AI Researcher

Neel Nanda Portfolio

Runs the Google DeepMind mechanistic interpretability team

Runs the Google DeepMind mechanistic interpretability team, after working on language model interpretability at Anthropic under Chris Olah; his site splits 51 numbered personal essays on agency, motivation, and career advice from a separate Mechanistic Interpretability section of paper walkthroughs, starter guides, and a glossary, plus open calls for his MATS research stream.

Mechanistic InterpretabilityTransformersAI SafetyDeepMind

Location

Unlisted — probably wherever the wifi is strongest.

Companies

Google DeepMindAnthropic

Credentials

  • Pure maths undergraduate degree, University of Cambridge (2020)

Skills

Mechanistic InterpretabilityTransformersAI SafetyDeepMind

Notable

  • Created TransformerLens, a library for mechanistic interpretability of language models (3.9k GitHub stars)
  • First author of "Progress measures for grokking via mechanistic interpretability" (2023)
  • Co-author of Anthropic's "A Mathematical Framework for Transformer Circuits" and "In-context Learning and Induction Heads"
  • Mentors interpretability researchers through his MATS stream, a full-time program run twice a year

What makes it work

A breakdown of the choices that make this portfolio stand out.

1

Two archives kept deliberately apart

The numbered posts (up to Post 51) are rough-draft essays on agency, friendship, and advice, written per the About page as an exercise in being less of a perfectionist; the technical work (the Othello-GPT follow-up, Toy Models of Superposition walkthrough, glossary) sits in its own Mechanistic Interpretability section, so neither audience has to wade through the other.

2

Mentorship infrastructure built into the site itself

Recruiting and supervising MATS applicants directly from the homepage turns a personal blog into a functioning pipeline for training new interpretability researchers, not just a publishing outlet.

3

Research strategy writing alongside technical writing

"Good Research Takes are Not Sufficient for Good Strategic Takes" addresses how to choose problems, not just how to solve them — a rarer and higher-leverage kind of writing for a field this young.

What AI Researchers can take from this

Specific, actionable tips to apply to your own portfolio — no generic advice.

  • If you write for two audiences (technical and personal), give each its own section rather than mixing them in one feed.

  • If your field is young, build a visible path (a program, a set of starter problems) for newcomers directly into your site.

  • Write occasionally about how you pick research problems, not only about your solutions — it signals judgment, not just execution.

Projects

Neel Nanda keeps the essays on making friends and beating procrastination in one archive and the reverse-engineered transformers in another, so the self-improvement readers and the interpretability readers never have to meet.

Open SourceadvancedOngoing open-source project

TransformerLens

PythonPyTorchMechanistic Interpretability

A library for mechanistic interpretability of language models: load a model, cache and inspect activations, and patch them to test hypotheses about what components do.

Turned a research workflow into shared tooling that other researchers build on.

  • Listed first under 'Stuff I've made' on his About page

  • Now maintained under its own GitHub organization

  • Built for hooking and patching activations

ResearchadvancedMulti-year research program

A Mechanistic Interpretability Analysis of Grokking

TransformersModular ArithmeticMechanistic Interpretability

Independent research that reverse-engineers a small transformer trained on modular addition to explain why it suddenly generalizes ("grokking") long after memorizing the training data.

A full reverse-engineering of a trained model's algorithm, start to finish.

  • Linked from the blog as an 'Interlude'

  • A companion paper-replication walkthrough sits in the MI section

  • Became a first-author paper

ResearchintermediateSingle long-form reference

A Comprehensive Mechanistic Interpretability Explainer & Glossary

Technical WritingMechanistic Interpretability

A searchable, detailed glossary of mech interp jargon, published December 2022 alongside a starter guide, a prerequisites list, and an annotated reading list.

Onboarding material for a young field, written by someone running a team in it.

  • Searchable format

  • Paired with 'Concrete Steps to Get Started'

  • Linked from both the About page and the MI section

Ready to build your portfolio?

Follow the AI Researcher roadmap — skills, projects, and timeline to get hired.

AI Researcher Roadmap