Jason Bohrer

I lead safety alignment for Muse Image and Muse Video at Meta Superintelligence Labs. I develop post-training methods to align our image and video generation models with Meta's safety policies, focusing primarily on the agentic planner model. I'm interested in designing alignment recipes that are robust to bias from other training objectives. I've built agentic data generation pipelines to support more data-efficient RL and more challenging evals. Previously, I led the development of safety classifiers to enable the launch of Muse Image.