Working explanations of transformer models and interpretability methods.
The Interpretability Map
The mechanistic interpretability literature as one navigable document: dependency graph, chronology, reading pathway, open problems.
Transformer Explainer
GPT-2 running entirely in your browser. Watch the computation move through attention and the softmax. Adapted from the Georgia Tech Polo Club original (MIT).
Latent Observatory
Search the learned features of real models by concept; inspect activation evidence and neighbouring features.
Circuit Traces
Attribution graphs on an open model: which features caused which, from prompt to prediction, through curated case studies.