Core nnsight concepts for neural network interpretability. Use when setting up models, tracing activations, saving values, or making basic interventions on model internals.
Optimize nnsight code for remote execution on NDIF. Use whenever the user is writing, reviewing, or refactoring nnsight code that runs against NDIF (any usage of `remote=True`, `model.session`, or models too large to run locally like Llama-70B/405B), even if they don't ask for "optimization" explicitly — naive remote code is often 100x slower than it needs to be, and almost all of that comes from two mistakes: too many requests and too much downloaded.
Causal intervention via activation patching to identify important model components. Use when determining which layers, heads, or positions are causally responsible for model behavior.
Gradient-based approximation to activation patching for scalable circuit analysis. Use when activation patching is too slow or when analyzing many components simultaneously.
Causal mediation analysis to identify which model components mediate specific behaviors. Use when investigating how information flows through the network and which neurons or layers are causally responsible for outputs.
Decode intermediate layer predictions using the Logit Lens technique. Use when analyzing what a model predicts at each layer, understanding information flow, or visualizing layer-wise processing.
Control model behavior through persistent edits and steering interventions. Use when modifying model outputs, applying steering vectors, or creating persistently modified model versions.