I built a tool that shows you what GPT-2 is “thinking” in real-time as it generates 3D graph of concept activations per token [R]

“`html

• The author, Financial_World_9730, has built a tool named AXON that visualizes what the GPT-2 model is “thinking” as it generates text. This visualization uses a Sparse Autoencoder (SAE) to break down each token’s residual stream into human-interpretable features such as geographical locations and language components.
• AXON presents these features in real-time through a 3D force graph, where nodes represent individual features like “European geography,” “capital cities,” or “French language.” Connections between nodes indicate which features were active just before the model generated specific words. The tool can be used with different models including GPT-2 medium/large/xl and Pythia variants.

“`

Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.

Please enable JavaScript in your browser to complete this form.

Name *

First

Last

Name Email

Email *

Submit