Musk stated earlier this year that his humanoid robot "Optimus 3" is expected to enter mass production next year. In the second half of the large model era, AI is no longer confined to digital space but must enter the real world to complete tasks for humans. Embodied AI and robotics have become the core breakthrough. Is the "iPhone moment" for embodied AI coming soon? On July 2, 2026, at the Tencent Cloud City Summit in Wuhan Optics Valley, we hosted a roundtable forum, engaging in an in-depth discussion with frontline expert representatives from enterprises, universities, and the government.
Guests:
Liu Chuanhou COO, Hubei Humanoid Robot Innovation Center
Yang Songhua Co-founder & CTO, Lunpai Technology
Ma Junjie Vice President, Beijing Kunlunxing Robotics Technology Co., Ltd.
Li Min Professor, School of Mechanical Science and Engineering, Huazhong University of Science and Technology
Moderator:
Wu Pengyang Senior Expert, Tencent Research Institute
【Core Views】
01 The "iPhone moment" is far from arriving; it's currently more like the "brick phone era".
Technical paths for embodied AI have not yet converged. Beyond VLA, there are various explorations, currently only meeting some basic functions.
02 "Whether it can move" has been solved; "whether it can be used" still requires effort.
The robot's "cerebellum" (motion control) is already quite good, but the "brain" (decision-making and generalization) is far from reaching the level for large-scale commercial use.
03 Data is the biggest bottleneck, with a significant gap.
Embodied AI likely requires tens of millions of data points to reach a moment similar to "GPT 2.0". Currently, globally there are only about 500,000, a gap of 200 times in magnitude.
04 The first hurdle for commercialization is the economics, which currently doesn't add up.
The cost of humanoid robots combined with VLA, world models, etc., is often significantly more expensive compared to traditional labor costs.
05 There are three categories of markets for embodied AI applications, with the largest having trillion-level potential.
Emotional value market (tens of billions), such as performances, exhibitions, companionship, and nurturing; Commercial service market (hundreds of billions), such as customer attraction, guidance, navigation, consultation, and shopping; Labor and operational productivity market (trillions), such as working in factories or doing household chores at home.
06 Humanoid robots still have a way to go, but embodied AI can proceed first.
Embodied AI applications can be implemented now, not necessarily requiring a humanoid form or the ability to perform labor. People would also anticipate companionship, like that from a pet.
07 The ultimate goal of AI is in the physical world, not behind a screen.
No matter how cute the AI on a phone is, it's still separated by a layer of glass. AI will ultimately enter the physical world and reach millions of households.
08 Using AI for efficiency gains is consensus, but key decisions cannot be handed over to AI.
AI has significantly improved efficiency in coding, research, teaching, etc., but risks of "deception" and "cognitive substitution" exist. Core judgment and decision-making must remain with humans.
Full interview content:
Q1: How far has embodied AI progressed?
Wu Pengyang: It's a great honor to discuss embodied AI with all the experts today, a topic of great public interest. In the marathon of humanoid robots over the past two years, robots are now running faster than humans, which has excited the public. So today's theme uses a rather provocative name—is the "iPhone moment" for embodied AI coming?
First question, how far has embodied AI actually progressed today? Guests can speak from their expertise, either from the technical angle—the development of underlying technology paths—or from the capability angle, such as running, jumping, or executing complex tasks.
Liu Chuanhou: First, the conclusion: I think embodied AI is still far from its iPhone moment.
To put it simply, embodied AI is currently in an era akin to the "brick phone," only meeting some basic functions. Current humanoid robots are divided into "brain" and "cerebellum." Full-body motion control (cerebellum) is already quite good, but the brain (decision-making and generalization) is far from being commercially viable. Since 2024, there has been continuous discussion about the brain of humanoid robots, but no significant breakthrough has been seen in industrial applications yet.
The technical path for the embodied AI brain has not converged. Currently, VLA models are hot, but they also face many problems; the road ahead is very long. Technology doesn't happen overnight; it requires sustained effort. Of course, there have been changes this year, such as combining world models and reinforcement learning with VLA models—could this create a new path? During this process, can new key data like tactile data be integrated into the models? These are under exploration. After all, for robots, it's not just about obstacle avoidance; they also need to make contact. However, no algorithmic model surpassing VLA has emerged yet.
The journey is long and arduous, but we must have confidence. Just like in the 1990s, it was hard to imagine phones solving so many problems today. It might be similar for robots.
Yang Songhua: I agree with Mr. Liu's view. Currently, embodied AI mainly consists of the physical body and the model. Looking at the physical body, events like last year's robot marathon and performances by various companies' robots during the Spring Festival Gala show that robots have basically solved the problem of "whether they can move." The next step the entire industry is focusing on is "whether they can be used," i.e., the brain problem.
Regarding the brain, as I originally worked on large language models, looking back at the development of large models, we (embodied AI) might not even be at the GPT 2.0 moment yet because our paths haven't converged. Like, we've been working on VLA models, but this year we've shifted to world models and various other models.
Because the paths haven't converged, everyone is collecting data, and issues like heterogeneity and synchronization of collected data are hard to solve. Currently, a major issue is the data gap. It's widely believed that at least tens of millions of data points are needed for embodied AI to reach a moment like GPT 2.0 or 2.5. But currently, globally, there are only about 500,000 data points combined, a gap of 200 times in magnitude.
So, embodied AI still has a long way to go, including in model architecture, data collection methods, and data scale. But I believe as long as we steadfastly follow this path, we will eventually realize the vision for the robotics industry.
Ma Junjie: From an industry practitioner's perspective, I'll explain the capabilities of embodied AI we understand in four aspects.
First is motion control capability, the cerebellum. As mentioned, this area is relatively mature. But achieving absolute maturity still has distance. The core lies in refining key modules, including electromagnetic shielding, thermal management, and motion control algorithms.
Second is mobility and navigation capability. Embodied AI exists in a 3D physical world. It first needs to know its location, where to go, perform automatic path planning, and dynamic obstacle avoidance. The technology is only relatively mature; end-to-end autonomous driving and VLA extended to the embodied AI industry can be used. But autonomous driving cars on roads face simpler scenarios than embodied AI. For example, this conference venue is more complex than public roads. Fully generalized application in such scenarios for embodied AI requires continuous refinement.
Third is multimodal interaction capability. Current large language models have good language interaction abilities. But next, in emotional companionship scenarios, limb interaction related to emotions and facial expressions might be used. These are still in their infancy.
Fourth is labor operation capability. As mentioned earlier, regarding related technical directions like world models, a preliminary consensus is forming. But exactly how to implement world models, each company is finding its own viable path. Widespread specific implementation might still require some time.
Li Min: What everyone can see is that the overall capability of robots is actually very good now. But for them to truly land and be usable, a crucial issue is stability. Whether robots can maintain stable operation over long periods like cars or other production tools—this still has a long way to go.
Another point about models: the so-called AI models or world models mentioned now, their paths and directions are not fully determined. It's still a state of a hundred schools of thought contending. An important reason is, compared to virtual AI like GPT, the most important issue is data. GPT data acquisition is relatively easier because it can obtain data from the internet. But for embodied AI, data acquisition is quite difficult, which makes data collection challenging for everyone. Whether it's real-world data, simulation data, physical interaction data, or tactile data, the lack of data actually hinders model development.
Overall, embodied AI still has a long way to go. But this industry is developing rapidly, changing day by day or week by week. The future is promising.
Wu Pengyang: Thank you to the four guests for providing some冷静 (calm) reflection amidst this fervor. To summarize, what stage has embodied AI reached? Comparing it to phones, it might be the brick phone era; comparing it to models, it might be GPT 2.0. Overall, it might still be some distance from the iPhone moment and needs more progress and observation.
Q2: Current state and future potential of embodied AI applications?
Wu Pengyang: Second question, how is the application development of embodied AI? What scenarios are relatively mature now, or show potential for规模化 (large-scale) application? Where are the applications with greater future imagination space?
Li Min: From a university perspective, the major directions now are entering factories and home or service scenarios. Overall, large-scale落地 (landing/implementation) applications are still premature. What people see now might be some demonstrative applications.
Wu Pengyang: Are there any relatively faster-progressing ones?
Li Min: Inspection might be done more, because inspection involves relatively fewer specific operations, so applications are slightly more. But truly replacing humans in factories has a long way to go, like the stability issue I mentioned earlier, as factories have very high requirements for efficiency and success rates.
Wu Pengyang: What application research are you mainly doing now?
Li Min: We are mainly focusing on engineering inspection, automotive industry, etc. There are leading companies jointly establishing labs. Actually, many enterprises have very significant demand for embodied AI; everyone is optimistic about it. But these companies don't know exactly how to use or implement it, so they rely on universities for research and demonstration.
Ma Junjie: I'll analyze from the market perspective, dividing it into three categories.
The first is the emotional value market (tens of billions) . Currently mainly performances and exhibitions, relying on motion control capability. Recently, some manufacturers are releasing companionship and nurturing robots, which might use multimodal interaction and mobile navigation capabilities. The market is in the early validation stage.
The second is the commercial service productivity market (hundreds of billions) . For example, customer attraction and guidance in various商业 (commercial) scenarios, guided tours in exhibition halls, guidance in service halls, and shopping guidance in offline stores. It mainly requires comprehensive capabilities of motion + navigation + multimodal interaction. If capabilities mature, this is a market worth hundreds of billions.
The third is the labor operation productivity market (trillions) . For example,广泛地 (extensively) working in factories or doing household chores at home. It重点依赖 (heavily relies on) the robot's labor operation capability. This market size is trillion-level. Why? Ten billion smartphones multiplied by 100,000 RMB (the lower average price of new energy autonomous vehicles), a simple estimate gives trillions. This market is enormous but also faces many challenges. In some relatively standard, specific scenarios that don't require much generalization, attempts can be made. Working in factories is possible, provided the ROI can be calculated. I believe with technological development, more scenarios will be found, making it more mature.
Wu Pengyang: Which scenario are you most optimistic about currently?
Ma Junjie: Definitely starting with the end goal in mind, but the path, as experts said earlier, is that we are in the brick phone era, and we need to do the right things at the right time.
Wu Pengyang: What scenarios are you working on currently?
Ma Junjie: For humanoid robots, the first two categories will be the main量产 (mass production) direction to锻炼 (hone) our business model moving forward, while also重点投入 (focusing investment) on some specific labor operation scenarios.
Wu Pengyang: The second scenario, the commercial service market, is larger. But we know robots in hotels already exist. What else can embodied AI do, what other valuable scenarios?
Ma Junjie: The value of commercial services is divided into two types: emotional value and functional value. The service robots seen in hotels now are mainly non-humanoid, focusing on functional value without emotional value. Adding a humanoid form to this combines emotional value with functional value, creating a very large market scenario.
Wu Pengyang: Emotional value happens to be Mr. Yang's expertise.
Yang Songhua: I'll focus on this point, which also concerns young people.
As experts mentioned, this year, robots entering factories and doing guided tours to replace repetitive labor in traditional industries is the mainstream trend. But this raises a problem: starting this year, moving towards commercialization means starting to calculate the economics. Then it's found that the cost of humanoid robots plus VLA models, world models, compared to traditional labor costs, often doesn't add up economically.
Wu Pengyang: Roughly how much is the difference?
Yang Songhua: If you want a robot to solve an application's value, building a robot to enter a factory or do household chores, it's often more expensive than a human, significantly more expensive, and sometimes unstable. This is the道理 (reason): to make something体现 (demonstrate) its practical and application value, even if you have money, you'll calculate the economics.
I believe within three years, robots might mainly provide companionship and emotional value to people. Emotional value has no上限 (upper limit). Like the company Pop Mart in China in recent years, reaching nearly a trillion market capitalization, essentially provides emotional value to young people and mothers. Recently, many domestic companies have been关注 (focusing on) bionic robots, generating billions or tens of billions of discussions online. People highly anticipate a robot, even if it can't work, just sitting there quietly watching you, accompanying you daily, chatting with you; we would also eagerly await it.
Wu Pengyang: Could it be a bit scary?
Yang Songhua: Emotional value doesn't necessarily require an extremely lifelike human face. We keep cats or dogs, or Pop Mart's Labubu, which has no human form, can also provide emotional value.
Our company has一直专注 (consistently focused) on emotional companionship robots. For example, busy parents who want their child to chat with a robot daily; or office workers who are tired and want a robot to provide emotional value, like having a cat or dog but without the burden of feeding or walking.
As for the future, we all hope robots will eventually enter千家万户 (millions of households). But I think the first step is also to cultivate user awareness, letting users first experience the companionship value of robots. Then, when our technology, models, and data scale mature, robots can gradually perform tasks like household chores.
Wu Pengyang: Current large model applications, including ChatBots we usually use, can also provide companionship and emotional value. Why add robots for companionship? What's the core value?
Yang Songhua: This actually goes back to an本质问题 (essential issue) of the industry. For example, AI on phones can also be cute, chat interestingly, many people are dating AI. But we believe the AI industry must ultimately enter the physical world.
Early AI could only do image recognition, translation. Then large models emerged for对话 (dialogue), answering questions. Now there are powerful agents helping solve workflow problems. But to achieve the ultimate scenes in科幻电影 (sci-fi movies), robots must, like everyone buys a car today, in the future, everyone might buy a robot, accompanying us daily, shopping, cooking, etc. This is the ultimate trend of the AI industry. Only this final entity, which you can see and touch daily, will truly bring a非常真实的亲切感 (very real sense of intimacy).
Wu Pengyang: Indeed, just touching a phone is still somewhat尴尬 (awkward).
Liu Chuanhou: Embodied AI and humanoid robots are related but also two different things. Humanoid robots本质上 (essentially) resemble humans more, able to truly integrate into human society. Embodied AI has a broader scope, not solely humanoid robots; it includes quadrupedal, wheeled robots, all belonging to the category of embodied intelligent robots.
The application of humanoid robots still has距离 (distance), but embodied AI applications can proceed first. Robots delivering meals in hotels now, automatic sweeping robots, are actually potential applications of embodied AI. Embodied AI applications can first切入 (enter) from other applicable scenarios, not necessarily aiming directly at specific humanoid forms.
Our Innovation Center is also actively exploring scenario applications of embodied AI on non-humanoid robots. Because China and overseas differ greatly. Overseas, like Tesla, pursues more first principles, wanting to create a human. But currently, most domestic companies are寻找场景 (looking for scenarios), doing direct applications. Many robotic companies' products are not fully bipedal humanoids; many are wheeled. Embodied AI applications can actually be applied now, without waiting ten or twenty years.
Last month, the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission jointly released a special action for humanoid robot and embodied AI real-scenario practical training. Our Innovation Center also actively applied for 5-6 partner product constructions. This year, from a strategic positioning perspective, actively promoting the application of humanoid robots and embodied AI in real scenarios is also a major national priority. For us, keeping up with or exploring feasible paths is very important.
Wu Pengyang: What scenario do you most希望突破 (hope to break through)?
Liu Chuanhou: In the long term, hoping to solve human problems. Things I don't want to do at home, or things people don't want to do—this is certainly what we want to solve. But the现实问题 (realistic problem) now is: robots are doing things we are good at, not替 (replacing) us in doing things we don't want to do. We want robots to cook, do household chores, but actually, we are doing chores, and robots are doing some脑力 (mental) tasks. This is hard to solve短时间 (in a short time).
Q3: Industrial ecosystem status and development needs for embodied AI?
Wu Pengyang: For embodied AI to突破 (break through), it certainly requires ecosystem, industry chain上下游 (upstream and downstream) linkage. From your respective enterprises and products, what is the current state of industrial ecosystem development? Are there any perceived lacking points, or hopes for breakthroughs in environment, systems?
Yang Songhua: Hubei's main advantage is that we are a traditional manufacturing powerhouse. In places like Optics Valley, there are many optoelectronics, 3C manufacturing enterprises, which确实能 (can indeed) provide a good environment for robot development零部件生产制造 (component production and manufacturing). Moreover, Hubei, especially Wuhan's biggest advantage, is the abundance of universities, with very high人才密度 (talent density). Many companies choose to place R&D in Wuhan because there are many成熟 (mature) engineers here. Like the data collection搞 (done) by Hubei Humanoid Robot Innovation Center Mr. Liu's team, it indeed helps解决 (solve) many university students' employment issues. Wuhan's main advantages lie in data and engineers.
For future development needs, compared to companies in Beijing, Shanghai, Shenzhen, it's mainly about the brain, models. Because the AI brain requires the top, smartest people. Wuhan might also need to多引进 (introduce more)优秀企业 (excellent enterprises) to play a leading示范作用 (demonstrative role).
Wu Pengyang: More specifically, for example, when you develop products now, which specific环节 (link) do you think large model companies need to provide something to you?
Yang Songhua: Currently, most companies doing multimodal likely still use traditional open-source models. We hope that some VLA models, world models can also be service-oriented, but currently it seems somewhat difficult. Hope large model companies can first provide a good base, like Tencent's Hunyuan base, which we can directly fine-tune and use. So开源 (open source) is still very important.
Ma Junjie: The industrial ecosystem is divided into upstream and downstream. Upstream includes AI and hardware; currently, China's industrial ecosystem is quite rich. But each area needs continuous advancement. For example, harmonic reducer core components heat treatment, testing equipment, etc., domestic supply chain technology needs strengthening; data, industrial ecosystem in AI need further integration.
Also, I want to mention downstream. The development journey of embodied AI might be similar to new energy autonomous vehicles, because downstream involves sales channels, financial solutions (like融资租赁 financing lease) , insurance, after-sales维修市场 (maintenance market), etc. As embodied AI increasingly appears in our lives, the downstream market will gradually develop; this is both a challenge and an opportunity.
Wu Pengyang: Which环节 (link) do you think is most challenging now?
Ma Junjie: Every link is important. To truly do these well, it's about培育成熟 (cultivating and maturing) the market, achieving a正向螺旋上升 (positive spiral上升 rise) between the market and the industry chain.
Wu Pengyang: Earlier you mentioned insurance; that's very new. How is robot insurance handled?
Ma Junjie: For example, if a robot performs, it's best to have an insurance方案 (plan). In case it doesn't perform well or an accident occurs, insurance provides保障 (protection). Actually, embodied AI insurance is similar to new energy vehicles; you can understand it that way. Everyone is exploring, and some细分行业 (niche industries) have already started doing it.
Liu Chuanhou: For the embodied AI industry, from the upstream industry chain, we梳理 (sorted out) about thirty-some companies within Hubei last year. But this year,重新梳理 (re-sorting) companies that can enter the industrial ecosystem, there are about一百四十 (one hundred forty)几家 (several). Of course, whether they can enter is another matter, involving产量问题 (production volume issues). How much终端产品 (end product) related to robots or humanoid robots Hubei can produce determines the scale of the entire upstream industry链 (chain) driven by the end product.
For the entire industrial ecosystem, where are the卡脖子 (chokepoint) areas? I think it might still be in data and models. I never worry about中国制造业 (China's manufacturing), production capabilities having any缺口 (gaps). But data and models might be a比较大的问题 (relatively big problem).
Earlier, it was mentioned how much data is needed to reach a level similar to current autonomous driving L3 or L4. Industry experts predict that autonomous driving reaching current levels requires roughly数百亿 (tens of billions) hours of data. But for robots or humanoid robots, the data volume might need to reach数千亿 (trillions) of hours. We currently have only a few million hours of data, still相差甚远 (far from enough). Models heavily依赖 (rely on) data; without data, model effectiveness in scenario implementation can have significant problems.
So I think the current薄弱 (weakness) is still in data. We need大量 (vast amounts) of data, and this data needs to be交易性 (tradable), allowing all institutions with model training needs to obtain it conveniently and cheaply. Only then can better models emerge,推动 (driving) the entire industry's development.
Wu Pengyang: If focusing on humanoid robots, because training human actions and behaviors is needed, somewhat like when large models asked people to do labeling, can普通 (ordinary) people also provide these actions, potentially allowing普通 (ordinary) people to participate in data provision?
Liu Chuanhou: Data is mainly divided into several types. Teleoperation data is the most expensive, but its advantage is it can be directly移植 (transplanted) to corresponding robots for simple adaptation. Human data; starting in February this year, NVIDIA推出 (introduced) a new technical direction, embodiment-less data collection becoming a new track. Vast amounts of embodiment-less data, especially human动作行为 (action behavior) data, can be collected for model training, solving significant problems.
Now, the Hubei Humanoid Robot Innovation Center is also actively exploring合作 (cooperation), 共建 (jointly building) data circulation应用 (application) platforms. Also, hoping to establish类似 (similar) crowdsourcing mechanisms, allowing普通 (ordinary) people to collect data in生活生产 (daily life and production) scenarios. If this can成功发动 (successfully mobilize) societal forces, having millions or even tens of millions of people collectively collecting data, I think it might accelerate embodied AI development. Because only then can sufficient data be generated,推动 (driving) model iteration.
Wu Pengyang: Professor Li, from a university perspective, how do you view the上下游 (upstream and downstream) ecosystem situation?
Li Min: Wuhan, like HUST, has many graduates从业者 (working) in互联网 (internet), relatively more, which is an advantage. Also,武汉传统制造业 (Wuhan's traditional manufacturing) is very strong, including now光伏 (photovoltaic),光芯 (optoelectronics chips)新兴制造业 (emerging manufacturing) are also very good.
Actually, a crucial驱动力 (driving force) for embodied AI development is scenarios and applications. This relates to an important point:生产数据 (producing data), which greatly promotes embodied AI because data itself is infrastructure.
From the university perspective,国家和政府 (the state and government) attach great importance to supporting this area. For example, yesterday, the National Natural Science Foundation of China and Hubei Province established a joint fund, supporting one billion per year for five consecutive years, providing technical research support in humanoid robots, embodied AI-related directions. Huazhong University of Science and Technology also牵头 (led) a humanoid robot突破计划 (breakthrough plan),联合 (jointly) with Tsinghua, Zhejiang University, Beijing Institute of Technology, Dalian University of Technology, and other universities to conduct a series of research, totaling ten topics.
Returning to the ecosystem, the humanoid robot industry chain includes many以前是做 (previously involved in)制造自动化 (manufacturing automation), including motors, perception. In some核心零部件 (core components), like电传动 (electric drive),物理交互信息 (physical interaction information) perception, etc., Hubei Province has its own advantages. For example, companies incubated by our team specialize in触觉感知 (tactile perception). Currently, many灵巧手 (dexterous hands) in the industry use tactile perception基本来自 (basically from) this team. Hubei Province has a good foundation in these areas, with strong后劲 (potential).
Q4: Current status and future possibilities of "Using AI to create AI"?
Wu Pengyang: How is the combination of large models and embodied AI? How do you use large models to help improve product and organizational operational efficiency in your work?
Liu Chuanhou: We have basically been using large models from 2023 until now. When ChatGPT emerged, we were already exploring how to apply it. But GenAI说实话 (to be honest) has a big problem—significant deceptiveness; its answers aren't necessarily correct. So using AI to create AI involves significant cognitive issues. But this doesn't阻止 (prevent) us from using AI's technical capabilities. Our team's R&D also uses it, like in coding, management levels. Our company's R&D are indeed深度用户 (deep users), with a relatively open attitude.
Yang Songhua: I am a 00s entrepreneur, still a Ph.D. student, also doing much research, writing papers; this field恰好是 (happens to be) one of my research areas.
The general method for training traditional large models uses all human internet data. Models like GPT-4.0 onwards basically use AI-generated data for training, or use a more powerful teacher model to train weaker models.
For model training in embodied AI, there are mainly three types.
The first, similar to the large model approach, uses a more advanced model as a teacher model to fine-tune a locally self-trained small model.
The second, can use current advanced models like VLA to clean, filter, segment data.
The third is world models. Initially, world model-generated data was used for training. Later, people thought,既然 (since) they can generate mechanical motion trajectory data, why not directly use prediction results as output? Like autonomous driving,最初也是 (initially also) used world models to generate一些极端场景 (some extreme scenario) videos as补充训练数据 (supplementary training data). I think using world models to train embodied AI, this possibility is still很大 (great).
Wu Pengyang: Application status of合成数据 (synthetic data) in embodied AI?
Yang Songhua: Actually, much data is合成 (synthetic) data like simulations. But simulation isn't really using AI to create AI, because simulation requires大量 (vast) human involvement writing物理规则 (physical rules). However, simulation is mainly used for VLA pre-training. For精细操作 (fine operations), teleoperation data is still primary.
Ma Junjie: I'll be brief. We are a newly established embodied AI company, also an AI-native company. AI has entered our daily office work and entire R&D, from product to design to interfaces, with AI participation. Overall, using AI is effective, including efficiency提升不少 (significantly improved), so we are坚定的 (firm) AI-native company.
Wu Pengyang: What do you think is the difference between传统企业 (traditional enterprises) and AI-native enterprises?
Ma Junjie: AI's core for startups like us is first efficiency, another is投入产出成本 (input-output cost). For example, basic编程 (coding) can achieve人力 (manpower) 1/3 cost with 3倍 (times) the effect. But I should add, AI currently enhances our efficiency, but in some areas like market research, AI sometimes会反复重复 (repeats) previous content. So our关键数据 (key data) and决策 (decisions) must not过分依赖 (over-rely on) AI; AI can be a参考 (reference).
Wu Pengyang: Which tasks are given to AI, which rely on humans? Is there a mechanism区分 (to distinguish)?
Ma Junjie: In mechanism, AI participates in每一项工作 (every task), but core decision points still rely on humans. AI output is based on通识 (common knowledge), while correct decisions often需要反通识 (require counter-common knowledge). I think很多时候 (often) it still relies on humans, especially face-to-face交流 (communication).
Wu Pengyang: How high is the degree of autonomous execution by智能体 (agents) in your company?
Ma Junjie: This is不好量化 (hard to quantify). It can only be said that whether in daily office work, organizing activities, or design, there is an助手 (assistant) that can大幅提升 (significantly enhance) each person's work efficiency. Somewhat类似 (similar to) agent workgroup模式 (mode).
Li Min: Young people easily accept new things;他们会想方设法 (they will find ways) to use各种工具 (various tools) to save time and effort. In AI usage, from the simplest weekly reports,汇报 (presentations), PPTs, to code, to research ideas and思路 (approaches), to论文撰写 (paper writing), drawing, etc., our students are using it. They use各种智能体 (various agents), and we encourage and support, including buying memberships.
We also use it in teaching. Because there are many students, teachers'精力有限 (energy is limited). Sometimes we build agents ourselves, letting students first interact and discuss with agents, agents汇总出 (summarize) key core points, then we针对性 (targetedly) conduct deeper交流 (communication) with students. This saves time. For writing papers, students are unfamiliar with论文结构框架思路 (paper structure framework approach); AI can assist. Of course, I require they cannot directly use AI-generated images for assignments; they must learn to create their own.
But there's an important issue:先进工具 (advanced tools) can lead to人的懒惰 (human laziness). Sometimes students don't know why, thinking AI results are correct, thinking process完全被替代 (completely replaced), directly submitting results. Using things is okay, but some思路 (approaches), formulas still need to be mastered, like原理是什么 (what is the principle); students sometimes find it hard to understand.
AI usage needs to have a度 (degree), must control every关键步骤 (key step). Only then can AI serve us. It is ultimately a tool for提升效率 (improving efficiency). But we need to recognize负面效应 (negative effects) brought by tools,尽量规避 (try to avoid) them, to use tools to the极致 (extreme).
Wu Pengyang: In your teaching and guiding students, are there any methods to ensure AI is used correctly?
Li Min: Sometimes we create teaching materials ourselves, letting students use applications within our limited scope, not directly交流 (communicate) with unrestricted AI.
Summary and Outlook
Li Min: From a物理 (physics) perspective, physical interaction plus cloud-based intelligence will ultimately推动 (drive) the arrival of physical AGI's终局 (endgame).
Ma Junjie: Physical AGI will eventually arrive; we will all努力 (strive) for it. Although there are多种路径 (multiple paths),条条大路可能都通罗马 (all roads may lead to Rome),看谁先走通 (see who succeeds first).
Yang Songhua: Embodied AI might not have a奇点时刻 (singularity moment) that全面点燃 (completely ignites) the mass market like iPhone or ChatGPT, but rather a渐进过程 (gradual process), breaking through from certain scenarios step by step, like a GPT moment for industrial scenarios, a GPT moment for human-computer interaction scenarios.
Liu Chuanhou: Embodied AI, the journey is long and arduous.
This article is from the WeChat public account "Tencent Research Institute" (ID: cyberlawrc), author: Tencent Research Institute
