The curator's reviewWord-association Tetris against a language model. Type a clue and the AI reorders the falling blocks by how strongly it associates them with your word. Think fast, type faster.
Semantris is word-association Tetris against a language model. Blocks fall, each carrying a word; you type any word you like, and Google's semantic-similarity model (2018-era universal sentence encoding) reorders the stack by how strongly it associates each block with what you typed. Associate hard enough and blocks clear. Think fast, type faster.
The gameplay is secretly a conversation. You are not typing definitions; you are probing what the model believes goes together, and it believes strange, wonderful things. Ocean reliably clears fish - fine. But the community discovered universal solvents: words like moist, mother, or death that light up half the board, because the embedding space runs thick with associations around them. Finding these is the metagame.
Played for twenty minutes, it becomes the best embeddings tutorial ever built. You feel the geometry: synonyms are close, categories are close, emotional register is a direction, and some words are simply well-connected celebrities of semantic space. Every intuition you need for understanding modern retrieval systems is available here as a reflex.
Now hosted at experiments.withgoogle.com after Google Research's own page went dark - a survivor of the great Google Experiment sunsetting, which nearly claimed it along with Blob Opera's old home. Playing a word game against an embedding space is like arm-wrestling a dictionary, and the dictionary is friendly. Bring your best universal solvent.
