Tom McGrath

Also known as Thomas McGrath, Thomas Michael McGrath

Tom McGrath is an interpretability researcher and AI lab leader whose work examines model internals, from AlphaZero to language models. Before joining DeepMind in 2019, he completed a PhD at Imperial College London with Nick Jones and Kevin Murphy; at DeepMind, he founded the mechanistic interpretability team and worked on AlphaZero and language models through late 2023. He coauthored studies of chess knowledge in AlphaZero, self-repair in language models, and copy suppression in GPT-2. In 2024, he co-founded Goodfire, where he is Chief Scientist; the company announced a $150 million Series B in February 2026, and he discussed its Intentional Design approach on a March 2026 podcast.

Now at: Goodfire
Badges: Founder
Links: Google Scholar, X, LinkedIn and GitHub

Timeline

By 2019

A PhD at Imperial

Before joining DeepMind in 2019, he completes a PhD at Imperial College London with Nick Jones and Kevin Murphy.
2019–2023

DeepMind's interpretability team

From 2019 to late 2023, he works at DeepMind on interpretability for large language models, AlphaZero, training-data science, RLHF data quality and self-annotation, and evaluation of generalist deep reinforcement-learning agents. As a senior research scientist, he founds DeepMind's mechanistic interpretability team.
2021–2022

Chess knowledge in AlphaZero

He is first author of Acquisition of Chess Knowledge in AlphaZero, submitted to arXiv in November 2021 with Demis Hassabis and chess grandmaster Vladimir Kramnik among its coauthors. Published in PNAS in November 2022, the paper uses linear probes to show when and where human chess concepts are represented in AlphaZero’s network.
July 2023

The Hydra Effect

In July 2023, he is first author of The Hydra Effect: Emergent Self-repair in Language Model Computations with Matthew Rahtz, Janos Kramar, Vladimir Mikulik and Shane Legg. The study shows that ablating one attention layer can cause another to compensate.
October 2023

Copy Suppression in GPT-2

In October 2023, he coauthors Copy Suppression: Comprehensively Understanding an Attention Head with Callum McDougall, Arthur Conmy, Cody Rushing and Neel Nanda. The paper shows that GPT-2 Small’s L10H7 attention head suppresses a token when it has appeared earlier in the context.
August 2024

Co-founding Goodfire

In 2024, he co-founds Goodfire with Eric Ho and Daniel Balsam and serves as its Chief Scientist. The San Francisco AI interpretability company announces a $7 million seed round led by Lightspeed Venture Partners in August 2024. After leading a research team at Google DeepMind, he goes all in on Goodfire.
February 2026

Goodfire’s $150 million Series B

In February 2026, Goodfire announces a $150 million Series B at a $1.25 billion valuation led by B Capital. The company identifies his AlphaZero paper as its first proof of concept for extracting novel chess concepts from a model and teaching them to a grandmaster.
As of October 2026

Chief Scientist at Goodfire

He is co-founder and Chief Scientist of Goodfire, an AI interpretability research lab in San Francisco. In March 2026, he joins Dan Balsam on The Cognitive Revolution podcast to discuss Goodfire’s Intentional Design approach, geometric latent-space methods, and reducing hallucinations.