Arthur Conmy
Research Engineer at Google DeepMind
London, England, Other
Summary
🤩
Rockstar🎓
Top SchoolArthur Conmy is a research engineer at Google DeepMind with eight years of experience in ML research and engineering, focused on AI safety, interpretability, and sparse autoencoders. He has authored and co-authored multiple high-impact papers (including an ICML 2024 Best Paper) and helped open-source compute-efficient Gemma models, demonstrating a knack for scaling research to practical releases. Previously at Redwood Research he contributed to circuit-level interpretability of GPT-2 and on GitHub has improved TransformerLens, refining experiments and test coverage for mechanistic analysis of GPT-style models. A Cambridge-trained mathematician who founded his university's competitive programming society, he blends rigorous theory with hands-on engineering across ML systems and tooling. Notably, his work surfaces both attack/defense insights for production language models and pragmatic engineering fixes that improve experiment fidelity.
8 years of coding experience
2 years of employment as a software developer
Bachelor of Arts - BA Mathematics, Bachelor of Arts - BA Mathematics at University of CambridgeThe Cherwell School
Stackoverflow
13reputation
486reached
0answers
1question
Github Skills (7)
transformer-models10
pytorch10
gpt10
python10
testing9
machine-learning9
inheritance6
Programming languages (13)
Github contributions (5)
TransformerLensOrg/TransformerLens
Sep 2022 - Nov 2022
A library for mechanistic interpretability of GPT-style language models
Role in this project:
ML Engineer
ML EngineerContributions:5 releases, 50 reviews, 594 commits in 2 months
Contributions summary:Arthur's contributions primarily involve modifying and testing an existing library for mechanistic interpretability of GPT-style language models. They refactored code, added tests for experiments, and modified existing experiment configurations, and removed a "hacky" fix. The work suggests a focus on refining the library's functionality and ensuring the accuracy of experiments related to model behavior analysis.
ArthurConmy/Automatic-Circuit-Discovery
Oct 2022 - Feb 2023
Contributions:48 reviews, 74 commits, 107 PRs in 3 months
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.