Subject
4 entries
Mechanistic Interpretability
Bookmarks
Neuronpedia Gemma Scope: interactive mechanistic interpretability for Gemma 2
Neuronpedia's Gemma Scope microscope lets you scan Gemma 2's internal features using Sparse Autoencoders — activating features, steering behavior, browsing SAE decompositions. Interactive mechanistic interpretability for a production model.
Linear Representations and Superposition in LLMs
A clear explainer on how LLMs encode thousands of features in relatively small embedding spaces via superposition — the Johnson-Lindenstrauss lemma applied to neural representations. Good primer on the theoretical foundations behind SAE-based interpretability.
ARENA Chapter 1.3.1: Toy Models of Superposition and SAEs
ARENA Chapter 1.3.1: an interactive curriculum chapter on Toy Models of Superposition and Sparse Autoencoders — building from the theoretical model to hands-on SAE implementation. Part of the ARENA AI safety education program.
Activation Steering Notebook (concentration_notebooks)
A Jupyter notebook exploring activation steering in neural networks from the concentration_notebooks repo. Saved as a reference in the mechanistic interpretability and model steering space.