Analysis updated 2026-05-18
Explore whether a small language model is thinking about something it never writes down.
Run guided interpretability experiments on Apple silicon without needing a CUDA GPU.
Swap an internal model clue and watch how the model's answer changes as a result.
Fit or apply a Jacobian lens on your own model and text corpus via the underlying library.
| smartaces/jacobian-lens-for-mac | 0marildo/imago | 0xdfi/glm-5.2-1m-4x-dgx-spark | |
|---|---|---|---|
| Stars | 3 | 3 | 3 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | easy | hard |
| Complexity | 3/5 | 2/5 | 5/5 |
| Audience | researcher | general | researcher |
Figures from each repo's GitHub metadata at analysis time.
Requires an Apple silicon Mac with 16GB RAM and downloads a ~9GB model on first live run.
This project is a Mac friendly version of Anthropic's Jacobian Lens, a research tool for peeking at what a language model seems to be thinking about internally before it finishes answering. The original version from Anthropic only runs on Nvidia CUDA graphics cards, so this fork rewrites it to run on Apple silicon Macs using Apple's own Metal based acceleration, and wraps it in a simple point and click local web app powered by a small AI model called Qwen. The web app walks a curious but non technical user through four guided experiments. The Focus Test shows whether you can tell what the model is thinking about even when it never says so out loud. The Missing Step Test looks for a reasoning step the model skips writing down. The Early Answer Test tries to see the model's answer before it actually states it. The Silent Warning Test looks for a problem the model notices internally but never mentions in its reply. The Missing Step tab also includes a cause and effect experiment, where you can swap one internal clue for another, such as changing a hidden association from spider to ant, and watch the model's final answer change as a result, offered as evidence that the clue genuinely influenced the answer rather than being a coincidence. To try it, you need an Apple silicon Mac with 16 gigabytes of memory recommended, about 10 gigabytes of free disk space, and Python 3.10 or newer. The app works immediately using saved example results, and only downloads the roughly 9 gigabyte Qwen model the first time you choose to run something live, asking for confirmation first. Beyond the web app, the repository still includes the original Jacobian Lens library for people who want to use it directly in their own code, including functions to apply an already trained lens to a model or to fit a new lens on your own model and text corpus. The project's own code is released under the Apache License 2.0, a permissive license that allows commercial and non commercial use as long as you keep the required notices. It downloads the Qwen model and lens separately from other sources with their own licenses.
A Mac-native, point-and-click explorer that lets you peek at what a small AI model seems to be thinking, ported from Anthropic's CUDA-only Jacobian Lens research tool.
Mainly Python. The stack also includes Python, PyTorch, Apple Silicon.
The project's own code can be used for any purpose, including commercially, as long as you keep the required license notices.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
Verify against the repo before relying on details.