Keyu Tian Raises $30M at $200M Valuation for World Model AI Startup
Keyu Tian, a Peking University researcher and lead author of a NeurIPS 2024 best paper, raised nearly $30 million at a $200 million valuation from investors including 5Y Capital and IDG Capital to launch a World Models AI startup. The 10-person company aims to create a novel AI-specific symbolic language to process video content and simulate real-world physical dynamics. This deal reflects a broader venture capital trend of betting heavily on elite AI research talent even before a startup has named products or business models. It also highlights the growing momentum behind World Models as the next frontier for AI applications in robotics, autonomous driving, and interactive video simulation. Tian's startup is building a dictionary of 200,000 visual symbols and training its model on 100 million hours of video, claiming it will reduce video generation costs by at least 90%. Tian was previously an intern at ByteDance who was dismissed and sued over a compute resource dispute, an incident he has since publicly acknowledged as an error in judgment.
## BACKGROUND
A world model in artificial intelligence is a machine learning system that constructs an internal representation of the physical environment to predict how spatial dynamics, objects, and actions evolve over time. Unlike text-focused large language models, world models rely on video and physics simulation to enable autonomous agents to reason and plan without trial-and-error in the real world. Visual AutoRegressive (VAR) modeling, pioneered by Tian and his collaborators, redefines visual generation as a coarse-to-fine 'next-scale prediction' task, greatly improving training and generation efficiency.