About

I work on multimodal interpretability and physical reasoning in video world models.

At Meta, I lead the internal interpretability community—a 150+ member group of researchers and engineers—and work on physically grounded video world models on the JEPA team. I am also completing my PhD at McGill University (Mila).

My research examines how video systems encode physical structure and causality, where those representations fail, and how to make such failures visible inside the model rather than relying on surface-level metrics. This work is motivated by the practical demands of deploying multimodal systems that must reason over physical dynamics, uncertainty, and long horizons, and has been published at leading machine learning conferences and workshops, including NeurIPS, ICML, and CVPR.

I have also served as an area chair, reviewer, and workshop organizer within the mechanistic and multimodal interpretability community.

Previously, I was the CTO at an early-stage startup, where I built and led the engineering team, drove recruiting, and led fundraising and technical strategy through our first institutional round. That experience informs how I think about building durable technology companies, particularly around alignment between research, infrastructure, and incentives.

I studied neuroscience and computer science at Princeton, with research at the Princeton Neuroscience Institute and Janelia Research Campus, where I investigated the visual system of mice. Alongside my research, I make interpretability, world models, and the AI industry legible to wider audiences through social media, with commentary appearing in The New York Times, TIME, and Bloomberg.

Publications and CV

See my official academic website at soniajoseph.github.io for a full list of papers and projects.

Elsewhere

Academic Website
Google Scholar
Twitter / X
LinkedIn
YouTube
Instagram