ByteDance-Sued PhD Launches $200M Video World Model Startup with 100M-Hour Dataset

Background and Startup Overview

According to a Bloomberg report, former ByteDance intern and Peking University PhD student Tian Keyu has publicly disclosed plans for a stealth AI startup focused on world models . The company, currently unnamed and without a public product, employs roughly 10 people, including several former ByteDance employees.

Funding and Valuation

The venture has secured nearly $30 million in funding from 5Y Capital and IDG Capital , implying a post-money valuation of approximately $200 million .

Technical Approach: Video World Model

The core technical direction is a video-oriented world model . The team plans to train the model on roughly 100 million hours of video data while aiming to reduce video generation costs by an order of magnitude (i.e., to one-tenth of current levels).

A key architectural element is a system of approximately 200,000 visual symbols . Tian has argued that human language cannot perfectly describe a video, so the team intends to first create a "visual language" for AI and then feed the massive video corpus into that representation.

Timeline

The full model is slated for release in 2027 . Before then, the team intends to showcase technical progress through public events.

Research Foundation: VAR (Visual Autoregressive Modeling)

Tian's prior work provides the research bedrock for this effort. He is a PhD candidate at Peking University advised by Professor Wang Liwei, with four top-tier conference papers, including two first-author papers at NeurIPS. His most prominent contribution is VAR (Visual Autoregressive Modeling) , published as a first-author paper at NeurIPS 2024 and awarded NeurIPS 2024 Best Paper .

VAR introduces a next-scale prediction paradigm for visual generation, moving away from traditional next-token prediction. The paper demonstrates that visual autoregressive models exhibit scaling laws analogous to those observed in large language models, suggesting that performance improves predictably with model and data scale.

Competitive Landscape

The article contrasts Tian's video-centric, visual-symbol approach with World Labs , founded by Fei-Fei Li, which focuses on spatial intelligence and 3D world understanding . While both operate in the world-model space, their technical routes differ: World Labs emphasizes 3D environment comprehension, whereas Tian's startup bets on large-scale video and a discrete visual vocabulary.

Past Controversy

In 2024, during his ByteDance internship, Tian became embroiled in a dispute over model-training compute resources. He admitted to terminating another intern's running program to free resources for other researchers, describing the act as "acting on behalf of heaven." A civil lawsuit filed by ByteDance resulted in a court ruling that Tian breached his contract, ordering him to pay 500,000 RMB in damages and forfeit all internship wages.

Current Status

Two years later, Tian has re-entered the AI industry. This disclosure marks the first public view of his team's technical direction: video as core training data, visual symbols as the foundational representation, and a explicit goal of slashing video world model generation costs by 10× . The claims remain unverified pending product release.