Embedding Space Navigator

Pick a topic and a prompt, and watch where the answer comes from. AI answers run on nearness, not truth.

Embedding Space Navigator

Watch a prompt steer an AI through

The model doesn't look up true answers. It moves toward ideas that sit near your words.

Vague

What should I do about security?

  1. Read the input
  2. Find the target
  3. Pick each word ↺
  4. Done

Loading the cloud…

  • Security (target)
  • Networking
  • Development
  • Governance

For the nerds

More settings to play with

How the model picks

The picture

What's simplified

Real models place words in thousands of dimensions; this squeezes them into three, and the positions here are placed by hand. Real models also don't use bubbles: they weigh every earlier word against every other (attention). The bubbles and the re-read pulse stand in for that.

A teaching model, not a measurement. It shows why context steers an AI's answer. It doesn't show how a real model computes.

Your data stays with you. Everything runs in your browser. Nothing you pick is sent to or stored on our server. (The site uses standard analytics to count page visits.)

Version 1.0 · what's new

Version 1.0 · 2026-11-02

First public release, after four review rounds.

  • Pick a topic set (cybersecurity by default) in the tool or with ?topic= in the link.
  • Each answer word feeds back into the input, and the aim is worked out again before every pick.
  • Themed 3D clouds, a follow camera, phases and a conversation mode.
  • A brightness slider in the player bar.