World Models & Robotics — Yann LeCun and DeepMind

Yann LeCun left Meta to found AMI and advocates world models and JEPA-style approaches over LLM-only routes. He argued text-only LLMs cannot capture real-world dynamics, stressed joint-embedding predictive architectures (JEPA/I-JEPA) and layered abstractions for efficient world prediction, and positioned video/multimodal training as essential. DeepMind and Google Robotics research emphasized combining VLA (vision-language-action) with simulation data, motion transfer across robot bodies, and the need for scalable robot data (sim + YouTube + generated data). Runway and Emu3.5 releases also signaled advances in video/world prediction models and real-time video editing.
Key insights
LeCun: world models, not LLM-only, required for real-world intelligence
Yann LeCun announced he left Meta to found AMI and argued that LLM-centric approaches cannot learn real-world dynamics from text alone. He promoted world-model architectures (JEPA/I-JEPA) that predict in abstract embedding spaces to ignore stochastic low-level details, and noted video and multimodal training supply richer, denser learning signals than text.
Robotics: VLA + simulation + motion transfer for cross-platform skills
DeepMind and Google Robotics researchers described combining Vision-Language-Action models with simulation data and motion-transfer techniques to enable cross-body generalization. Gemini Robotics 1.5 introduced textual 'thinking' traces for robot plans and motion-transfer to share skills across different robot bodies, while emphasizing touch/force modalities for dexterous tasks.
Emu3.5 and Runway GWM: multimodal world models for next-state prediction
Beijing Institute's Emu3.5 framed multimodal learning as unified token prediction for 'next world state' across text, image and video. Runway released Gen 4.5 and a GWM family (GWM1, GWM Worlds, GWM Robotics) for frame-level prediction, real-time editable avatars, and robotics-focused simulated environments that include failed scenarios for robust learning.
Data and scalability remain the core robotics bottleneck
Robotics teams emphasized data cost as the bottleneck: physical robot data collection is expensive at scale; strategies include scalable simulation, leveraging human videos (YouTube), and synthetic data generation (video models) to provide the breadth needed for real-world generalization.