Ryan Lagasse

I care about understanding AI systems mechanistically—not just observing that they work, but knowing why. The problems I keep coming back to: making interpretability tools useful at training time (not just post-hoc analysis), scalable oversight that actually scales, and architectures that support genuine reasoning and memory. Right now that means interpretability research on looping transformers and building multi-agent planning systems. The through-line is simple: alignment without mechanistic grounding is just vibes, and I want to build the science that changes that.