Rumi Salazar
I'm a PhD student at the University of Melbourne studying singular learning theory and linear logic for AI alignment, supervised by Daniel Murfet and Nora Ganter. I collaborate with Timaeus (now Resolution), a research organisation aiming to make fundamental scientific progress on technical AI alignment.
Currently I'm focused on Turing machines, and more recently Boolean circuits, in the context of singular learning theory and differential linear logic. The Turing machine work is described in the update below. More broadly, I'm interested in agent foundations and the science of deep learning to answer fundamental questions about alignment that matter for reducing existential risk from powerful AI systems.
Before my PhD I studied at UNSW Sydney, where I wrote an honours thesis on the relationship between symplectic reduction (a quotient in symplectic geometry) and the GIT quotient in algebraic geometry, supervised by Daniel Chan. Prior to that, I investigated the question can one hear the shape of a drum? in a three-month research project supervised by Michael Cowling.
Updates
September 2026. My new paper, Interpretability for Turing Machines, with Billy Snikkers, Daniel Murfet and Will Troiani, is on the arXiv. We show that susceptibilities, an interpretability technique developed for neural networks, can detect algorithmic structure in Turing machines. This work builds upon a long chain of prior work: Turing machines are encoded as proofs in linear logic, then modelled in vector spaces, which is a semantics for differential linear logic. That gives derivatives of Turing machines for which one can apply the tools of calculus. Singular learning theory then turns this into a geometry of program synthesis, in which programs are singularities of a real-valued function over a space of noisy Turing machines. This is the setting of our paper.