This development coincides with the release of H-JEPA by AMI, a system advised by LeCun and co-authored by Michael Rabbat and others. This first end to end learned hierarchical world model uses a 3 level structure of stacked JEPAs to break coarse goals into concrete subgoals. The approach increased success on the Visual AntMaze benchmark from 18% to 73% and reduced required planning compute by building on the SIGReg/LeWM architecture.
Tuesday, October 20, 2026
Story
Meta's 8B Parameter JEPA Establishes Robot World Model Scaling Laws
AI13 days ago
The capacity for artificial intelligence to power physical robots now follows a predictable growth curve as researchers increase the size of their underlying architectures. Meta has demonstrated this phenomenon by training an 8B parameter JEPA model, according to findings shared by chief AI scientist Yann LeCun, which establishes that robot world models obey scaling laws similar to those found in large language models.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.