wtf is j-studio?

J Studio is a desktop application, built in Python, for looking inside how a language model "thinks" and testing changes to what it says. The tool is aimed at researchers who want to poke at decoder only language models, meaning models that generate text one word at a time, like the ones behind many chatbots. Its workflow borrows from the classic PC tool Cheat Engine: pick a running model session, scan for concepts inside the model's internal representations, narrow down results, then queue up changes such as injecting a concept, replacing one, or suppressing one. The interface wraps that scanning process in a modern dark themed workbench. Advanced options stay hidden until needed, and the current session and model details are shown as status labels so it is clear what is being tested. Chat, a rules editor, and an interactive visualization panel called J-Lens all share one visual style. Under the hood the tool talks to a real Hugging Face and ROCm backend for running models, alongside a demo mode that produces the same interface with fixed, repeatable data for testing. PyTorch and Transformers, the libraries used to load and run the models, are kept separate from the interface code. Features include a streaming chat view that can be paused and resumed, a table of detected concepts with activation charts, editors for injecting, replacing, or suppressing concepts with automatic strength tuning, and saved projects in a versioned JSON format. A rules editor runs checks in an isolated JavaScript sandbox process for safety. To run it locally you need Python 3.11 or newer, then install with pip and launch the module. It defaults to a specific cached model checkpoint, though other Hugging Face models can be used with a flag. The project only supports decoder only models for now, but the code is written so more model types could be added later. This is described as an independent research tool with no official connection to Anthropic or any cited researchers.