Door: mechanistic interpretability¶

Start here¶

• The decoder running example on the IOI prompt.
• Intervene on a site and measure the effect: ablation, the hook check, and the same edit bound live for sweeps and generate().
• See the graph and inspect any operation.

Going further¶

• The mech-interp kit: residual decomposition, direct logit contributions, per-head results, the component edge graph, patching grids.
• Intervention reference and the intervention explainers.
• Recipes: ACDC on IOI, pretrained SAEs, tuned lens, conditional steering.
• Performance: choosing a fast path.

Coming from another tool¶

TransformerLens, nnsight, pyvene, SAELens, plain forward hooks. Where another tool is the better choice: where to go instead.