Neel Nanda
Google DeepMind mechanistic interpretability team lead
Explains how to inspect and intervene on model internals, separating a useful hypothesis about a mechanism from evidence that a causal test supports it.
Why read Neel Nanda?
Read Nanda to translate interpretability terminology into an experiment. His explainer distinguishes a feature, a circuit and an intervention on a model’s activations. Begin with those definitions, then ask what would have to change inside the model for its answer to change.
His attribution-patching article makes that question concrete by comparing a clean prompt with a changed prompt and replacing internal activations. The faster gradient approximation is useful for finding hypotheses, but he explains when it fails and why the hypothesis needs a stronger follow-up test. TransformerLens, which he created, provides tools for inspecting and changing activations.
Selected work
-
A Comprehensive Mechanistic Interpretability Explainer & Glossary
What counts as a model feature or circuit? His explainer introduces the vocabulary and gives an intervention a concrete meaning: change an internal activation and continue the model run.
Cite this selection
XFollowList, “A Comprehensive Mechanistic Interpretability Explainer & Glossary” reading notes in the Neel Nanda profile. https://xfollowlist.com/people/neel-nanda#work-08eb1a82b45d
The linked work is the original source; this permalink identifies XFollowList’s reason to read it. Original work published .
Profile-source checking date and references · Dated reading map
-
Attribution Patching: Activation Patching At Industrial Scale
How do you test which internal component affects an answer? His method explanation compares activation patching with a faster gradient approximation and describes cases where that approximation is unreliable.
Cite this selection
XFollowList, “Attribution Patching: Activation Patching At Industrial Scale” reading notes in the Neel Nanda profile. https://xfollowlist.com/people/neel-nanda#work-85aa9cf8692b
The linked work is the original source; this permalink identifies XFollowList’s reason to read it. Original work published .
Profile-source checking date and references · Dated reading map
Projects & roles
TransformerLens
Mechanistic interpretability library created by Nanda, now maintained by its project team; exposes activations for inspection and intervention.
Sources & further reading
Profile sources checked . This describes checking these references, separately from the publication dates of selected works.
- His current own biography says he runs Google DeepMind’s mechanistic interpretability team and links @NeelNanda5.Link to this source note
- Current name and bio agree with his own-site identity and team role.Link to this source note
- The byline identifies Nanda; he credits the earlier Anthropic team for work underlying this personal write-up.Link to this source note
- The current project README credits him as creator and names separate maintainers.Link to this source note
Cite this profile’s source check
XFollowList, Neel Nanda profile sources. Sources checked 2026-10-11. https://xfollowlist.com/people/neel-nanda#sources-checked