Neel Nanda

Neel Nanda is a mechanistic interpretability researcher who studies neural networks through their internal components. He won bronze for the United Kingdom at the International Mathematical Olympiad in 2015 and gold medals in 2016 and 2017; he completed a pure mathematics degree at Cambridge in 2020 and explored AI safety through internships. At Anthropic and DeepMind, he worked on language-model interpretability, coauthored an RLHF assistant paper in 2022, and joined DeepMind’s mechanistic interpretability team in 2022. He coauthored a 2024 evaluation of sparse autoencoders, created TransformerLens, and presented on sparse autoencoders at Stanford in May 2025. As of October 2026, he is a Staff Research Scientist at Google DeepMind and leads its mechanistic interpretability team.

Now at: Google DeepMind
Works: 77
Badges: Bridge
Links: Website, Google Scholar and X

Publication numbers from OpenAlex, as of 22 October 2026. Citations and h-index are hidden because this OpenAlex record may be incomplete or mix in other authors.

Timeline

2015–2017

International Olympiad medals

Representing the United Kingdom, he wins bronze at the International Mathematical Olympiad in 2015 and gold medals in 2016 and 2017.
By 2021

Cambridge and safety internships

He completes a pure mathematics degree at Cambridge in 2020 after summer internships in quantitative finance at Jane Street and Jump Trading. In 2021, he explores AI safety through internships at Oxford’s Future of Humanity Institute, DeepMind and UC Berkeley’s Center for Human-Compatible AI.
2022

Anthropic RLHF research

At Anthropic, he works as a language-model interpretability researcher under Chris Olah. In April 2022, he coauthors Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback, which uses preference modeling and RLHF to fine-tune language models as helpful and harmless assistants.
2022–2023

DeepMind and grokking research

After independent interpretability research, he joins DeepMind’s mechanistic interpretability team as a researcher in 2022. In 2023, he coauthors Progress measures for grokking via mechanistic interpretability; the paper uses reverse-engineered model components to define continuous progress measures.
2024

Evaluating sparse autoencoders

In 2024, he coauthors Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control with Aleksandar Makelov and George Lange. The study evaluates sparse autoencoders on GPT-2 Small’s indirect-object-identification task and finds they are less successful than supervised features in controlling the model.
May 2025

A Stanford sparse-autoencoder talk

At Stanford on May 1, 2025, he presents Sparse Autoencoders: Assessing the evidence. The talk assesses how well sparse autoencoders work, their biggest problems, and next steps for the field.
By May 2025

TransformerLens library

By May 2025, TransformerLens is a library for mechanistic interpretability of language models that Nanda creates.
Since 2025

Leading mechanistic interpretability

In 2025, he takes over Google DeepMind’s mechanistic interpretability team when its lead steps down. He has no prior team-lead experience. As of October 2026, he is a Staff Research Scientist at Google DeepMind and leads the team, which uses model internals to understand neural networks and make them safer.