A former intern at ByteDance Ltd has secured nearly US$30 million in funding from prominent venture capital firms, including 5Y Capital and IDG Capital, to establish a stealth artificial intelligence (AI) startup focused on advancing the development of world models. The company, which does not yet have a name or public product, is valued at approximately US$200 million following the recent investment round.
Tian Keyu, who completed his Master’s degree at Peking University in 2025, leads the new venture. Over the past year, Tian has concentrated on building AI models that simulate real-world physics and mechanics. These models aim to power a variety of applications such as robotics, interactive video, and autonomous driving. This niche research builds on the pioneering work of AI experts like Fei-Fei Li and Yann LeCun.
Tian gained recognition in the AI community as the lead author of a paper awarded one of the best presentations at the 2024 NeurIPS conference. The research introduced a novel, efficient image-generation method operating similarly to large language models but predicting clusters of pixels rather than words or text.
Prior to launching his startup, Tian spent three years at ByteDance, where he was dismissed in 2024. ByteDance stated at the time that he was terminated due to a “serious violation” but did not elaborate. The company refuted online allegations that Tian had deliberately sabotaged foundation model training efforts. Tian admitted to shutting down another intern’s program to free up computing resources for his colleagues, describing his actions as a form of vigilante intervention. A court later ordered Tian to pay 500,000 yuan (approximately US$75,000) for breaching his contract, including forfeiture of his entire salary.
In the new venture, Tian is developing a tokenized visual language for AI systems to process video data more effectively. He explained that human language falls short in describing video content precisely, so his team has created a symbolic dictionary containing 200,000 unique tokens designed specifically for AI interpretation. This approach aims to reduce the computational cost of video generation significantly, by at least an order of magnitude.
The startup currently employs around 10 people, including other former ByteDance employees. Tian plans to showcase the technology at live events before releasing a full model in 2027. At this stage, the focus remains on advancing the underlying technology rather than developing consumer products or business models.
Investors involved in the funding round confirmed their participation but declined further comment. Tian’s efforts reflect a broader surge in venture capital interest in emerging AI technologies, especially those seeking to expand the capabilities of visual and world-modeling AI beyond traditional language-based approaches.