Skip to content
Ben Kennedy Explainers · Latent Space Learning
The machine as a friendly device

A model stores
a position.

Watch separate concepts become numerical vectors, settle into a learned coordinate system, form neighborhoods, and answer a similarity query. Then take control and inspect the geometry yourself.

00.0 / 17.0 sec
Reference projection Fast, authored coordinates are active. MiniLM computation is optional.
Preparing the vector instrument Three.js field · GSAP sequence
Receiving concepts
Guided · 3D
The visualization could not load. Open the page while connected to the internet, or serve it from a local web server.
Sequence
Step 01 · Token input

Concepts begin without geometry

Before encoding, each item is a separate symbol. Its useful position has not yet been learned.

Use one finger to orbit, two fingers to pan, and pinch to zoom. Tap a vector to open it in the inspector.

The word is not stored.
The relationship is.

A vector preserves useful relationships as coordinates. Search works because concepts with related meaning tend to occupy nearby directions in the learned space.

Ben Kennedy builds privacy-first AI for regulated industries. Founder of Kennedy Applied Sciences, creator of StratoSort, PhD candidate in Artificial Intelligence.