Interview with Tan Tieniu, Standing Committee Member of the Chinese People's Political Consultative Conference and Party Secretary of Nanjing University: AI Development Needs to Remove the Bubble

On March 5th, the government work report was released, mentioning “artificial intelligence” multiple times, and embodied intelligence was once again included in the report.

Focusing on hot topics in artificial intelligence and embodied intelligence, Daily Economic News (NBD) reporter (hereinafter referred to as NBD) interviewed Tan Tieniu, member of the 14th National Committee of the Chinese People’s Political Consultative Conference, Academician of the Chinese Academy of Sciences, and Secretary of Nanjing University, during the National Two Sessions.

Tan Tieniu previously served as Vice President of the Chinese Academy of Sciences; in August 2022, he received the Jujing Sun Award, the highest award in the field of pattern recognition internationally—awarded for the first time since its establishment in 1988 to scholars outside North America and Europe.

This year marks Tan Tieniu’s 40th year working in artificial intelligence. From early image recognition to later biological feature recognition and video analysis, he has continuously pioneered new research directions. He was among the earliest scholars in China to work on iris and gait recognition, with his research results widely applied in coal mines, criminal investigations, and other key fields.

In the interview, Tan Tieniu lamented that the progress of AI technology in recent years has indeed been “beyond imagination,” “unexpected.” He recalled, “About ten years ago, we still regarded natural language interaction between humans and robots as an important goal. Now, this problem has basically been solved.”

But he also warned, “This does not mean that AI is万能 (all-powerful) now; it still has many ‘cannot’—although, without AI, many things would be impossible.” Tan Tieniu emphasized that developing AI should be “rational and pragmatic, not follow the trend blindly, adapt to local conditions, and focus on implementation,” ensuring AI is used for good, truly promotes new productive forces, and supports Chinese-style modernization.

1

It’s still too early to talk about industry maturity

NBD: This year’s Spring Festival Gala stage featured robot performances again, sparking nationwide discussion. What signals do you think this release conveys? Is this concentrated exposure a sign of industry maturity?

Tan Tieniu: Humanoid robots are indeed a hot area of technological revolution and industrial change, widely appreciated by the public. But we need to see the underlying meaning and not just enjoy the spectacle.

First, it’s worth noting that from last year’s “YangBot” to this year’s “WuBot” [Bot is short for robot], the level demonstrated by Yushu humanoid robots is impressive. They went from unsteady walking to somersaults within a year, with rapid progress. This fully reflects our achievements in自主创新 (independent innovation), proving that Chinese people can also lead global technological innovation trends. So, we should maintain confidence in自主创新, and in this wave of humanoid robot enthusiasm, at least in movement and control capabilities, we are already among the world’s front runners.

Image source: video screenshot

But we also need to view this objectively. Humanoid robots are not equivalent to AI. Seeing them perform flips and other cool actions, one might think they possess high intelligence. This needs clarification: humanoid robots are not equal to AI.

Robots and intelligence are two highly related but conceptually different ideas. Robots do not necessarily have intelligence; they are more like carriers that can host AI capabilities. Only robots with certain intelligence are called intelligent robots, and the same applies to humanoid robots.

Currently, the popular humanoid robots mainly demonstrate advances in control and movement capabilities, similar to drone formations, which are pre-programmed, pre-trained in known processes, actions, and scenarios. They do not fully represent AI progress. If during a performance, a prop’s position is suddenly changed, the robot might fail to respond. If it can autonomously find the prop, that would be a higher level, a true manifestation of AI.

Therefore, it’s still premature to talk about industry maturity. I believe that if humanoid robots only dance and do flips, they will ultimately be fleeting. It’s essential to find killer applications. Many orders after the Spring Festival Gala are not surprising, but the novelty and curiosity won’t last. The key is whether it’s a real necessity and can solve actual problems. Without killer applications, they will eventually be eliminated by history.

A lesson from history is worth noting. Japan started early in humanoid robot research, launching the globally renowned “ASIMO” in 2000. But after 22 years, due to high costs and limited practicality, it failed to find killer applications and eventually exited the stage.

Of course, Yushu robots have far surpassed ASIMO in movement control, but their intelligence level remains limited.

2

Smart Humanoid Robots

Mass adoption in households will take at least 5 more years

NBD: What fields do you think these killer applications will appear in?

Tan Tieniu: There are many, such as manufacturing, inspection—like inspecting roads, high-speed rail, high-voltage lines. But inspection tasks require high standards; robots need to have “sharp eyes,” meaning strong visual capabilities and fast processing speeds. This involves not only control and movement but also environmental perception and understanding. Some applications are already being implemented, but open scenarios still pose challenges.

NBD: What key obstacles remain before robots truly enter homes and factories? When do you think highly intelligent robots will reach thousands of households?

Tan Tieniu: Robots are already entering homes, like common vacuum robots with certain intelligence. But for humanoid robots to truly enter homes, they need to help with more chores and achieve seamless human-machine collaboration. There are many hurdles. I believe at least five more years are needed.

The reason is that robots must have strong scene perception abilities. They need to understand their surroundings, know their own location, and interpret human intentions and actions. They shouldn’t block pathways or start pouring hot water when someone is about to take a cup. They must judge human behavior and intentions, which is very difficult. If they don’t understand what others want to do, collaboration is impossible, and misoperations could pose risks.

Another key shortcoming is dexterous manipulation, especially the “hands.” Current tactile sensing capabilities are far from enough to accurately perceive smoothness, material, temperature, humidity, etc. When will humanoid robots be able to play ping-pong with humans and win? That’s when I’d truly admire them, but we are still far from that.

NBD: Industry optimism suggests that in 3-5 years, intelligent robots will enter households. Do you think this optimism is overly optimistic or influenced by hype?

Tan Tieniu: Some believe there’s a bubble in the industry, and I agree. I see three bubbles:

First, expectation bubbles. People have high hopes for AI and humanoid robots. The recent progress has indeed been beyond expectations, but that doesn’t mean AI is万能 (all-powerful). Rapid development leads some to believe that general AI (AGI) will be achieved within two or a few years, which is overly optimistic.

Second, investment bubbles. OpenAI has burned through huge amounts of funding and has yet to turn a profit.

Third, valuation bubbles. Despite not being profitable, OpenAI’s valuation has soared to hundreds of billions of dollars—obviously inflated. Some AI companies, with only decent products, are valued at tens of billions, which is clearly overhyped. Media hype and self-media amplification also contribute to the bubble.

Herbert A. Simon, Nobel laureate in economics and Turing Award winner, predicted during the first AI wave in 1965 that machines would do all human work within 20 years. That prediction has not come true. It shows that in the hype, rationality is needed.

3

Achieving Artificial General Intelligence (AGI) remains a long-term challenge

NBD: You once said “Elon and others are overly optimistic,” and believe AGI is still far off. Yet, the industry’s pursuit of AGI remains intense. Between “rational pragmatism” and “technological idealism,” how should China’s AI development pace be managed?

Tan Tieniu: The key is how to define AGI. My definition: AI that can match and surpass human intelligence (wisdom). It should at least be comparable to humans, capable of doing everything humans can do. If defined this way, I think it’s difficult to realize in the foreseeable future.

The reason is that humans have insights, common sense, can infer from one thing to another, and understand context—especially grasping implied meanings and reading between the lines. Current AI sometimes lacks even basic common sense because it is trained on big data and does not truly understand causal relationships and physical laws of the material world.

“Intelligence” currently lacks a unified definition, and the mechanisms behind human intelligence and wisdom are not fully understood. Surpassing something that is not fully understood is logically questionable. Superficial surpassing is possible, but surface-level capabilities cannot be exhaustively tested. There is also a misconception that AI already has consciousness and emotions; in reality, it only mimics them. Mimicry does not mean possession or understanding.

I have two doubts about AGI:

First, is AGI actually needed in practical applications? Simply put, “general” means capable of doing everything. In my view, the answer is no, because specialization is more effective. We aim to cultivate versatile talents, but that doesn’t mean everyone can do everything well—there’s no such thing as a “generalist.”

Therefore, why not develop a group of highly specialized AI agents, each with a clear focus, working together? Even in household scenarios, cooking, cleaning, and elderly care could be handled by different specialized agents. They would be multi-purpose, not truly general.

Second, can AGI be achieved? Since AGI must surpass human intelligence, and human intelligence mechanisms are not fully understood, how can it be surpassed? So, I believe that AGI remains a long-term or even distant goal.

4

Embodied Intelligence is the inevitable path to approaching human-level intelligence

NBD: There’s a view that embodied intelligence is a necessary stage for achieving AGI. What’s your opinion?

Tan Tieniu: Of course, if the goal is to infinitely approach human intelligence and wisdom, embodied intelligence is a path—or rather, an essential route. However, the term “embodied intelligence” is sometimes misused or overly labeled.

Embodied intelligence has two core elements: first, having a physical body—an observable, tangible physical entity; second, continuous interaction with the environment, becoming smarter through “learning by doing.” If interaction with the environment is only to perform predefined tasks, that cannot be called embodied intelligence.

There are misunderstandings that having a physical entity and some intelligence is enough to be called embodied intelligence. This is incorrect. We need to distinguish between robots and AI, robots and intelligent robots, intelligent robots and embodied intelligent robots.

Simply put, a robot is a hardware entity. An intelligent robot is a robot with added intelligence capabilities. For example, a typical industrial robotic arm is pre-programmed and lacks intelligence; an intelligent mechanical arm that can autonomously avoid obstacles and adjust its path while grasping objects is an intelligent robot.

What’s the difference between embodied intelligence robots and intelligent robots? First, if a robot is physical, it is embodied. Embodied intelligent robots must become smarter through interaction with the environment, learning continuously and acquiring abilities not pre-programmed. If its capabilities are fixed and preloaded, it’s just an intelligent robot, not embodied intelligence.

Why is embodied intelligence considered the path to approaching human intelligence? Because human intelligence evolved this way. To approach human-level intelligence, the most direct and effective way is to learn and evolve like humans do, which could lead to surpassing human intelligence.

My understanding of embodied intelligence has evolved over recent years. Initially, I was skeptical because human and animal natural intelligence develop gradually through experience and adversity. Only through wind and rain, seeing the world, can one grow skills. Isn’t this the core of embodied intelligence? The character “智” (wisdom) itself contains “知” (know) and “日” (day), symbolizing daily personal experience. Wisdom, intelligence, and smartness all mean gaining through experience and practice.

Therefore, the essence of embodied intelligence is to gain dynamic improvement through interaction with the external world. If there is only interaction but the intelligence level remains fixed, it can only be called an intelligent robot, not embodied intelligence. For example, a robotic hand picking up a cup involves interaction, but if it doesn’t learn how to grip more tightly, lacks tactile sensing and feedback, it cannot be called embodied intelligence.

5

Stacking computing power and data alone is unsustainable

NBD: You often mention intelligence. Can this intelligence be understood as the capability of large models? What role do large models play in embodied intelligence? Is there a risk of over-reliance on large models?

Tan Tieniu: Several concepts need clarification here. Large models are not equivalent to AI; embodied intelligence is a development path, a way to develop AI, and an essential route to approach human intelligence.

Large models are the core technology of the current AI boom. They are based on deep neural networks, simulating the layered information processing of the human brain, learning from coarse to fine, from broad to precise. Large models can be roughly understood as huge neural networks with massive parameters, trained on vast data, mimicking human neural networks. They are just one way to realize AI, not the whole picture. Mimicking human intelligence doesn’t necessarily require simulating the human neural network; that’s the most direct approach.

Last April, I proposed that relying solely on stacking computing power and data to develop AI is unsustainable. This has been increasingly validated: first, performance improvements are diminishing; second, the required computing power is unsustainable; third, data availability is reaching limits. Almost all usable data on the internet has been utilized. Every physical system has its limits, and new approaches are needed.

The surge in GPU prices reflects the public’s perception of the reliance on raw computing power. [Image source: Meijing Media Library]

DeepSeek’s success is partly because it did not fully depend on stacking computing power and data but instead innovated algorithms, achieving comparable or better results with fewer chips and less data.

Large models cannot keep growing indefinitely; there are scale limits. Therefore, alternative approaches are necessary. Embodied intelligence is one such path—it does not rely solely on existing internet data but dynamically acquires data through interaction with the environment, such as perceiving material and smoothness when grasping a cup.

6

In the next 3-5 years,

Focus on breakthroughs in sensing technology and brain-computer interfaces

NBD: In the next 3-5 years, what disruptive breakthroughs in AI and embodied intelligence do you think are most worth paying attention to?

Tan Tieniu: I believe several directions are worth watching.

First, breakthroughs in fundamental structures and new machine learning paradigms. Relying solely on data is unsustainable; new models combining data and rules are needed. Integrate data-driven and rule-based approaches—use rules for deterministic parts, data for uncertain parts. Also, explore new models that combine data and knowledge, driven by both, which will be key in the next 3-5 years.

Second, breakthroughs in sensing technology, especially high-sensitivity, multifunctional sensors. This directly impacts the capabilities of end-effectors like dexterous hands, crucial for embodied intelligence.

Third, new machine learning methods that are low-cost and efficient, reducing dependence on massive computing power and data. Insights from brain and cognitive sciences could lead to new intelligent models, possibly moving beyond Transformer architectures, opening new development paths that do not rely on large models.

Additionally, multi-agent systems and human-machine collaboration are important areas to watch. Brain-computer interfaces that facilitate better interaction and cooperation between humans and machines may also see breakthroughs.

7

To avoid the “AI gap” caused by disparities in AI endowment conditions

NBD: What’s your advice regarding the societal anxiety that AI will replace humans, especially concerns that embodied intelligence might replace blue-collar jobs? We note that Nanjing University is promoting “1+X+Y” AI literacy education. Can this general education meet the talent needs of the AI era?

Tan Tieniu: It’s inevitable that AI will replace some jobs—that’s normal in technological progress. But overall, it’s not about destroying all human employment.

The World Economic Forum’s “Future of Jobs Report 2025” predicts that between 2025 and 2030, about 92 million jobs worldwide will be replaced, but 170 million new jobs will be created. Historical experience shows that technological progress replaces some jobs in the short term but increases employment in the long run, leading to structural optimization.

However, the new jobs created may not be accessible to those displaced. If people do not pursue lifelong learning and re-skilling, unemployment risks increase. Conversely, proactive planning, on-the-job training, innovative curricula, and new training models can help workers adapt to new demands.

Image source: video screenshot of human-machine collaborative inspection

Therefore, Nanjing University’s talent training emphasizes “three adaptations”: first, adapt to national needs—adjust majors accordingly; second, adapt to the era—this era’s most prominent feature is intelligence. In 2024, we will launch the country’s first compulsory AI literacy course for all majors, and train teachers, because AI will淘汰 (eliminate) those who do not learn AI; third, adapt to student development—personalized education.

This is the fundamental logic of our educational reform, not just for hype.

NBD: Do you have other suggestions or reflections on AI development?

Tan Tieniu: I believe there are several points to emphasize—

First, pay special attention to regional and industry disparities in AI empowerment, to avoid widening the “AI gap,” which could exacerbate regional and sectoral imbalances and intensify social conflicts.

Second, expanding domestic demand is the primary driver of economic growth. We should vigorously promote AI-enabled consumption, creating new consumption scenarios, such as home services, elderly care (“aging”), and education (“children”). For example, if companion robots truly become empathetic, safe, and affordable, they could be killer applications. But many issues remain, including standards, ethics, and safety, which need to be addressed gradually during development.

Reporter | Zhang Rui

Editor | Wen Duo

Visual | Chen Guanyu

Layout | Wen Duo

Overall Coordination | Yi Qijiang