resources

Mechanistic Interpretability

What is it?

Mechanistic interpretability focuses on understanding how a model processes information internally to produce its outputs. This involves analyzing the internals of the model, such as the behavior of neurons, attention heads, and weight patterns, to identify interpretable mechanisms or algorithms that explain the model’s decisions.

https://en.wikipedia.org/wiki/Mechanistic_interpretability

Resources

https://www.anthropic.com/research/team/interpretability

https://transformer-circuits.pub/2021/framework/index.html

https://transformer-circuits.pub/

https://transformerlensorg.github.io/TransformerLens/

https://distill.pub/2020/circuits/zoom-in/

Code / GitHub resources

https://github.com/gauravfs-14/awesome-mechanistic-interpretability

https://github.com/TransformerLensOrg/TransformerLens

https://github.com/callummcdougall/sae_vis

Papers

Bridging the Black Box: A Survey on Mechanistic Interpretability in AI