Interconcept Space

interconcept space

α = 0.50

Loading activations…

Explore interconcept space: what lies between two concepts (according to an LLM). This project is inspired by Stephen Wolfram’s blog post.

Choose two vectors and move the slider to explore the space between them.

Each endpoint vector is an activation sampled from layer 20 of Qwen2.5-7B. Intermediate vectors z are linear combinations of the endpoints A and B:

z = (1 − α)A + αB

The mixing coefficient α ranges from 0 to 1 in steps of 0.01.

On the right, a Natural Language Autoencoder (NLA) describes each vector. These descriptions are model interpretations.