Latent Space Games is three small browser pieces that treat a language model’s embedding space as a place you can walk into. A model holds every word it knows in a high-dimensional space, where nearby points mean nearby meanings. We never get to stand inside it - hundreds of dimensions, no doors. These are doors.
Each game runs the same hidden machinery, sentence embeddings computed entirely in the browser, and turns it into something you can touch. None of them phone home. The model is the toy.
Embedding Space. You type a word and it becomes the anchor at the center. Every word after it lands at a distance set by its cosine similarity to that anchor - closer means the model thinks they are more alike. There is no goal, only the slow drawing of a constellation: water pulling in tight to ocean, grief drifting out past fire, a name you did not expect landing right on top of another.
Semantic Battleship. A word is hidden and you hunt it by meaning. Each guess is scored by semantic similarity - warmer guesses run hotter and deal more damage - until the secret is sunk. You learn to think the way the model does: not in synonyms but in neighborhoods, circling the hidden word by feel until the sonar runs hot.
Monster Meme Memory. You feed a bot memories and then talk to it. It keeps what you give it and reaches for the nearest stored fragment when you speak. It remembers everything and understands almost nothing - tell it the sea in Beirut was grey, ask it later about the sea, and it hands the line back to you, a little proud, a little off. The gap between holding and recalling is the whole piece.
Together they are a way of playing with how a machine thinks it understands us. Small, fast, made to be opened mid-conversation and closed again. They live in the oulipo.xyz Kitchen Lab, where you can play all three.