Summary of Key Points
Chen Boyuan, a post-2000s graduate from Peking University, founded the AI company Inverse Matrix, which focuses on developing the "Universal World Base Model" (Physis). This model aims to enable AI to truly understand physical laws such as gravity and causality, allowing it to solve real-world physical problems in various scenarios, including industrial applications and embodied intelligence. The team received tens of millions of dollars in investment from firms like Hillhouse shortly after its establishment and is currently valued at over 5 billion yuan, with plans for a financing round exceeding 100 million dollars. Chen Boyuan also serves as the head of the Behavioral World Model Innovation Center at the Zhiyuan Research Institute, working together with them to advance the implementation of this technology. Their model differs from others in the industry that focus on specific use cases by using unique approaches such as physical latent space learning and interactive validation to create the infrastructure needed for AI to interact with the real world.
I. The "Powerful" Financing of the Post-2000s Team: Behind the 5-Billion Yuan Valuation is a Commitment to Solid Technology
Chen Boyuan, the founder of Inverse Matrix, is a recent graduate from Peking University's Yuanpei College. While still in school, he won the best paper award at the international conference ACL for his research on reinforcement learning (along with the DeepSeek team). He founded the company in early 2026 before completing his studies and secured over tens of millions of dollars in investment from Hillhouse and Peking University-related funds. With a current valuation of over 5 billion yuan, Inverse Matrix is one of the highest-valued AI teams among those founded by post-2000s individuals.
However, they are not in a hurry to raise more capital: "We prioritize making breakthroughs in technology." This is also the reason for their collaboration with the Zhiyuan Research Institute, which places more emphasis on research rather than commercialization. The institute has previously incubated companies like ZhispAI (now listed) and Yuezhi Dianmian, both valued at over 10 billion yuan. As the founder of Inverse Matrix, Chen Boyuan also leads the Zhiyuan Innovation Center, enabling mutual support: Inverse Matrix conducts cutting-edge research, while Zhiyuan applies the validated technologies to broader scenarios.
II. The Universal World Base Model: Taking AI from "Watching Videos" to "Understanding Physics"
Many companies in the industry talk about world models, but most of them only achieve superficial results—such as generating smooth videos (like OpenAI's Sora) or creating game interactions (like Google's Genie). These models are considered W0/W1 level, meaning they can only handle tasks related to media and gaming.
Inverse Matrix aims for a W2+ level universal world base model that truly understands physical laws. For example, in industrial settings, AI needs to understand constraints like gravity, contact, and causality for tasks like controlling robotic arms or predicting collisions in autonomous driving. Their goal is to create a single model that can handle all physical scenarios, just as ChatGPT can handle various tasks across different domains (e.g., finance and law), because the underlying physical principles are universal.
III. Technological Challenges: A Different Approach from ChatGPT
While ChatGPT succeeded by amassing large amounts of data and parameters, world models cannot follow this path due to three major obstacles:
1. Lack of physical data: While internet text can be easily collected, real-world physical data (e.g., data on robotic arm movements or glass fractures) is difficult to gather in large quantities.
2. Discrepancy between pixels and reality: Videos contain mostly textures and lighting that are irrelevant to physical laws, wasting parameters when used in models.
3. Correlation does not equate to causality: ChatGPT learns statistical associations (e.g., "it rains" and "people use umbrellas"), but physics requires understanding cause-and-effect relationships (e.g., "pushing a cup" causes it to fall).
Inverse Matrix's solutions include:
- Physical latent space learning: Allowing the model to directly grasp abstract physical principles without focusing on individual pixels.
- Introducing action-based interactions: The model needs to not only observe the world but also take actions (e.g., simulating the act of pushing a cup to predict its outcome).
- Reinforcement learning for validation: Physical laws can be verified using standard methods, similar to how math problems have correct answers, allowing the model to continuously improve through reinforcement learning.
They have observed that as they collect more data and engage in more interactions, the model's physical prediction accuracy improves, potentially leading to a breakthrough similar to ChatGPT-3.
IV. Insufficient Data? They Use a "Three-Tier Pyramid" and Simulation to Generate Data
To address the lack of physical data, Inverse Matrix has developed a "data pyramid":
1. Base layer: Learning about fundamental properties (e.g., water flows and glass breaks) using extensive real videos and complex interaction data.
2. Middle layer: Understanding cause-and-effect relationships (actions leading to outcomes), focusing on first-person data (e.g., from the perspective of "I push a cup"), which naturally contains causal information.
3. Top layer: Learning about rare or extreme scenarios (e.g., glass explosions or object occlusions) using simulation engines to create realistic environments.
The ratio of first-person data to multi-view data (external observations) is 9:1 or even 100:1, as first-person data is more effective for learning causal relationships. They have also found that cross-scenario data can reduce the need for specialized data by up to 20 times while improving model performance.
V. Not Just a Post-2000s Era, but the Future of Physical AI
Although the Inverse Matrix team is primarily composed of post-2000s individuals, Chen Boyuan believes that "Physical AI belongs to everyone who believes in this direction and is willing to invest in it for the long term."
Why is Physical AI important? Because the next step in AI development (AGI) must involve interacting with the real world. Current large models can only process text and images, but future AI should be capable of controlling robots, conducting industrial simulations, and predicting scientific experiments (e.g., controlled nuclear fusion). These developments are closely related to new productivity trends such as low-altitude economics, commercial aerospace, and intelligent manufacturing, all of which require AI that understands physical laws.
Inverse Matrix's goal is not to create a model for a specific industry but to build an infrastructure for the entire physical world—just like electricity, which powers various real-world applications. This news highlights how post-200s teams are using innovative approaches to push AI from the virtual realm towards the real world, with Physical AI potentially becoming the next transformative technology.