Arthur Conmy

Arthur Conmy is an AI researcher whose work spans mechanistic interpretability, model extraction and AI alignment. He studied mathematics at Trinity College Cambridge, earned upper first class honours, and completed a software engineering internship at Meta in 2021. At Redwood Research, he helped develop ACDC; from 2023 to 2026, he worked at Google DeepMind on Gemini post-training and interpretability tools. In 2024, he coauthored an ICML Best Paper and contributed to Gemma Scope; in 2025, he coauthored SAEBench, a benchmark for sparse autoencoders. As of October 2026, he is a Member of Technical Staff at Anthropic, focused on aligning models during training.

Now at: Anthropic

Timeline

2021

Cambridge mathematics and Meta

He studies mathematics at Trinity College Cambridge and earns upper first class honours. In 2021, he completes a software engineering internship at Meta.
2022–2023

Automating circuit discovery

He works at Redwood Research in 2022–2023. At NeurIPS 2023, he and four coauthors present ACDC, an algorithm that identifies transformer-circuit edges; it selects 68 of 32,000 edges in GPT-2 Small for a Greater-Than task.
2023–2026

Senior Research Engineer

From 2023 to 2026, he serves as a Senior Research Engineer at Google DeepMind, working on Gemini post-training and production interpretability tools.
2024

An ICML Best Paper

In 2024, he coauthors Stealing part of a production language model, which recovers the embedding projection layer of black-box production language models through API access. The work earns an ICML Best Paper award in 2024.
July 2024

Gemma Scope for Gemma 2

In July 2024, he contributes to Gemma Scope, Google DeepMind’s open suite of sparse autoencoders for interpreting Gemma 2. The team also releases Mishax, a tool built for interpretability work.
2025

SAEBench for sparse autoencoders

He coauthors SAEBench, a standardized evaluation suite that measures sparse autoencoder performance across eight metrics, including interpretability and feature disentanglement. It supports comparisons across architectures and training methods.
2026

Mentoring through MATS

He mentors the MATS Summer 2026 stream, which evaluates interpretability techniques for current and future AI safety problems. Its projects include reasoning-model interpretability, eliciting strange behaviors from models in the wild, and model organisms.
As of October 2026

Member of Technical Staff

As of October 2026, he is a Member of Technical Staff at Anthropic, where he works on aligning upcoming models during training.