Michael Pearce
Goodfire
Goodfire
Michael currently works at Goodfire applying interpretability methods to scientific models. He worked on mechanistic interpretability with Lee Sharkey during MATS 6.0. During MATS, he developed methods to decompile more interpretable transformer architectures, used Meta-SAEs to find interpretable features within SAE latents, and applied information-theoretic approaches like minimal description length to guide the selection of SAE hyperparameters and architecture. He has competed a PhD in theoretical physics at Stanford focused on understanding the dynamics of evolving populations and complex ecosystems.
The Summer 2024 cohort marked a significant expansion, supporting approximately 90 scholars with 40 mentors—the broadest mentor selection in MATS history. This cohort incorporated MATS as a 501(c)(3) nonprofit organization, formalizing its institutional structure. The program expanded its research portfolio to include at least four governance mentors alongside technical research streams, reflecting growing interest in AI policy and technical governance work. The 10-week research phase continued in Berkeley, with scholars conducting work across mechanistic interpretability, evaluations, scalable oversight, and governance research.Notable outputs from this cohort include research on targeted manipulation and deception in LLMs trained on user feedback, which was accepted to NeurIPS workshops, and contributions to an AI safety via debate paper that won best paper at ICML 2024. One scholar co-founded Decode Research, a new AI safety organization focused on building interpretability tools.
Michael Pearce
