Tian Keyu, who is facing an 8 million yuan lawsuit from ByteDance, has unveiled his startup plan: 200,000 visual symbols to reduce video production costs by 90%.

Beating AI News reports that former ByteDance intern Tian Keyu has disclosed his startup project focused on world models. His team has raised nearly $30 million from investors including Source Code Capital and IDG Capital, with a post-money valuation of $200 million. The team currently has around 10 members and has not yet launched any public products. World models are designed to enable AI to learn how objects move and change in the real world, with applications in video generation, robotics, and other sectors. Tian, the first author of the best paper at NeurIPS 2024, previously researched how AI generates images gradually from coarse to fine. Building on this work, his new team has developed a "vocabulary" of 200,000 visual tokens for video—these operate similarly to the tokens used by large language models to process text, allowing AI to handle video in a more computationally efficient manner. The team plans to train its model on 100 million hours of video. Tian states this approach could cut the cost of generating one second of video to one-tenth or even lower than current levels. The team intends to hold a public demo first, with the full model scheduled for release in 2027. He also told Bloomberg that he once shut down another intern’s program, claiming it was to free up computing power for other researchers, and admitted the action was improper. ByteDance fired him in 2024 and subsequently sued for 8 million yuan. Court documents reviewed by Bloomberg show the court ordered Tian to pay 500,000 yuan for breach of contract.