Follow up: GPT-2's vocabulary as a hyperbolic tree — 32,070 tokens in a Poincaré ball you can fly through [P]

Explain Like I'm Normal

A developer created an interactive 3D visualization of GPT-2's vocabulary using hyperbolic geometry to better represent the naturally hierarchical structure of token embeddings. Unlike 2D maps that compress data poorly, this Poincaré ball allows you to navigate through 32,070 tokens as if they were a vast, interconnected forest. It runs directly in the browser and uses Möbius translations to let users explore semantic relationships via their phone.