cluster; burning means you're standing on the token.
Slop Fact: Real language-model embeddings live in hundreds of dimensions, not two, which is why we flattened the universe of meaning into a single square and called it a day. Every distance here is a lossy projection — much like every claim a model makes about understanding you.