llm debugging

Mechanistic Interpretability Myths Put to the Test

For years, the conventional wisdom in mechanistic interpretability was grim: neural networks are tangled knots, and before you can understand one you first...