Summary of Key Points
At the 2026 World Artificial Intelligence Conference (WAIC), "world models" became a focal point, recognized by the industry as the foundational technology for AI to understand the laws of the physical world and achieve embodied intelligence (such as robots' autonomous actions). Multiple companies released related products with diverse technical approaches, but there is consensus that the scale of data is the key to breakthroughs—it will take the accumulation of tens of millions of hours of multi-modal "task-related data" to bring about significant improvements. In terms of applications, the gaming industry is set to benefit first, and in the long run, embodied intelligence is the core direction. The "ChatGPT moment" for physical AI is expected to occur in 2-5 years.
I. World Models: The "Brain Simulators" for AI to Understand the Physical World
Simply put, a world model is like giving AI a "physical world sandbox within its mind." Previous AI systems (like ChatGPT) were good at processing text but did not understand the real-world logic—things like a cup breaking when dropped or the force required to push a table. World models enable AI to "simulate" physical laws, allowing it to predict the consequences of actions in the same way humans do, such as knowing how much force to use when a robot picks up a cup without causing it to break. This is essential for autonomous behavior in the real world.
Why are they so popular now? Because to realize embodied intelligence (such as robots and autonomous driving), AI must first "understand the world." Previous AI was more theoretical; now, it needs to be practical, and world models serve as the bridge between digital AI and the physical world.
II. Companies Racing to Develop World Models: Diverse Approaches, All Betting on the Future
Several companies showcased their world model solutions at WAIC, each with its own focus:
- Kunlun Wanwei Matrix-Game3.5: Uses gaming scenarios for training AI. It utilizes over 1,200 game scenarios as a training ground and innovates with "patch-level memory" (inserting small, coordinate-based blocks into images to help AI remember spatial locations consistently). The core architecture is open-source, aiming to attract more participants.
- Ant Lingbo LingBot-VA2.0: Features a "natively designed" approach without combining different models. Previous robot "brains" were composed of digital and physical models; now, the design starts from the need for interaction with the environment, including dynamic modeling and real-time execution, making it more realistic.
- 3D Generation Companies (such as Yingmou Technology): Focus on creating "scene-level 3D" environments. While earlier 3D generation focused on individual objects (like a cup), they can now generate entire room scenes, providing AI with a more authentic training environment, moving from simply "viewing images" to practicing in virtual spaces.
Although the approaches vary, all companies see world models as the core of future AI competition.
III. The Biggest Barrier: Data! Tens of Millions of Hours of Task-Related Data Required for Breakthroughs
Industry experts agree that the key to advancing world models lies in the quantity and quality of data:
- Current Level: Most companies have only around 100,000 hours of training data; Kunlun Wanwei used more than 200,000 hours this time, but it's still far from a breakthrough.
- How Much Data is Needed?: Fang Han (CEO of Kunlun Wanwei) estimates that at least 10 million hours of data may mark a turning point, and potentially up to 100 million hours.
- What Kind of Data Matters: It's not just any video or image; it needs to be "task-related data"—for example, recording a robot's actions, the state of the object, and sensory feedback (such as grip strength). Chen Yilun from Shizhihang emphasizes that watching videos alone is insufficient for learning physical laws; multi-modal data (visual, tactile, and force data) is essential.
- China's Advantage: China has a strong presence in data collection hardware (sensors, robots), and there are many service scenarios (such as home services) that facilitate faster data accumulation. He predicts that some companies could reach 1 million hours of data by the end of this year, tens of millions by the end of next year, and 100 million hours in 3-5 years.
IV. Application Prospects: Gaming Industry to Benefit First, with Robots Taking the Lead in the Long Run
- Gaming Industry to Benefit First: Open-world games (like Genshin Impact) previously required teams of hundreds of people working for years; now, world models can make virtual worlds grow in real-time with player actions—for example, if you cut down a tree, the surrounding grass will grow taller due to changes in sunlight. In the next 3-5 years, world models will become a fundamental component of the gaming industry.
- Embedded Intelligence as the Long-Term Direction: Robots need to be integrated into homes and factories to perform tasks effectively, which requires understanding the environment through world models. However, this is still in its early stages; for example, robots often drop objects, indicating a need for more data and technological improvements.
V. The "ChatGPT Moment" for Physical AI: 2-5 Years Away, Depends on Multiple Factors
Experts predict that the breakthrough of physical AI (AI capable of working in the real world) will take another 2-5 years. The consensus is that it won't happen with a single model or data source; rather, it will require a combination of data scale, model architecture, simulation training, real-device feedback, and hardware to form a cycle—using simulation scenarios to train AI, testing it on real devices, and then using the feedback to refine the models continuously.
Just as ChatGPT didn't become popular overnight but was the result of years of data and model development, the same is true for physical AI. This "evolutionary wheel" needs to start spinning.
Conclusion
World models are a crucial step for AI to move from the "digital world" to the "physical world." The industry is in a phase of rapid development, with data being the biggest obstacle. China has an advantage in data collection. The gaming industry will benefit first, but robots represent the long-term goal. The "ChatGPT moment" for physical AI is still several years away, but the direction is clear: to make AI truly understand the world and be capable of performing practical tasks.