Research
AI safety research that borrows tools from nonlinear dynamics to study how training changes a model's behavior.
Current project
Does the assistant persona lose stability before reward hacking?
Nov 2026
A pre-registered test of whether a persona-stability probe gives early warning of reward hacking during RL training. Pre-registration in progress; no results yet.
AI safety background
Physics
Background in physics, focused on nonlinear dynamics and complex systems. The current project takes its central tool, critical slowing down, from that field.
Apart Research hackathon
Took part in the Secure Program Synthesis hackathon, May 2026.
Report for Horizon Omega
Wrote a report on AI guardrails for Horizon Omega.
AI Safety Montreal study group
Organizer of AI Safety Montreal's study group on the Iliad curriculum.
Seminar on safe agentic workflows
Gave the seminar “Safe Agentic Workflows: From First Guardrails to Adversarial Oversight”, October 2026.