what is jspace fr?

jspace is an interactive web tool for looking inside a large language model to see what it is quietly preparing to say at each layer and each word position, before it actually generates anything. It is an implementation of a method called J-lens, based on an Anthropic research paper about how meaning shows up across the internal layers of language models. The tool works with open source models such as Qwen3 and SmolLM2, or in principle any causal language model hosted on Hugging Face. Under the hood, the method estimates how each internal layer's activity relates to what the model outputs in its very last layer, then translates that internal activity into the vocabulary the model actually speaks in, so you can read out a probability distribution over words even from early or middle layers, not just the final one. Because computing this exactly would be far too expensive, the tool instead uses a randomized estimation technique that is proven, through the project's own tests, to converge to the correct answer without bias. To use it, you install the project's dependencies, then start a local web server. In the browser you pick a model, type in some text, and run the analysis. If a model has not had its internal statistics precomputed yet, the tool falls back to a simpler comparison method called the logit lens while you kick off the background precomputation, which can take anywhere from about thirty minutes to several hours depending on model size, and automatically checkpoints its progress so it can resume if interrupted. There is also a command line option to run this precomputation directly. The web interface shows a grid where rows represent layers and columns represent word positions, with each cell showing the most likely word at that point and a background color reflecting its probability. Clicking a cell reveals a ranked list of alternatives, and you can flip between the new J-lens method and the older logit lens to compare them directly. Running this tool requires a fairly capable machine: Python 3.12 or newer, the uv package manager, and ideally a CUDA capable graphics card, since precomputing on a CPU alone is very slow. The project includes a test suite that checks the math and the caching behavior.