The Danger of Jailbreaking Robots: Meta’s Research on Physical Intelligence

Can AI truly understand the physical world? In this interview from NeurIPS 2025, we sit down with Sonia Joseph (‪@meta‬ & ‪@mcgillu‬) to discuss the next frontier of artificial intelligence: Physical Intelligence.

From the limitations of current Video-Language Models (VLMs) to the work being done by the JEPA team at Meta, Sonia breaks down why predicting pixels isn't the same as understanding physics. We also dive into the unique intersection of industry and academia in Montreal and cover a critical safety concern for the future: the very real danger of "jailbreaking" robots.

In this video:

Physical Intelligence: Why current models struggle with causal dynamics.
JEPA & World Models: Moving beyond pixel prediction to abstract representations.
The "Newton VLM": Can AI derive laws of physics from video data?
Robot Safety: Why a jailbroken robot is infinitely more dangerous than a jailbroken LLM.

Timestamps:
00:00 - Intro & Sonia Joseph's Background
00:33 - From Neuroscience to AI Interpretability
01:25 - Can AI Learn Physics? (World Models)
02:24 - Inside Meta’s JEPA Team: Physical Reasoning
03:44 - What are Video VLMs? (Video-Language Models)
04:50 - Working at the Intersection: Meta vs. Mila (Academia)
05:52 - Predictions: When Will We Have the "ChatGPT Moment" for Robotics?
07:11 - The Danger of Jailbreaking Robots

Thank you to ‪@lambda-ai‬ for making this interview series possible!

#NeurIPS2025 #ArtificialIntelligence #PhysicalAI #MetaAI #Robotics #MachineLearning #WorldModels #JEPA