Mechanistic interpretability

Opening up language models to see how they work.

This site will host my personal research and blog posts on mechanistic interpretability: the study of what neural networks actually compute, layer by layer, feature by feature.

Large language models are usually treated as black boxes: text goes in, text comes out. Mechanistic interpretability tries to do better, by finding the internal features a model represents, tracing the circuits that connect them, and testing those explanations with experiments that can prove them wrong.

What to expect

• Layer atlases: what each layer of a real model does, measured rather than guessed.
• Circuits: how models carry out specific tasks like arithmetic or copying text, traced with activation patching.
• Features and steering: finding the directions a model uses to represent concepts, and changing its behaviour by adjusting them.
• Robustness: what happens to a model when its weights are pruned, perturbed, or merged.
• Hybrid architectures: interpretability for models that mix attention with linear-attention layers, where many standard tools don't apply directly.

The first posts are in preparation. Everything here is independent work, done on local hardware with open-weight models.