
Description
Lucid is a browser instrument for reading what a language model holds in mind before it answers. It applies the Jacobian lens to a model's internal representations, which Anthropic calls the J-space, and shows which concepts enter the model's workspace at each layer. You type a prompt, and Lucid shows the top concepts for each layer and position, along with concepts the model holds but never says. Rank tracking follows pinned tokens through the layers. The public instrument runs on small open models, including Qwen 0.5B to 3B and Pythia 1.4B, with more available on request. The lens is fitted once per model, and reading a prompt costs a single forward pass. Every session can be exported as a shareable slice page. Lucid is part of Earthpilot Laboratory and presents itself as a research project, with sections for explore, write, research and methods. Its card test is inspired by the Zener cards used in early studies of mind reading. No account or install is needed, and the reading room is open to anyone. It suits researchers and curious readers who want to see how a model represents concepts inside.…
Highlights
- Reads concepts layer by layer
- Runs on small open models
- Rank tracking for pinned tokens
- Shareable session pages
- No account or install
Ratings & Reviews
No ratings yet
Be the first to rate Lucid.
People also visit
Lucid FAQ
What is Lucid?
A browser tool that shows the concepts a language model holds before it answers, layer by layer.
Which models can it read?
Small open models, including Qwen 0.5B to 3B and Pythia 1.4B, with more on request.
Do I need an account?
No. The site says there is no account and no install.
Who makes it?
Earthpilot Laboratory, which presents Lucid as a research project.
What are the best Lucid alternatives?
Popular alternatives to Lucid include Perplexity and Ecco. See all Lucid alternatives










