Drag a token. Watch the geometry think.

What you're seeing. Real LLM embedding spaces have thousands of dimensions, not two. The intuition transfers: tokens that appeared in similar contexts during training end up near each other. Drag the orange marker — when it lands inside a cluster, the model treats it as semantically related to the words there, and the nearest-neighbor list updates live. Switch to Vectors mode to see each token rendered as an arrow from the origin — that's closer to what's actually stored in memory (a 4,096-dim vector per token). Switch to Trail mode and your path is preserved as you move; this is how a sentence travels through embedding space token by token. The pulsing animation on the top-1 neighbor is the visual signal that this would be the predicted class if you asked a nearest-neighbor classifier.