Special Topic: 2026 Sustainable Global Leaders Conference & the Second Green Industry and Sustainable Consumption Expo. Reporting by Sina Finance, Zhou Wenmeng.
The 2026 Sustainable Global Leaders Conference was held in Shanghai from October 20 to 22. Thomas C. Südhof, winner of the 2013 Nobel Prize in Physiology or Medicine and a neuroscientist, held a dialogue on stage with Jin Yaochu, an academician of the European Academy of Sciences, an IEEE Fellow, and founder of the Trustworthy and General Artificial Intelligence Laboratory at Westlake University.
In an exclusive conversation with Sina Finance after the meeting, Jin Yaochu pointed out: "The human brain has characteristics such as extremely low power consumption, dynamically adjusting functional connectivity, and closed-loop operation. Embodied intelligence needs to deeply learn from the human brain's operating mechanisms to develop autonomous planning and closed-loop control frameworks, and cannot completely follow the existing large model paradigm."
In his view, although AI large models have indeed driven the rapid progress of embodied intelligence, current embodied intelligence technology still has obvious shortcomings in completing complex and fine operations. External expectations for embodied intelligence are too high, and the iPhone moment for embodied intelligence will not arrive very soon.
Speaking about the increased safety risks in the physical world that may result from AI plus embodied intelligence, Jin Yaochu believes that simply "braking" AI development is meaningless. A better route is to research intrinsically safe and reliable artificial intelligence rather than stop technological development.
Robot "brain" research should learn from the human brain's operating mechanisms
Sina Finance: How does brain-inspired technology work in the "brain" of embodied intelligent robots? What is its relationship with existing neural networks, and how do they divide work and collaborate?
Jin Yaochu: Although today's large model neural networks are inspired by neuroscience, there are still major differences from the human brain's essential processing mode. First, the human brain mainly relies on spike signals, event-driven processing, and sparse activation to handle information. This is the key to the human brain's extremely low power consumption: 86 billion neurons and 100 trillion synaptic connections, with total power consumption of only 20 watts. Large models use analog signal processing, and computing power and energy consumption are enormous, which is very important for edge devices such as robots. Second, the human brain dynamically adjusts functional connectivity according to tasks and flexibly changes working modes through neural plasticity and neuromodulation. This is a capability currently missing from large models. Third, most embodied intelligence today uses end-to-end data-driven methods. No matter what kind of data and models you use, it is equivalent to open-loop control. But the human brain works in a closed loop. The reason humans are intelligent is that they have the ability to actively perceive, model, and predict the external world and themselves, without needing to precisely model and reproduce the external world. When performing tasks, they can perceive surprises in real time, correct promptly, and replan. At the same time, they have memory mechanisms such as long-term memory, short-term memory, and working memory, as well as bottom-up and top-down dual-channel information processing logic. Therefore, we need to learn more deeply from the human brain's operating mechanisms to develop better AI, and we cannot completely follow the existing large model paradigm.
Sina Finance: Has cutting-edge technology such as brain-inspired computing already entered daily life and achieved large-scale deployment?
Jin Yaochu: Brain-inspired computing is a very cutting-edge direction in artificial intelligence. Cutting-edge means new technical ideas, but it also means it has not yet reached the stage of large-scale direct deployment. I began doing research related to brain-inspired computing around 2000. After returning to China, I carried out multiple collaborations with State Grid in the field of brain-inspired computing. On one hand, we work on forecasting new energy power generation and load forecasting. On the other hand, our current key direction is to combine brain-inspired computing with embodied intelligence. We recently also obtained a major national science and technology project to tackle the deployment and application of brain-inspired embodied intelligence technology. In terms of energy consumption and information processing paradigms, brain-inspired computing differs from current large models and has its own advantages, but its overall technological maturity is even weaker than traditional AI. Truly large-scale entry into public life has not yet arrived, but it may be possible to overtake on a different track, starting with pilot applications in industries such as embodied intelligence.
Shortcomings remain in fine operations, and the iPhone moment will not arrive soon
Sina Finance: If we make a time prediction, approximately how long will it take for brain-inspired plus embodied intelligence to be deployed?
Jin Yaochu: Time prediction itself is very difficult. But the country now attaches great importance to brain-inspired technology, and scientific research projects clearly put forward a deployment orientation, promoting technology from research to industry. More than two years ago, we began research and development with Wenzhou Electric Power for fine operation scenarios. Existing embodied intelligence technology has obvious shortcomings in fine operations. We hope to introduce brain-inspired technology to improve robots' autonomous operating ability to handle subtle changes and long-cycle complex tasks. Conservatively, in about 3-5 years, large-scale deployment and application can be achieved in some limited scenarios.
Sina Finance: Can applying brain-inspired computing research to embodied intelligence improve robot generalization? When will the embodied intelligence industry welcome its iPhone moment?
Jin Yaochu: In my personal view, the current route of end-to-end, world models, and world action models is still insufficient to achieve truly autonomous embodied intelligence. Embodied intelligence cannot simply apply the machine learning concept of "generalization"; what it pursues is autonomous task capability. Even if you have a perfect world model and model all the details of objects in a room, the robot will not naturally learn to do work. When people do things, they do not fully reproduce all details in the environment. For example, when I am chatting with you in this room, I will not pay attention to how many chairs are nearby, nor how many switches are on the wall. I only focus on communicating with you, the position of the camera, and so on. I will only model and predict external information related to my task. At the same time, I also need to model the body itself and perceive body information such as my own limbs and the weight and hardness of grasped objects. This is what we call the internal model: first, modeling and predicting the external environment related to the task; second, modeling and predicting the robot body itself. Relying on these two points can truly improve the ability of embodied intelligence systems to autonomously complete tasks, and rely on online feedback for continuous iteration. Large models have indeed driven rapid progress in embodied intelligence, but we cannot expect every robot to be directly controlled by a large model. Public expectations for embodied intelligence are partly too high. We can refer to electric vehicles, whose technological development took twenty to thirty years. I do not think the iPhone moment for embodied intelligence will arrive very soon. Movement abilities such as running, jumping, and dancing are already done quite well, but fine operations still have a long way to go. Household service and elderly care companion robots will enter homes more slowly, because embodied intelligence safety issues must also be solved at the same time.
Sina Finance: Regarding the concept of internal models, what is the core difference from what everyone now discusses as self-feedback learning and world models?
Jin Yaochu: Self-feedback learning leans more toward the self-iteration and evolution of agents, and is not completely equivalent to what we call internal models. The industry is now gradually beginning to pay attention to online closed-loop correction. World models are more about precise modeling of the external environment, while internal models focus only on the task-related external environment plus body perception, and do not need high-precision modeling of all details. This is the key difference between the two, and there is still relatively little discussion of this in the industry.
Braking AI is useless; we need to develop trustworthy AI
Sina Finance: Will AI large models plus physical robots gaining stronger action capabilities amplify safety risks? At the industry and association levels, how should constraint standards be established?
Jin Yaochu: Risks will indeed be amplified. AI agents can perform soft operations, while embodied intelligence directly acts on the physical world. Once errors occur, they can cause personal injury, and the safety consequences are more serious. Currently, relatively few people in the industry work on safety. Many people think technology should be built first and safety added later. I believe we need intrinsic safety: safety must be integrated into the system during the research and development stage, rather than putting a safety shell on after the technology is completed. Our laboratory has set up a dedicated group to study embodied intelligence safety and embed human safety common sense into robot systems. For example, let the robot itself know not to put flammable materials near the stove, and not to place glass cups at the edge of a table. The industry has not yet reached the stage of large-scale standard setting. At this stage, the focus is on researching risk identification and enabling robots to actively avoid risks. After the technology matures later, industry norms and standards can be promoted.
Sina Finance: Some voices overseas have proposed that AI development needs "braking." What do you think?
Jin Yaochu: Simply braking has no practical meaning. Just as with the earlier call to pause large models for six months, there was great controversy. A better route is to simultaneously research intrinsically trustworthy, safe, and reliable artificial intelligence, advancing technological innovation and safety capability building in parallel, rather than stopping technological development.
Sina Finance: You serve as chairman of the Zhejiang Province AI Agent Industry Development Council. How do you view AI agents? What ideas does the alliance have for solving the difficulty of AI deployment?
Jin Yaochu: The core difference between AI agents and large models is that, in addition to reasoning, they have perception and action capabilities. Actions can be digital-level decisions or physical mechanical actions. With autonomous decision-making and action capabilities, the applicable boundaries and industrialization possibilities of AI technology will be greatly broadened. This is also the original intention of establishing our industry alliance. The alliance connects academia and industry. We will visit enterprises and hold academic conferences. For example, we have already jointly built an industrial artificial intelligence research center with enterprises, which was facilitated by this alliance. We hope to promote the deployment of agent technology into usable products: both daily work scenarios such as document processing, and industrial scenarios such as process parameter optimization and complex industrial device design.
Comments