What is jspace-scope?

jspace-scope is a self hosted web app that lets you watch the inner workings of an open weight language model while you chat with it. Language models silently think about words they never say out loud, and this project reads those hidden words out of the network using a technique from Anthropic research called the Jacobian lens. It shows a live chat where every token you type or the model writes comes with a readout of what the model is thinking about behind the scenes. Beyond just watching, the app also lets you reach into the model's computation and change those hidden thoughts. You can replace a thought, for example swapping France for China so the model answers Beijing instead of Paris, erase a thought entirely, or inject a new one and watch how the reply changes. You can even turn off the model's whole hidden workspace and see it stay fluent but lose its ability to reason step by step. The app has four different views. The simple view is the default and shows a live thought cloud plus supporting metrics. The lab view lets you run two versions of the same prompt side by side to compare what changes. The theater view is a full screen display meant for projectors and demos. The expert view exposes the full technical grid of data for deeper inspection. There are also fourteen built in one click experiments, grouped into watching the model think, editing its thoughts, and stress testing it, each of which explains what happened afterward in plain language. Running it requires Python 3.10 or newer, and ideally an NVIDIA GPU with at least 6 GB of video memory, though it can run on CPU more slowly. A Docker option is also available for those with an NVIDIA GPU and the NVIDIA Container Toolkit installed. The first launch downloads the model and a matching lens file from HuggingFace, which takes a bit of setup time before the app is ready to use.