Turn the dials on a language model’s brain

mindcontrol runs Gemma 2, a 2.6 billion parameter language model, on your own graphics card. Along the bottom are dials wired into its working memory: turn Pirate up and it starts talking like one, mid-sentence. Every word it writes can be clicked, to see what it nearly said instead and what each of its 26 layers was thinking at the time. Pick another word and the reply splits into a new timeline.

• Your graphics card, not a server every calculation runs in the page, and nothing typed leaves the machine
• Written from scratch a Rust engine compiled to WebAssembly, with twelve GPU shaders written by hand
• Eleven dials directions in the model’s working memory, found by comparing how it writes when asked to be a pirate and when not
• See inside for any word: what it nearly said, its best guess after each layer, and what it was looking at
• Checked, not guessed every layer matched against the reference implementation before a single dial was turned

How it was builtSource on GitHub

The model is a 1.6 GB download the first time, then kept in your browser. It needs a desktop browser with WebGPU, such as a recent Chrome or Edge. Gemma is provided under and subject to the Gemma Terms of Use.