Samuel Marks

I received my PhD from Harvard in May 2023. While at Harvard, I studied p-adic Hodge theory and prismatic cohomology under Mark Kisin.

Nowadays I work in machine learning, specifically mechanistic interpretability of large language models, with the hope that understanding the inner workings of AI systems will somehow help with aligning them to human values. More specifically, I am a postdoc at Northeastern University under David Bau.

email: smarks [at] math [dot] harvard [dot] edu

Research papers

• The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets (with Max Tegmark). Under review.
• Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback (with Steven Casper, Xander Davies, and many other co-authors). Under review.
• Laurent F-crystals and Lubin-Tate (φ_q, Γ)-modules. Under review.
• Derivatives of p-adic Siegel Eisenstein series and p-adic degrees of arithmetic cycles (undergraduate thesis).
• p-adic properties of Hauptmoduln with applications to Moonshine (with Ryan Chen and Matthew Tyler): SIGMA 15 (2019).

Teaching

• Decision theory tutorial (Math 99r, spring 2022).
• Vector calculus and linear algebra II (Math 22b, spring 2021).
• Algebra I: Theory of groups and vector spaces (Math 122, fall 2020).
• Modular forms tutorial (summer 2020).

Exposition

• Prisms and distinguished elements, given at STAGE in spring 2021.
• The action of Frobenius and rationality of zeta, given at STAGE in fall 2020.
• p-adic modular forms à la Serre, given at STAGE in spring 2020.