Sonia Joseph: Interpretable Foundation Models for the Physical World
Interpretable Foundation Models for the Physical World — with Sonia Joseph
Join us for a conversation with Sonia Joseph, who recently left Meta to build a frontier neolab for the physical world, focused on interpretable foundation models and scientific simulation.
At Meta, Sonia built and led the company’s internal interpretability community while also conducting research on the JEPA team. Her work there included mechanistic interpretability for physical reasoning in video world models. These experiences shaped her conviction that interpretability and the physical world belong together — not only as a way to debug physical foundation models, but as part of how such models can be trained, verified, and trusted throughout their lifecycle.
Sonia’s new lab is exploring foundation models as learned latent-space scientific simulators, with potential applications across areas such as climate science, astrophysics, supply chains, power grids, robotics, and other physical systems. The lab is also focused on safety and verification for physical AI, while supporting open research and stronger connections between industry and academia.
In this session, Sonia will discuss the ideas behind interpretable physical AI, world models, and latent-space scientific simulation, as well as her vision for the new research lab she is building.
We’ll leave plenty of time for an open Q&A with Sonia.
More details:
https://www.linkedin.com/feed/update/urn:li:activity:7487938191652577281/