←→ parent / child ↑↓ siblings 1–4 choose T tokens M map G grow
Every time you pick one future out of several, you put information into the text. Choosing 1 of 2 is 1 bit, 1 of 4 is 2 bits, 1 of 8 is 3 bits: log2 of the number of options. The meter adds these up along the path on screen and sets them next to the words the model wrote.
It counts choices, not effort. It treats every option as equally weighted, so it is the most a choice among that many options can carry.
“I could steer the text anywhere I wished without having to write a word.”janus, Testimony of a Cyborg (2023)
janus, who posts as repligate, defined bits of selection in her essay Quantifying curation (2021), crediting the idea to a suggestion by Gurkenglas.
A small base model can run inside your browser and grow new futures from any point in the tree, or from a prompt you write.
Checking what your browser supports…
Top-p is fixed at 0.95. Probabilities shown are the model's own, before temperature. Downloaded once, then kept in your browser's cache. Nothing you type leaves your browser.
Your version becomes a new sibling, marked as written by you. The model's version stays in the tree.