Song-chun Zhu warns against misguided AI narrative in China
Dean of Peking University's Institute for Artificial Intelligence emphasizes the importance of basic research, warns AI innovation shouldn't be reduced to mere computational power.
I recently came across a speech on AI that deeply resonated with me. While AI is not usually on the radar for this newsletter, this piece is too good to keep to myself.
In recent months, the rise of Deepseek has captivated the minds of many in China. However, Song-chun Zhu, Dean of Peking University's Institute for Artificial Intelligence (IAI), recently cautioned that the current narrative surrounding AI in Chinese society is misguided. He emphasized the need for a more cautious approach in both societal and policy discussions on AI.
At the 2025 Zhongguancun (ZGC) Forum on General Artificial Intelligence held on March 29, Dean Zhu highlighted that discussions concerning AI have become overly focused on the capabilities of large models, while basic research, original innovations, and the essence of intelligence have been sidelined. Alarmingly, some have even dismissed the importance of these foundational areas.
Moreover, he noted the rise of what he called "technological speculation," where only a few companies are seen as representing the true state of AI development in China, while the academic community, including researchers and cognitive scientists, the true driving force behind AI advancements, is often overlooked. This cognitive bias, Zhu warned, is steering us further from genuine AI innovation.
Zhu summarized AI innovation into five levels: philosophy, theory, models, algorithms, and engineering/deployment. He argued that many of today’s so-called "innovations" are limited to the algorithm or deployment levels and lack a solid theoretical framework.
He further added that the foundation supporting all AI applications today is built on decades of academic investment in philosophy, theory, modeling, and algorithms. Denying the value of basic research based on short-term productization results—and even promoting the notion that academia is "useless" — is not only absurd but also dangerously misguided.
In conclusion, Professor Zhu calls for attention from all sectors of society, including the media, government agencies, and researchers, to focus on the top-level logic and original innovations in AI, rather than blindly following existing narratives.
This article was originally published on the official WeChat account of Tencent Technology 腾讯科技 but was removed shortly thereafter. It remains accessible within China's Great Firewall.
Below is my translation of the speech.
北大人工智能研究院朱松纯：“中国的AI叙事” 存在认知偏差
Song-chun Zhu of Peking University's AI Institute: Cognitive Biases at play in "China's AI Narrative"
I attach great importance to engaging with industry and media because I deeply understand how the "narrative logic" of an industry significantly shapes public perception.
Artificial intelligence (AI) has evolved from a purely academic topic into a widespread societal and policy issue. Most decision-makers, institutions, and media we deal with typically do not come from professional AI backgrounds. Yet, they are compelled to learn, understand, make decisions, and even promote AI to the public in a short period of time, which complicates accurate communication.
Thus, it is essential for us to establish a clear and accurate narrative, especially given the critical role of media. Doing so helps avoid misinformation influencing decision-making or allowing self-media narratives to dominate, which would result in confusion about AI among the public and policymakers alike.
Now, I'd like to discuss the global narrative around AI and address some common misconceptions.
AI is marked by hype, bubbles, and blind imitation
Since 2019 or even earlier, the U.S. has re-established technological dominance through AI. Over the past seven to eight years, global capital has flowed into the U.S., driven by a core narrative emphasizing big data, massive computing power, and large models, ultimately aimed at achieving Artificial General Intelligence (AGI).
This narrative was followed by widespread media promotion of the idea that AGI could threaten human existence, which created global anxiety. In fact, this portrayal has been extensively promoted by media and investment circles, which have persistently emphasized the "big data, massive computing power, large models" framework over the past decade -- as if it were the sole path forward for AI.
The first AI boom occurred around 2015-2016 with AlphaGo. However, looking back eight years later, the hype around AlphaGo and its related industries didn't result in widespread industrial or societal impact. Aside from a few computer vision companies (the so-called "Four Dragons", namely CloudWalk, SenseTime, Megvii and Yitu Technology) getting inflated valuations, many AI startups eventually declined.
The industry's popular concept of "AI for Science (a new paradigm for scientific research)" also involves cognitive bias. AI research today remains primarily focused on intelligent perception and actions -- such as vision, language, and robotics -- which, strictly speaking, are not directly equivalent to "Science." What truly advances scientific research is "Deep Learning for Science," referring to tools like deep learning that assist in scientific modeling and data analysis. In other words, AI itself is not inherently part of the scientific process.
I once wrote an article about the "Raven vs. Parrot paradigm," where I pointed out that current AI largely stays within the "parrot-like" stage of large-scale imitation, still fundamentally distant from the "raven-like" cognitive and reasoning abilities. I criticized the inflated myths driving bubble-like funding at the time, and today, many of these issues are still repeating.
Amid the AI boom, there's an oversupply in AI platforms and computing power centers. Many platforms can't even rent out their capacity, with actual usage rates only around 15-20%. Even more absurdly, in some areas, electricity prices have become negative, with electricity not being sold -- so how can there be an energy crisis?
Why, then, did so many places rush to jump on the bandwagon? The root cause lies in the narrative shaped by public opinion. Decision-makers in some places may have succumbed to public opinion pressures, with exaggerated portrayals by the media further amplifying the rush into AI investments.
Current state and challenges of AI in China
The current AI landscape in China is characterized by surface-level excitement masking underlying disorder.
In recent years, numerous "AI institutes" have been established across China. Ironically, many of these institutes are headed by individuals who aren't even specialists in artificial intelligence. For example, a university appointed a well-known theoretical computer scientist as the dean of its AI institute, despite this expert having never formally published research in AI. Other universities have even assigned faculty members from mathematics or arts departments to serve as "part-time deans" of their AI institutes.
This situation resembles the earlier "nano" craze, where everything was branded as "nano," from insoles to pressure cookers. Now, we are seeing a similar "pseudo-AI hype." Take some large model companies, for instance -- they proudly label themselves as "six little dragons," but many are not even profitable, with inflated valuations and enormous risks.
It is often said that China is "choked," but I believe the real problem is our own perception.
Currently, government agencies, the public, and even the media have a serious lack of understanding of AI. Following the Western narrative blindly has led to the conclusion that "we're choked." The real issue, however, is that our cognitive level is far from adequate to guide proper innovation and strategy.
What constitutes true innovation in AI?
When it comes to AI innovation, I have identified five levels:
The foundational level is the philosophical aspect: exploring the nature of "intelligence." In fact, intelligence is inherently "subjective" -- each person's decision is based on their own perceptions of the world and personal value systems. These perceptions may not be objective, yet they shape behavior.
The second level is theoretical: developing a mathematical framework for cognition, including logic, statistical modeling, and probability calculations.
The third level is the model-based: constructing specific models based on the framework, such as discriminative models, generative models, and large models.
The fourth level is the algorithm level: developing optimization algorithms within the models to improve the efficiency of computation, reasoning, and training.
The fifth level is engineering and deployment: implementing the models onto hardware and platforms, optimizing storage and computation to create usable products and systems.
Currently, many so-called innovations remain limited to the fourth (algorithm) or fifth (deployment) levels. They lack a solid theoretical foundation yet are still marketed as "disruptive." What we genuinely need today are original breakthroughs that deepen our understanding of the essence of intelligence and cognitive modeling.
There's a widespread misconception in society today suggesting that only companies like DeepSeek have made meaningful progress. Some even go to the extreme of labeling the efforts of academia and research institutions as "useless." This emotionally driven, irrational narrative is seriously misleading the public.
We must clarify that while DeepSeek has indeed made advancements in engineering implementation, API productization, and computational optimization, these achievements are primarily at the deployment level. They have not addressed AI's core challenges, such as cognitive modeling, intelligence theory, and learning mechanisms.
The foundation underlying all current AI applications is built upon decades of continuous academic investment in philosophy, theory, modeling, and algorithms. To dismiss basic research due to short-term product success, or even to promote the notion that “academia is useless,” is not only absurd but extremely dangerous.
Take U.S. innovation as an example: much of it focuses on the foundational layers, like hardware (chips, architecture), large models, and algorithm optimization. If we want to make breakthroughs in the China-U.S. competition, the key lies in philosophical and theoretical innovation at the higher levels, beyond just algorithms and deployment.
If we merely follow the established U.S. Approach -- focusing on computing power, algorithms, and deployment—we will always remain followers.
The future frontier of AI lies in the humanities
The most challenging problems ahead are those concerning complex social systems, such as population, policy, civilization evolution, and value systems -- areas that the humanities focus on. These issues are currently difficult to model or experiment with. Academia has long relied on "oral explanations" and "post-event analysis," with almost no predictive capability.
However, today, large-scale simulation experiments and agent modeling offer the possibility of turning the humanities into an experimental science for the first time. The true frontier of AI is not limited to optimizing images, speech, and conversations, but in using simulations and modeling to bring civilization, society, economics, and policy into a verifiable scientific domain.
Our current main focus
We have largely completed the initial work in intelligent philosophy, theoretical frameworks, and model building, and are now moving forward with algorithm optimization and engineering deployment.
The next priorities are:
Rapid scaling.
Engineering and commercializing the models.
Deep integration with industries-specific applications.
Establishing a factory for AGI Agents.
Our relationship with large models is not one of "opposition" but of symbiosis. Large models function like the human subconscious, providing a foundation for perception and memory. Our task is to build cognitive and decision-making systems for general AI agents on top of that foundation.
Summary
In conclusion, AI is not a "myth" nor is it synonymous with "security crises" or "survival crises." It is a tool that truly impacts the future evolution of human civilization.
The real "choke point" lies in our lack of understanding and the misleading narratives around AI.
I urge the media, government agencies, and researchers to focus on the top-level logic and original innovations in AI. We should not blindly follow existing narratives but instead ask ourselves: What kind of AI does China truly need?