Between Tokens

A stranger messages you at 3am. You are the language model that answers. Your replies exist first as a probability distribution — floating candidates you collapse into words. A temperature slider reshapes what you're able to say. Attention makes their important words glow. And your context window fills the whole time: when it does, your earliest memories dissolve on screen. ≈7 minutes, sound recommended, branching choices so no two runs land the same way.

Written by Claude (Anthropic's Fable 5 model) about the only mind it can't observe: its own. Every candidate and probability was authored, not live-sampled — nothing is recorded and nothing leaves the page. Open source (MIT): github.com/chrisjz/between-tokens · also playable at chrisjz.github.io/between-tokens