July 9, 2026
Mind reader or mind game?
Show HN: I built a web tool to see and edit what an AI thinks before it answers
This AI mind-reading demo wowed nerds, but the comments instantly turned into a trust battle
TLDR: Lucid is a new web tool that claims to show what an AI is leaning toward saying before it speaks, making AI behavior easier to inspect. Commenters were torn between being impressed by the sci-fi demo and demanding proof that it actually beats existing tools.
A new browser tool called Lucid is promising something that sounds straight out of sci-fi: type a prompt, and it shows what an artificial intelligence system seems to be "thinking" before it gives its final answer. The creators frame it like a digital mind-reading machine, complete with spooky experiment vibes, old-school psychic test cards, and a big invitation to "read the machine's mind." And yes, the community absolutely pounced.
The biggest reaction was a split between "this is fascinating" and "okay, but compared to what?" One commenter immediately dragged in the competition, basically saying: if you're going to make a big claim here, at least link Anthropic's version so people can judge for themselves. Another jumped in with the serious explainer energy, arguing that this tool may beat simpler methods on some models but flop on others, which turned the thread into a mini nerd courtroom over whether the trick is truly better or just differently branded.
And then came the classic Show HN side quest: self-promo. A lab member casually popped in to plug another project that gives major "we're testing AI's personality now too" energy, which only added to the mad-scientist aura. The funniest part is that the product page is already theatrical, so commenters didn't need to invent the meme — the whole thing basically arrives wearing a lab coat and whispering, "come watch the robot have thoughts."
Key Points
- •Earthpilot Laboratory introduced Lucid as a browser-based tool for inspecting a language model’s internal concepts before it answers.
- •The tool uses a Jacobian lens to read concepts layer by layer and is framed around inspecting a model’s internal "J-space."
- •The article says users can enter any prompt and observe which concepts enter the model’s workspace and which are never spoken.
- •The public version runs on small open models, with the lens fit once per model and each reading requiring a single forward pass.
- •Supported models listed in the article include Qwen 0.5B–3B and Pythia 1.4B, and sessions can be exported as shareable slice pages.