wtf is jspace-viz?

jspace-viz is a visual tool for looking inside open weights language models while they process text. It lets you type any prompt and watch, layer by layer, what the model seems to be thinking at each word position, before it produces its final answer. There is also a free online demo you can try in your browser without installing anything. The tool is built on an idea from a 2026 research paper about how language models may hold a shared, readable set of concepts across their internal layers, something the paper calls a global workspace. To make those inner layers readable, the tool uses a technique called the Jacobian lens, which translates the model's internal signals into words using a mathematical transformation learned from example text. This is more accurate than an older method called the logit lens, especially for early and middle layers of the model. The tool lets you switch between the two methods to compare them side by side. When you type a prompt, the screen shows a grid where each column is one word or token of your input and each row is one layer of the model. Each cell in the grid shows what word the model would say if it had to answer right at that point, colored to reflect how confident it is. A separate panel highlights concepts the model appears to be holding in mind that are not simply words copied from the prompt or the next word it will say. Clicking any cell shows a detailed ranking of possibilities, and charts track how word like or concept like the model's internal state is at each layer. The project works with several existing language models, including small ones like GPT-2 and larger ones such as DeepSeek and Llama variants, and it includes scripts for fitting the required lens data to a new model on a personal computer. It is aimed at researchers and technically curious people who want to explore how language models process information internally, and setup requires some comfort with Python and command line tools.