What it is
A web app that explores how a single prompt branches into many ChatGPT completions, visualized as trajectories through embedding space reduced to 3D with PCA.
Gabriel’s notes
MIND is a browser-based visualization that explores how a large language model’s responses branch from a single prompt by repeatedly sampling completions (e.g., “Intelligence is …”) at a high temperature, then mapping each partial phrase as a trajectory through embedding space and projecting that space into 3D using Principal Components Analysis (PCA). It pairs a 3D “cube” trajectory view with a tree view that shows branching probabilities and lets you highlight paths by hovering over words. (Some implementation details are described by the project’s write-up, but the exact on-page documentation at the canonical URL is Unknown / not confirmed because the page content could not be reliably parsed in this pass.)
Quick take: This is the rare AI visualization that actually helps your intuition instead of just decorating your screen with neon spaghetti. If you’ve ever tried to “feel” temperature and token probabilities, this makes it concrete.
I saved this under AI because it’s a hands-on way to understand next-token sampling as a branching process—and to see how “semantic space” gets abused (lovingly) for visualization.
Good fit if you want to:
- Build intuition for why the same prompt can fork into wildly different continuations (especially at higher temperature).
- Explain embeddings + dimensionality reduction to non-technical teammates without immediately losing them.
- See completions as paths (prompt → partial phrase → longer phrase) instead of as isolated outputs.
- Steal interaction design ideas for your own “LLM observability” dashboards (hover-to-highlight is doing real work here).
- Prototype new prompts by quickly scanning the “shape” of the response space.
Pricing snapshot (auto-enriched)
Unknown / not confirmed. I didn’t find pricing info on the canonical page in this pass. It appears to be an openly accessible web app, but whether it’s always free (or rate-limited) is not confirmed.
Work-use / compliance snapshot (auto-enriched)
Unknown / not confirmed. The project write-up describes generating the dataset using the ChatGPT API and embeddings, which implies the usual “your prompts/completions are data” questions apply. What I could not confirm here: whether the web app logs your interactions, whether it stores prompts you try, or what retention/deletion controls exist. If you’re using it at work, assume you should avoid sensitive inputs unless you’ve validated the operator’s data-handling posture.
Alternatives (auto-enriched)
- OpenAI Playground — Good for rerunning prompts and comparing outputs, but whether it can replicate this specific trajectory + tree visualization is Unknown / not confirmed.
- Embedding visualizers (general) — Many tools project embeddings into 2D/3D, but the “branching trajectory from one prompt” framing is the distinguishing feature here; specific best-in-class options are Unknown / not confirmed in this pass.
Before you adopt it:
- Decide whether you want a “wow” demo or a repeatable measurement tool—those are different products.
- If you replicate the method: budget for API sampling costs (hundreds of completions adds up fast).
- Validate privacy expectations before anyone pastes real customer or internal info into a public web app.
Sources
- https://moebio.com/mind
- https://sites.uac.pt/amendes/visualizacao/look-into-the-machines-mind/
- https://moebio.com/
- https://en.wikipedia.org/wiki/Principal_component_analysis
- https://platform.openai.com/