Tian Keyu’s unnamed AI lab reportedly raises about $30M

Tian Keyu’s unnamed AI lab is developing a token language intended to help AI systems process and predict video. The prospective users are teams building robots, interactive video and autonomous-driving systems, but the ten-person startup has not released a product. Bloomberg reported that it raised about $30 million from 5Y Capital, IDG Capital and others at a $200 million post-money valuation; 5Y and IDG confirmed investing without giving further details.

The technical idea is to translate visual information into symbols that a model can compress and predict, instead of forcing motion and spatial relationships through human language. Tian said the team has assembled a dictionary of about 200,000 machine-readable symbols and wants to train on 100 million hours of video. Those are development claims, not evidence of a commercial deployment.

The work grows out of Tian’s image-generation research. He was first author of a paper selected as one of the NeurIPS 2024 best papers; its autoregressive method predicts groups of pixels from coarse layouts toward finer detail, borrowing a sequence-model approach associated with language models. The startup is trying to extend that progression from still images to video and, eventually, models of how scenes change.

A world model is supposed to represent how objects move, interact and respond to actions. That could let a robotics or self-driving team rehearse scenarios in simulation, or let a video tool preserve a scene as a user changes it. Those use cases explain the market the lab is chasing, but they are prospective: Bloomberg said it still has no product.

The founder also arrives with a complicated record. ByteDance dismissed Tian in 2024 after he stopped another intern’s program; Tian later acknowledged the action was improper, and Bloomberg reported that a court found him in breach of contract. The investment therefore backs both a prize-winning research direction and a founder whose conduct at his previous employer drew a legal judgment.

For now, there is no disclosed customer base or business model. Tian told Bloomberg that he is prioritizing the technology rather than a consumer product or monetization, and said the team plans live demonstrations before releasing a full model in 2027.