Lucid Web Tool Lets You Visualize and Edit a Language Model’s Pre‑Answer Thoughts
Lucid Web Tool Lets You Visualize and Edit a Language Model’s Pre‑Answer Thoughts
Lucid lets you watch a language model think before it answers
Takeaway: Lucid is a web‑based tool that visualizes the internal concepts a language model activates layer‑by‑layer before producing an output, using a Jacobian‑based "lens"; this makes the model’s hidden reasoning observable and editable.
What Lucid does
Lucid places a Jacobian lens in the browser. When a user types a prompt, the lens extracts the top concepts from each hidden layer of the model and displays them as a series of cards. The cards reveal which concepts enter the model’s internal "workspace" (the so‑called J‑space described by Anthropic) and which concepts remain hidden and never surface in the final answer.
"Lucid watches a language model think. It reads the concepts a model holds before it speaks, layer by layer, using the Jacobian lens." – Lucid homepage
How the Jacobian lens works
The Jacobian lens computes the derivative of the model’s output logits with respect to the hidden‑state activations. This corrects for the basis shift that occurs between early and late layers, unlike a plain logit lens that assumes a fixed basis. The result is a more faithful mapping of internal concepts to the eventual output.
"The Jacobian matrix corrects for the shift in basis from initial to final layer compared to logit lens which assumes that the residual remains in the same basis across layers." – Comment by @krackers
Supported models and cost
Lucid currently runs on small open‑source models such as Qwen 0.5B–3B and Pythia 1.4B. The lens is fitted once per model, and each reading requires only a single forward pass, making it inexpensive to use.
User workflow
- Enter a prompt – any text you would ask a model.
- Observe the cards – each card shows the top concepts for a specific layer and token position, with rank tracking for "pinned" tokens.
- Edit if desired – the interface allows you to modify the internal representation before the model generates its answer.
- Export – every session can be saved as a shareable "slice" page for later analysis.
No account or installation is required; the tool runs entirely in the browser.
Why the pre‑answer phase matters
Focusing on the model’s internal reasoning before it produces text provides a more practical avenue for alignment than merely tweaking system prompts. By seeing what the model is thinking, users can identify hidden biases or misconceptions early.
"I love the focus on the pre‑answer thinking phase. It feels way more practical than just trying to steer the output with a generic system prompt." – Comment by @mune2gu‑chan
Related projects from Earthpilot Laboratory
- Personality Bench – a companion site (https://persona.earthpilot.ai) that runs personality tests on frontier models, offering another lens into model behavior.
"also from our lab is Personality Bench: https://persona.earthpilot.ai which runs every frontier model through personality tests." – Comment by @ada1981
How to get started
Visit https://lucid.earthpilot.ai, click explore or write, and type any question you would ask a mind. The interface will display the evolving concept cards, and you can interact with the "docent" chatbot to ask for explanations of any term.
Bottom line: Lucid provides an accessible, low‑cost way to peek inside a language model’s hidden workspace, making the model’s pre‑answer reasoning transparent and manipulable for researchers, developers, and curious users alike.