Orion Reblitz-Richardson

Independent alignment researcher · Distiller Labs

Orion Reblitz-Richardson

I build tools for measuring how language models represent moral content and whether their decisions and actions follow it, from pretraining checkpoints through post-training. My current work is DeepSteer, a PyTorch-native toolkit that tracks how these representations form during training and steers them when they drift.

Previously, I spent nearly a decade at Meta where I led interpretability research and PyTorch infrastructure teams. I co-authored Captum, one of the most cited interpretability libraries in the PyTorch ecosystem, and contributed to the launch of the PyTorch Foundation with the Linux Foundation. I also worked on ExecuTorch, Meta's on-device inference framework.

I hold an M.Eng. and B.S. in Electrical Engineering & Computer Science from MIT.