Summary of Key Points
World models represent a new frontier in the field of AI, with the goal of enabling machines to make intuitive predictions about the world in a way similar to humans (such as guessing the plot of a comic or predicting a goal during a football match). This is a crucial step towards achieving Artificial General Intelligence (AGI). Countries around the world are competing fiercely in this area, but for the first time, Chinese companies cannot find direct equivalents from the United States. Chinese researchers play a central role in technological development. China has advantages in terms of low data costs and a fast hardware supply chain; however, it faces challenges with insufficient computing power. To implement these models effectively, three major issues need to be addressed: computing power, data, and model architecture. Additionally, security considerations must be taken into account at the levels of algorithms, application boundaries, and product design.
Detailed Explanation
1. Understanding World Models
A world model is essentially an AI system that can predict future scenarios based on data, similar to how humans mentally fill in gaps between consecutive frames in a comic or anticipate the direction a ball will go when watching a football match. In simple terms, through data training, AI can develop the ability to make intuitive predictions, which is essential for creating general-purpose robots that can perform human-like tasks and is a necessary part of the path to AGI.
2. Global Competition
For the first time in this field, Chinese companies do not have direct competitors from the US. Chinese researchers are becoming a key force in this race:
- In the past, Chinese tech companies tended to follow the US lead (e.g., Google in the US and Baidu in China). However, with the shift towards world models, the situation has changed:
- In 2020, OpenAI shifted from focusing on physical AI (robots) to language models, but research in other areas continued. Meta developed vision models, and Chinese scholars like He Kaiming and Xie Saining were at the forefront of these efforts.
- 2025 marks a turning point: NVIDIA made its Cosmos model open-source, and Chinese teams such as Shengshu Technology and Moke Robot started their work almost simultaneously, with some even achieving better results than Cosmos. For the first time, Chinese companies no longer need to compete against specific US counterparts, as everyone is on the same starting line.
- Chinese researchers are highly influential: The core authors of NVIDIA’s Cosmos model include Chinese individuals, and Harvard Professor Dylan Du is also involved in research on robot world models. Many startup teams in Shenzhen and Beijing are led by Chinese people.
3. China’s Advantages and Disadvantages
Advantages:
- Low data costs: There are hundreds of data collection companies in China that can provide human-centered videos (e.g., footage of people cooking or opening doors) at lower costs compared to North America.
- Fast hardware supply chain: If a robot breaks down, repairs can be done quickly (within 10 minutes in Shenzhen), whereas in the US, it might take days. This rapid response allows Chinese teams to iterate on their products more rapidly.
Disadvantages:
- Insufficient computing power: Chinese companies only have 1/4 to 1/5 of the computing resources available to their North American counterparts. Major companies allocate most of their computing power to language models (such as ChatGPT), leaving limited resources for robot models.
4. Challenges in Implementation
Three major hurdles must be overcome to apply world models to robots:
- Computing power: Leading robotics companies only have a few thousand computing cards, which is insufficient for large-scale training.
- Data: There is a lack of data that combines videos with corresponding actions (e.g., videos of people lifting their hands and the corresponding commands for robots), with the goal of accumulating hundreds of millions of hours of data.
- Model architecture: An efficient model is needed to process large amounts of data effectively. Currently, there is a shortage of cutting-edge researchers and resources for data collection.
5. Ensuring Security
Security is critical when deploying robots in everyday scenarios:
- Algorithms: Develop safe algorithms that instruct robots to stop if they come into contact with people (e.g., the SafeDojo model developed by the Chi Xiaowei team).
- Application boundaries: Limit the actions of robots (e.g., preventing them from lifting heavy objects or moving outside designated areas) using underlying control systems.
- Product design: Equip robots with protective layers (e.g., inflatable materials) to reduce injury in case of collisions.
Conclusion
World models represent the next major challenge in AI development. For the first time, China is on par with the rest of the world in this field. Although it faces limitations in computing power, its advantages in data and supply chain give it a significant advantage. By addressing these implementation issues, China has the potential to emerge with world-class AI companies. The role of Chinese researchers in this area will become increasingly important as the field progresses.