Neel Nanda
Runs the Google DeepMind mechanistic interpretability team; creator of the TransformerLens library
- Runs the mechanistic interpretability team, Google DeepMind12
Neel Nanda runs the Google DeepMind mechanistic interpretability team, which works on reverse engineering the algorithms and structures that trained neural networks have learned. Before that, Nanda did independent interpretability research and worked at Anthropic as a language model interpretability researcher, contributing as a core research contributor to Anthropic's 2021 "A Mathematical Framework for Transformer Circuits". Nanda created TransformerLens, an MIT-licensed library for mechanistic interpretability of GPT-style language models, and is the last-listed author of Gemma Scope, an open suite of sparse autoencoders trained on Gemma 2 models.1435
Work
TransformerLens (external site: github.com)Open-source code
Created the library, which the repository README describes as a library for mechanistic interpretability of generative language models, "created by Neel Nanda"; it is released under the MIT license.31
Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2 (external site: arxiv.org)Published research
Last-listed author of the 2024 report introducing Gemma Scope, an open suite of JumpReLU sparse autoencoders trained on all layers and sub-layers of Gemma 2 2B and 9B and on select layers of Gemma 2 27B.5
Progress measures for grokking via mechanistic interpretability (external site: arxiv.org)Published research
First author of the 2023 paper, which reverse-engineers the algorithm learned by small transformers trained on modular addition and uses that understanding to define progress measures for studying grokking.6
Earlier roles
This profile covers public professional work only, from the sources below. It is not an endorsement or a ranking. To correct or remove it, use the correction form.