In July this year, both inside and outside the venue of the World Artificial Intelligence Conference (WAIC) in Shanghai, a term was repeatedly mentioned with unprecedented frequency: the world model.
The year 2026 is widely regarded by the industry as the first year of the world model. As breakthroughs of large language models (LLMs) in the digital world gradually move toward commercialization, the world model, which focuses on compressing time, space, physical causality, and actions, is viewed as the ultimate path toward artificial general intelligence (AGI).
However, the path from the laboratory to the physical world is fraught with numerous hurdles in computing power, data, model architecture, and industrial synergy.
On the morning of July 19, during the CEO Breakfast Meeting hosted by Guancha.cn at WAIC, the theme of the second session was set as "The Decisive Summit of the World Model."
Four guests from diverse backgrounds, including Chi Xiaowei, founder of general-purpose robotics company Muka Robotics; Wang Zeyuan, executive head of the world model at Shengshu Technology; Tan Huajie, CEO & CTO of EvoPhys.ai; and Zhang Chen, partner at Chengwei Capital, sat together to address a core question: How can China climb this decisive summit?
Why Is the World Model the Decisive Summit for AI Implementation?
Before discussing how to climb it, we must first answer a question: Why is the world model so important that it is called the decisive summit of the AI industry?
Zhang Chen, partner at Chengwei Capital, provided a definition from the perspective of the capital market. He believes that a model is essentially a lossless compression of world knowledge. While LLMs have successfully compressed linguistic knowledge, generating capabilities such as coding, writing, and common-sense reasoning, the world model aims to compress multi-dimensional factors beyond language, including time, space, physical causality, and actions.
This higher-dimensional compression could potentially become the ultimate solution for AGI. Zhang Chen stated that the world model is called the decisive summit because it represents the compression of the highest dimension of intelligence. Language models only compress human cognition of language, whereas there are numerous forms of intelligence beyond language that remain uncovered by current models.
From the perspective of technological implementation, Tan Huajie summarized the connection between embodied AI and the world model using a twin relationship. He believes that for robots to truly enter the physical world, they must possess a worldview to understand it, which is inseparable from the empowerment of the world model. Conversely, to verify its own effectiveness, the world model must be tested in real environments; thus, embodied AI becomes the sole means of evaluating the world model. This mutually prerequisite and mutually validating relationship dictates that the world model is a necessary prerequisite for the implementation of embodied AI, rather than an optional add-on.
Chi Xiaowei, on the other hand, approached the issue from more specific implementation bottlenecks. He noted that the implementation of world models in the embodied field generally faces three major shortcomings: insufficient computing power, a lack of human-centric scalable data, and the absence of efficient model architectures.
In China, leading language model providers occupy the vast majority of computing resources, leaving extremely limited computing power for robotics and world model-related fields. Regarding data, although massive amounts of video data exist, how to transform this data, which lacks action labels, into effective resources for training remains a challenge. In terms of model architecture, he believes this is essentially a competition of efficiency, requiring the discovery of architectures that can maximize the utilization of limited computing power and data.
Wang Zeyuan further elaborated on the core value of the world model from the perspective of technical routes. He proposed that while robot hardware will inevitably become diversified in the future, intelligence will definitely converge, necessitating a more general embodied model to drive all different hardware platforms. The world model is precisely the key to achieving this general intelligence. It can endow robots with the ability to understand and predict the world, much like the human brain, without requiring pixel-level precise restoration.
Wang Zeyuan mentioned that the best world model is always the world itself, but what AI needs to do is simulate and predict the trends of the world within the algorithmic space.
How Are Teams in China and Chinese Researchers in Silicon Valley Racing?
In the critical track of the world model, which is regarded as the path to AGI, Chinese researchers are participating with extremely high density and depth. From academic foundation and technical architecture to industrial implementation, their presence can be seen at almost every critical node. According to the MacroPolo 2025 report, 47% of the world's top AI researchers completed their undergraduate studies in China, with Tsinghua University, Peking University, and the University of Science and Technology of China contributing nearly 60% of the top talent.
As the world model becomes the new focus of global AI competition, what position does China hold in this race?
Drawing on his personal experience in research and entrepreneurship, Chi Xiaowei outlined an industry landscape that has shifted from benchmarking to having no benchmarks. He recalled his early research between 2018 and 2020. At that time, OpenAI had just acquired the famous simulator MuJoCo and began conducting extensive reinforcement learning research, but later abruptly cut all physics AI-related divisions to go all-in on language models.
However, beneath the surface, research on world models never stopped. In 2018, Jürgen Schmidhuber published the first paper titled "World Model." In 2023, Meta released Segment Anything, a large-scale visual foundation model.
Throughout this research process, Chinese scholars have always been at the forefront. Research teams represented by MMLab have contributed a significant amount of important work in the fields of computer vision and machine learning. From Kaiming He to core researchers in the world model direction, Chinese figures are ubiquitous.
Chi Xiaowei specifically mentioned a critical turning point: the year 2025.
In this year, Chinese companies in the world model field no longer have benchmarks. In the past, the habitual practice of Chinese tech companies was to find a U.S. counterpart and answer the exact same problems as that U.S. company, but in the Chinese version. However, in the world model field, this paradigm has been broken. In the summer of 2025, NVIDIA open-sourced the Cosmos series of models, training multiple versions of foundation models using nearly 10,000 GPUs. During the same period, China's Shengshu Technology also open-sourced a series of works using generative models for general-purpose robot manipulation. Yizhuang Humanoid Innovation Center, where Chi Xiaowei himself is located, keenly perceived the value of the world model from late 2024 to early 2025, conducting its first large-scale pre-training with over 10,000 hours of data, completing it even earlier than NVIDIA's Cosmos.
Chi Xiaowei believes that China has both disadvantages and advantages in this race. In terms of computing power, the resources obtained by Chinese companies may be one-fourth to one-fifth of those available to their North American counterparts, with costs far higher than their competitors. On the other hand, China has a huge advantage in data acquisition costs. There may be over 200 to 500 related companies nationwide, making the data supply chain extremely rich.
More crucially, China possesses an extremely efficient hardware ecosystem synergy system. Chi Xiaowei gave an example: when collaborating with a certain hardware manufacturer in Shenzhen, if a robot encountered a problem, a single phone call or a less-than-ten-minute taxi ride would bring a hardware team to the site to solve the issue. He believes that this kind of soil condition could potentially give birth to truly world-class model companies.
Starting from the practice of Shengshu Technology, Wang Zeyuan further elaborated on the technical route choices of the Chinese team. He introduced that Shengshu Technology proposed the world-action model MOTHS based on the MOT architecture as early as 2025. This model includes three expert modules for language, action, and video, which can be pre-trained using data from different modalities respectively. Subsequently, the MotoBrain version, developed based on the more powerful video generation model Vidu Q3, significantly improved model performance with a larger number of parameters and a stronger foundation model, achieving unprecedented results on public evaluation benchmarks that remain unsurpassed to date.
Wang Zeyuan mentioned that this technical route—first enabling robots to gain basic cognition of the physical world through massive video data, then giving robots proprioceptive awareness through first-person perspective data, and finally adapting through fine-tuning on specific hardware—has preliminarily verified the possibility of cross-task and cross-hardware transfer of world models.
What Kind of Mindset and Courage Do Researchers and Investors Need?
In the process of climbing this decisive summit, while the technical route is certainly important, what is more critical is whether the participants possess the determination and conviction commensurate with this peak.
Tan Huajie and Zhang Chen shared their respective psychological journeys from the perspectives of an entrepreneur and an investor.
Tan Huajie mentioned that as early as 2024, while interning in the field of embodied AI, he began to ponder a question: Has the current AI development path been inverted?
He discovered that current AI first ingests massive amounts of linguistic corpora from the internet to acquire high-level language reasoning capabilities, and then attempts to use data to achieve physical world interaction, which is completely opposite to the biological evolution process. Tan Huajie believes that the root of this dilemma—easy to know but hard to do—lies in the fact that models do not start learning from the most native physical interaction data.
He proposed two core intuitions: First, all physical cognition should originate from the most native world interactions; the model should be like a blank sheet of paper, starting to learn from the alignment of actions and states. Second, complete memory is a prerequisite for the emergence of intelligence. If a model only has a three-second memory and cannot understand scenes and causal chains, it cannot be called a world model.
However, when Tan Huajie proposed these ideas, they were not recognized. He encountered the predicament of insufficient computing power and rejected proposals. It was not until he connected with the Lighthouse Project of Moore Threads and obtained 30 domestic computing devices benchmarked against the H100. Tan Huajie recalled that the team's first reaction upon getting the computing power was not to complain about the difficulty of using domestic computing power, but rather, "Finally, we can verify our proposal." During that period, the team worked in 24-hour shifts to guard the computing power and ultimately built the first version of the model. Tan Huajie admitted that during his subsequent entrepreneurial journey, he was constantly asked by investors, "What makes you different from others?" He gradually learned to persuade them with intuition and conviction, feeling that his proposal was the best in the world.
Zhang Chen, from an investor's perspective, proposed a complete investment philosophy. He used "climbing a mountain in the fog" to describe the state of investing in world models. Investors cannot see the summit, cannot clearly see who is on the optimal path, and can only see a few companions around them seemingly moving upward. He believes that this high degree of uncertainty means that traditional investment logic has failed.
Zhang Chen observed that this wave of AI presents an unprecedented dichotomy and unity. On the one hand, AI has entered the win-rate curve of value realization, with upstream computing infrastructure profitability significantly enhanced and the Annual Recurring Revenue (ARR) of model companies growing rapidly. On the other hand, the world model is still in the odds-ratio curve stage; everyone is exploring, there are no benchmarks, and it is full of uncertainty. These two curves running in parallel have led capital of different attributes—including U.S. dollars, RMB, state-owned capital, and insurance capital—to flood into the same track.
Zhang Chen proposed that investing in world models requires a Plan C. Plan A is a technology-first model, like large language models, where top-tier technology is developed first, and then commercialization scenarios are sought. Plan B is to choose a vertical scenario to enter and deeply cultivate it with engineering capabilities. However, Zhang Chen believes that neither of these two approaches works in the world model field because the resources required for world models far exceed those for language models, and it is highly unlikely to claim to be a world model relying solely on a single vertical scenario.
His proposed Plan C uses vertical scenarios as anchor points, rooting technology downwards and opening up the industrial closed loop upwards. To this end, his institution established the C Engine AI incubator, hoping to accompany scientists at the earlier stage from minus-one to zero, turning an idea from the laboratory into an investable asset.
Zhang Chen mentioned that investing in world models is itself an act of courage, because it is easy to speak incorrectly and easy to be proven wrong. He described the current stage as a game that rewards the brave, where entrepreneurs abandon safe paths and go all-in on a track with a one-in-ten or even one-in-a-thousand chance of survival. Researchers need to explore unknown directions for a long time rather than quickly monetizing research assets. Investors, on the other hand, need to dare to act when others are doubtful and dare to step back to think calmly when others are blind. He quoted a viewpoint he likes: "Investors like to think about the endgame, but when you think about the endgame, you can never enter the game."
At the end of the panel discussion, the guests also explored the issue of safety governance. Chi Xiaowei proposed a three-layer defense line: the model algorithm layer simulates safety intuition through data, the physical control layer sets action boundaries through compliant control, and the product design layer reduces risks through hardware improvements. Tan Huajie mentioned that safety governance should be front-loaded to the data level, filtering data through pipelines that align with human values before pre-training. Wang Zeyuan, drawing an analogy to the development history of LLMs, believes that safety governance requires intelligence first before considering safety. The experience of LLMs from adversarial attacks to safety alignment can provide references for world models.
Zhang Chen summarized that after the world model enters the physical world, in addition to the safety guardrails at the model layer, protective measures must also be added in the physical world, just like adding millimeter-wave radar and tactile sensors to a lawnmower. He believes that with technological development, the issues of asynchrony between the model layer and the hardware layer, as well as safety governance, will be resolved as the industrial closed loop gradually takes shape.
Zhang Chen stated that those who reach the top of the decisive summit do not rely on telescopes, but on the courage to dare to be the first in the world.
Looking at it now, for China, this decisive summit is both a challenge and an opportunity. On this new track without benchmarks, the answers can only be written by the climbers themselves.