Summary of Key Points
This news article focuses on the third entrepreneurial venture of Mu Zelin, a post-95s serial entrepreneur, with his company Moushen Intelligence. Moushen Intelligence specializes in the "World Action Model," which is essentially the "brain" of robots, and has raised nearly 1 billion yuan in funding to date. Unlike most teams that emerge from laboratories and focus on publishing research papers first, Mu Zelin has emphasized commercialization from day one, rejecting "showroom-style fake orders" in favor of only accepting real orders with large-scale demand. The company's technical approach addresses the current shortcomings of embodied intelligence: mainstream VLA (Visual Language Action Models) function like rote memorized actions that fail to adapt to different scenarios. In contrast, the World Action Model enables robots to understand physical principles and adjust their behavior accordingly. The team is led by Mu Zelin, who oversees business operations (sales, financing), and Chen Tao, a tenured professor at Fudan University, who manages the technical development. They believe that creating such a model requires a systematic approach involving more than 40 doctors working across different research areas, as it cannot be achieved by a single genius.
I. The Challenges and Iterations of a Serial Entrepreneur: From AI Customer Service to Robot Brain
Mu Zelin's entrepreneurial journey has been one of steady progress:
- First venture: He worked in education in college to gain industry knowledge.
- Second venture: After graduation, he founded Muxin Intelligence, which provided AI customer service services, achieving annual revenue of 100 million yuan and a profit of 20 million yuan. However, they discovered that the model architecture was flawed—relying on data from small and medium-sized enterprises led to a significant decrease in data value when larger models emerged. Moreover, as a project-based company, it struggled to reach new heights.
- Third venture: They set three criteria: focus on AI, develop multi-modal capabilities (not just text processing), and create products that can be sold individually (rather than as part of projects). They eventually settled on the World Action Model because it meets these criteria and aligns with the technical expertise of Chen Tao's team at Fudan University.
Mu Zelin’s lesson is that the model architecture is more crucial than the amount of data collected; rushing into commercialization too early can limit growth. The goal is to create sustainable solutions rather than short-term profits that may lead to the company being sold off.
II. Technical Approach: Why VLA Fails and the World Action Model Is the Future
Here’s a simpler explanation:
- Problems with VLA: It’s like memorizing a poem by heart—robots can perform tasks like picking up a plate from a table of a fixed height, but they fail if the table is lowered (because VLA models only remember the action trajectory without understanding physical principles, such as the need to bend more when the table is lower). This also wastes computing resources on irrelevant information (like identifying patterns on the table).
- Advantages of the World Action Model: It’s like learning grammar rather than memorizing a poem—it understands how actions change with the environment (for example, adjusting bending posture based on the table height). The model generates universal action trajectories that can be applied to different robots, solving issues related to cross-scenario and cross-robot compatibility.
- Mu Zelin’s assessment: Pure VLA approaches will be overshadowed by the World Action Model, which is still in its early stages (20%-30% mature) but focuses on the fundamental relationship between environment and action. The time needed to deploy new models has decreased from several months to just 2-3 weeks by 2028.
III. Commercialization: Rejecting Showroom Robots, Accepting Only Real Orders
Mu Zelin’s approach to commercialization is practical:
- Profit from day one: A company is not a laboratory; it needs to survive market cycles, and technology must be ready for widespread use.
- Filtering real orders: They consider the scale of demand—for example, only companies with 10,000 employees would likely buy 10,000 robots (rejecting those who want just one for a showroom display).
- Setting barriers: They charge a initial fee for research and development; only customers with substantial needs can afford the cost, while those with fake demands won’t pay.
- Avoiding project-based models: They aim to sell products individually (e.g., 1 billion robot brains) rather than relying on one-time project payments.
Mu Zelin’s goal is to build Moushen Intelligence into a company with annual revenue of 1 billion yuan, rather than a small, profitable startup.
IV. Team and Competition: A Well-Structured R&D Team Is a Competitive Advantage
- Clear division of labor: Mu Zelin handles business, while Chen Tao focuses on technology (algorithms and engineering). Mu Zelin listens to technical meetings but doesn’t get involved in the details.
- Systematic R&D: In collaboration with Fudan University’s laboratory, they have over 40 doctors working across four research areas. If one person leaves, only a quarter of the knowledge is lost, highlighting the importance of a team effort.
- View on competition: Their main competitors are those who started working on the World Action Model two years ago (few both domestically and internationally). New teams established after 2026 are likely just following the trend; the World Action Model requires a systematic approach that’s too late to start from scratch. They’re not worried about young geniuses, as long-term commitment and teamwork are essential.
V. Financing and Industry Insights: Why Choose State-Owned and Industrial Capital?
- Reasons for financing: This round of funding mainly came from state-owned entities (Shenzhen Newspaper Group, Shaanxi High-Tech Investment) and industrial capital (Anyu Fund). The benefits include policy support and resource coordination, as well as direct access to real-world applications and orders. These investors are more patient, making them suitable for long-term R&D projects like the World Action Model.
- Industry outlook: Technological paths will converge; the World Action Model follows a principle-based approach similar to the Transformer architecture in large language models. Society will ultimately reward those who make genuine efforts, not just speculators.
This article illustrates how a tech entrepreneur with a business acumen finds the right direction in the AI industry—balancing technical innovation (systematic R&D) with market considerations (early commercialization and order selection), striving to find a balance between idealism (long-term technology development) and realism (surviving through profit generation).