A Chinese AI start-up founded by a former Nvidia senior manager has unveiled a novel physics-based world model aiming to overcome limitations of data-driven AI approaches used by US tech giants OpenAI and Meta.
- Fysiverse simulates environments using embedded physics, not just data correlations
- Physics-based approach enables more accurate predictions of object interactions
- Startup founded by Nvidia veteran Zhang Lihua with institutional backing
What happened
Fysics AI, a Shanghai-based start-up founded in 2024 by Zhang Lihua, a former Nvidia senior manager, announced the launch of its new AI model called Fysiverse. This world model incorporates the laws of physics directly into its programming to simulate environments more realistically. The company shared the news via a WeChat post, highlighting that the model aims to correct common issues in current world models such as unrealistic physical interactions and reasoning errors in atypical scenarios.
Fysiverse differentiates itself from existing models used by major US firms like OpenAI and Meta, which rely heavily on video data correlations, black-box learning methods, or 3D geometry predictions. Instead, it integrates physics-based simulation to reconstruct and predict real-world scene dynamics in a way that is computable and controllable.
Why it matters
Current AI world models have key limitations, including failures in reproducing physical causality and difficulty in validating their internal rules. Video-based models often capture correlation without understanding cause, while black box approaches produce knowledge that is hard to interpret or verify. Even 3D modeling techniques struggle to handle realistic dynamics like material contacts and motion.
Fysiverse addresses these problems by embedding physical laws at its core, enabling it to generate dynamic scenes that evolve according to real-world physics. This promises improvements in AI applications such as robot training, autonomous driving simulations, and content generation where reliable physical interaction modeling is critical.
What to watch next
The development of Fysiverse may influence future AI research by encouraging more physics-informed approaches in world modeling. Observers will want to track the startup’s ability to scale its technology and integrate with existing AI ecosystems, as well as how it competes with dominant players like OpenAI and Meta.
Additionally, adoption by industries involved in robotics, autonomous vehicles, and virtual environment creation will be key indicators of the model's practical impact. Funding and partnerships also bear watching, as Fysics AI has attracted investors including MPC and MetaX, signaling confidence in its approach within China’s expanding AI sector.