ByteDance seeks 8 million in compensation from Tian Keyu, who publicly reveals startup plan: 200,000 visual symbols to cut video costs by 90%.

动察 Beating AI News: Former ByteDance intern Tian Keyu has disclosed his world model startup plan. His team has raised nearly $30 million from institutions including 5Y Capital and IDG Capital, at a post-money valuation of $200 million. The team currently has about 10 people and has not yet publicly released a product. A world model aims to let AI learn how objects in the real world move and change, and can be used in fields such as video generation and robotics.

Tian Keyu is the first author of the NeurIPS 2024 Best Paper, and previously studied how to enable AI to generate images progressively from coarse to fine. Following this direction, the new team has designed a "vocabulary" of 200,000 visual symbols for video. These symbols function similarly to the tokens used by language models when processing text, allowing AI to process video in a way more suitable for computation. The team plans to train the model on 100 million hours of video.

Tian Keyu said this method is expected to reduce the cost of generating one second of video to one-tenth of the current level or even lower. The team plans to give a public demonstration first, and then release the full model in 2027.

He also admitted to Bloomberg that he had shut down another intern's program, claiming it was to free up computing power for other researchers, and acknowledged that his approach at the time was inappropriate. ByteDance dismissed him in 2024 and subsequently sued for 8 million yuan in damages. A judgment reviewed by Bloomberg shows that the court ordered Tian Keyu to pay 500,000 yuan for breach of contract.