Thomas McGrath

I’m chief scientist and co-founder of Goodfire. We’re an AI interpretability startup.

I was at DeepMind from 2019 to late 2023, where I worked on:

• Interpretability for LLMs (e.g. the Hydra Effect, Copy Suppression) and AlphaZero.
• Science of training data.
• RLHF data quality and self-annotation.
• Evaluation of generalist deep RL agents.

​