Neel Nanda

Neel Nanda has helped turn mechanistic interpretability from a specialist pursuit into a practical research field. Through open tooling, explanatory teaching and large-scale internal-feature maps, he works to reveal what language models represent, how they compute and when their stated reasoning diverges from what they know.

Neel Nanda leads Google DeepMind’s mechanistic-interpretability team. A Cambridge mathematics graduate, he entered the field through independent research and built TransformerLens, an open library that lowered the practical barrier to inspecting Transformer internals. His recent work moves interpretability from small demonstrations towards contemporary models. In December 2025, his team released Gemma Scope 2, a public suite of sparse autoencoders and transcoders covering every layer of Gemma 3 models from 270 million to 27 billion parameters. The team reported that the release required roughly 110 petabytes of stored data and involved training more than one trillion total parameters. Nanda also co-authored two 2025 studies on hidden knowledge. A proof-of-concept Taboo model showed that black-box prompts and mechanistic methods could recover a concealed word. A later benchmark trained three model families to apply knowledge they denied possessing; prefill attacks worked best, while logit-lens and sparse-autoencoder methods improved auditor success but were less effective. He additionally joined a 29-author review defining open problems in the field. The record is notable for tools and limits together: make models inspectable, then state clearly what inspection still cannot establish.