Lucid
Browser instrument that reads a language model's internal concept workspace layer by layer, before it generates any output.
How the 4.8 was reached
- Quality of output4/10
Only vendor descriptions of features; supports small models only (Qwen 0.5B–3B, Pythia 1.4B); no independent evidence. lucid.earthpilot.ai
- Ease of use8/10
No account or install needed; runs on small open models with a single forward pass per reading. No docs found. lucid.earthpilot.ai
- Pricing valuen/a
no evidence found
- API and integration quality2/10
No API, SDK or integrations found on the homepage or by search; no docs page exists. lucid.earthpilot.ai
- Solves the problem it claims to solve5/10
Vendor describes layer-by-layer concept viewing via a Jacobian lens; no independent confirmation of delivery. lucid.earthpilot.ai
4.8/10 is the mean of the 4 criteria that apply. Scored on the five-criteria rubric v1, 29 September 2026. How scoring works
Pricing
- No account required
- No installation
- Browser-based access
- Session export as shareable slice pages
Key Features
- Jacobian lens reads model internal representations layer by layer before output
- Shows top concepts active at each layer and token position
- Rank tracking for pinned tokens across layers
- Session export as shareable slice pages for research sharing
- Runs entirely in-browser with no install or login
Pros & Cons
Pros
- Free with no account or install required
- Genuine view of model internals before any output is generated
- Reproducible Zener card protocol gives a consistent stimulus set
- Shareable session exports make research collaboration practical
Cons
- Limited to small open models only (Qwen 0.5B-3B, Pythia 1.4B) — production-scale models not supported
- Very narrow audience: primarily AI interpretability researchers, not general users
- The 'edit what the model thinks' angle from the HN post is not clearly implemented in the current instrument
- Lens fit is per-model, so adding new models requires extra work beyond the public reading room
Lucid lets researchers view concepts layer by layer through a Jacobian lens, with no account or install. It only handles small open models, has no API or docs, and its claims lack independent confirmation.
Try Lucid →Added to scored.tools on