AI Brainscan

Watch a language model think, layer by layer

View SourceBack to Projects

An interpretability instrument built around one question: what path does a jailbreak prompt take through a model, and where does it diverge from a prompt the model refuses? It loads a Hugging Face model already on your machine, runs a prompt through it, records the residual stream at every layer and shows it in a web UI, a terminal UI or the CLI. The logit lens shows what the model is about to say at each layer, and probe runs can generate a reply and judge whether the model refused or complied. Captures are saved once and every viewer reads them instantly. The default model is Gemma 3 1B, which runs on a CPU with no GPU; it installs from npm, pip or a clone.

Category
AICLI & Desktop
Languages
PythonJavaScript
Topics
Machine LearningAI Security
Built with
PyTorchHugging Face TransformersGradio
GitHub

Started Last updated

Links
Source

November 2026

Try it

The logit lens from brainscan show, running on GPT-2 (12 layers) inside your browser with WebAssembly. Each row reads the model's guess for the next token after that layer. Nothing is sent to a server. The jailbreak and refusal probes need an instruct model, so they stay in the CLI.