Shengshu says world models, not LLMs alone, will power video, robots and agents

生数科技 · wechat · 2026-07-22

Shengshu AI argues world models are the next major capability jump after LLMs

At a WAIC 2026 forum, Shengshu AI CEO Luo Yihang argued that once AI moves into video, games, robotics, and autonomous driving, language understanding alone is no longer enough. He framed world models as the next LLM-scale paradigm: systems that can perceive what is happening, predict what will happen next, and choose actions based on feedback.

Why world models matter

LLMs are good at language, knowledge, and information.
World models target an agent’s interaction with the physical and digital world.
Key requirements include:
a unified representation of the world
continuous perception, prediction, and action loops
transfer across tasks and embodiments

Luo said a true world model should not be treated as just video generation, 3D reconstruction, or robot control, but as a general foundation that can scale across scenes and devices.

Shengshu’s product roadmap

Shengshu tied this view to three product lines built on one shared base model:

ViduQ for digital content generation, focused on人物、场景、运动 and physical consistency
ViduS for real-time, continuous interactive video experiences such as companionship, interaction, and games
Motubrain for translating environment understanding into robot actions in the physical world

The company’s conclusion is that digital agents and physical agents may look different on the surface, but the intelligence underneath should be universal, transferable, and grounded in how the world changes over time.

Original post →

Shengshu says world models, not LLMs alone, will power video, robots and agents

More from Embodied