👀China’s AI Looks Vibrant, but It’s Fundamentally Chaotic, Scientist Warns

Song-Chun Zhu (朱松纯), a prominent Chinese computer scientist, warns that China’s current AI narrative has “cognitive bias” at the 2025 ZGC Forum in Beijing.

Hi, this is Tony! Welcome to this issue of Recode China AI (for the week of March 24, 2025), your go-to newsletter for the latest AI news and research in China.

Song-Chun Zhu (朱松纯), a prominent Chinese computer scientist and the founding director of the Beijing Institute for General Artificial Intelligence, delivered a thought-provoking speech at the 2025 Zhongguancun Forum (ZGC Forum) in Beijing on March 29. He warned that China’s current AI narrative suffers from “cognitive bias.” Despite the early success of companies like DeepSeek and Monica AI (creator of Manus), Zhu argued that China still lacks original breakthroughs in foundational AI technologies. He voiced similar concerns in a previous interview translated by CSET.

Song-Chun Zhu, Founding Director of the Beijing Institute for General Artificial Intelligence. Source: zhusongchun.net

In my view, Zhu’s speech was an call for attention—and support—for academic research. And I get where he’s coming from. The rise of DeepSeek has reinforced a popular belief: that China’s best bet to catch up with the U.S. in AI is to follow closely behind American innovation, then rely on engineering scale and cost-efficiency to catch up. In that narrative, industry and commercial labs take center stage—leaving little room for fundamental academic breakthroughs.

While I don’t fully agree with Zhu—especially when he seems to pit industry labs against academic ones—his words made me reflect on my own work. Most of my newsletter articles tend to highlight industry achievements—they’re more headline-worthy and accessible. But that also means we may be overlooking the deeper worries of front-line researchers in China. That’s why I’ve decided to translate his latest speech, which was edited and published by Tencent News. I believe it deserves more attention.

Here’s the full translation below.

---

Key Highlights

• We need to cautiously address society and policy’s perception of AI, as this cognitive bias is increasingly distancing us from genuine AI innovation.

• The current state of AI in China is vibrant on the surface but fundamentally chaotic. At present, government institutions, the public, and even the media severely lack a deep understanding of artificial intelligence.

• Much of today’s so-called AI innovation is stuck at the algorithm and deployment levels, lacking genuinely original breakthroughs in cognitive modeling and the nature of intelligence.

• While companies like DeepSeek have achieved success in engineering implementation, true breakthroughs in AI still depend on the academia’s long-term dedication to understanding intelligence fundamentals, cognitive modeling, and theoretical foundations.

• The most challenging future problems for AI lie within social complexity systems, such as demographics, policies, civilizations, and value systems—areas traditionally addressed by the humanities. Due to difficulties in effective modeling and experimentation, academia mostly relies on retrospective analysis, with prediction capabilities nearing zero.

I particularly value communication with industry and media because I deeply understand the importance of the industry's narrative logic to societal cognition.

Artificial intelligence has shifted from an initially academic issue to a comprehensive social and policy-related matter. Most decision-makers, institutions, and media we face today do not come from AI backgrounds but are forced to rapidly learn, understand, make decisions, and even publicly promote AI, causing immense difficulty in accurate information dissemination.

Thus, we must clearly establish the correct narrative logic, especially for the media, which carries a significant responsibility. This is crucial to avoid erroneous public opinions influencing decision-making, or worse, allowing self-media to lead the narrative, resulting in confusion among the public and policymakers regarding AI.

Here, I want to discuss global AI narratives and misconceptions:

AI: Hype, Bubbles, and Blind Trends

Since 2019 or even earlier, the US has reaffirmed its technological hegemony through AI. Global capital has flowed toward the US for the past seven to eight years, centered on the narrative: big data + massive computing power + large models eventually leading to Artificial General Intelligence (AGI).

Subsequently, narratives about “AGI posing existential threats to humanity” gained widespread hype, creating global anxiety. This is essentially a media-driven narrative. Over the past decade, media and investors have repeatedly promoted the trinity of big data, computing power, and large models, as though it represents the sole future of AI.

In 2015-2016, AlphaGo triggered the first AI boom. Yet, after eight years, apart from inflating valuations for some computer vision companies (the so-called Four Little Dragons), it has not resulted in extensive industrial or social impact. Many AI startups eventually declined.

The currently popular narrative “AI for Science” also contains cognitive biases. AI's primary research directions remain in intelligent perception and actions like vision, language, and robotics, which do not directly equate to science. What truly advances scientific research is “Deep Learning for Science”—using tools like deep learning to support scientific modeling and data analysis—not AI itself becoming an intrinsic part of science.

I once wrote an article introducing the idea of the “Crow vs. Parrot Paradigm,” criticizing the current AI trend of large-scale mimicry (parrot-style) that remains fundamentally different from true cognition and reasoning (crow-style). I called it a “bubble-driven funding myth.” Looking back now, many of the same patterns are repeating.

In the midst of the AI boom, it’s easy to see an oversupply of AI platforms and computing power centers. Many of these platforms can’t even be rented out, with actual usage rates as low as 15% to 20%. Even more absurd, in some regions, electricity prices have turned negative—there’s more power than demand. So how can there be an energy crisis?

But why did so many places still rush to launch these projects? The root cause lies in the narrative shaped by public discourse. In some cases, local decision-makers may have felt pressured by the prevailing rhetoric, while the media’s “overhype” only added fuel to the fire.

China’s AI: Current Situation and Confusion

Today’s domestic AI sector is vibrant superficially but fundamentally chaotic.

In recent years, China has established numerous “AI institutes,” ironically led by figures who lack real AI expertise. For instance, one prominent university appointed a renowned theoretical computer scientist as a part-time AI dean, despite no formal AI research publications. Other institutions have simply assigned professors from mathematics or art departments to serve as part-time AI institute deans.

This resembles the earlier craze over “nano” concepts—such as nano-shoe insoles and nano-pressure cookers. Currently, we are experiencing a similar pseudo-AI frenzy. Some LLM companies self-brand as “Six Little Dragons,” yet many remain unprofitable with dangerously inflated valuations.

The commonly heard phrase “we are being choked by others” actually reflects our cognitive limitations rather than genuine external constraints. Currently, government bodies, the public, and even media severely lack AI understanding, blindly adopting Western narratives, ultimately leading to misguided conclusions about being technologically “choked.”

What Truly Constitutes AI Innovation?

I have long categorized AI innovation into five levels:

• Philosophical Level: Discussing the fundamental essence of “intelligence.” Intelligence is subjective, with decisions rooted in individuals' cognition and value systems.

• Theoretical Level: Establishing mathematical frameworks for cognition, such as logic, statistical modeling, and probabilistic computing.

• Model Level: Constructing specific models (discriminative models, generative models, large models) based on these frameworks.

• Algorithm Level: Developing optimized algorithms under specific models to enhance computational, inferential, and training efficiency.

• Engineering & Deployment Level: Implementing models on hardware and platforms, optimizing storage and computation to create practical products and systems.

Many current innovations only reach the fourth (algorithm) or fifth (deployment) levels, lacking theoretical foundations yet boasting about being “revolutionary.” Genuine innovation demands original breakthroughs in understanding the essence of intelligence and cognitive modeling.

There’s a serious misconception in society today—one that suggests only companies like DeepSeek have delivered meaningful results. Some even go so far as to claim that academia and research institutions are just “freeloaders.” This kind of emotional and irrational rhetoric is misleading the public.

We must clarify that while DeepSeek has indeed achieved success in areas like engineering implementation, API productization, and compute optimization, its contributions are primarily at the deployment level. It has not addressed AI’s core challenges—such as cognitive modeling, intelligence theory, or learning mechanisms.

The foundation beneath all current AI applications has been built on decades of academic research in philosophy, theory, modeling, and algorithms. To dismiss this foundational work just because of short-term product successes—and worse, to promote the idea that academic research is useless—is not only absurd, but deeply dangerous.

Take the U.S., for example. Much of its innovation has focused on foundational layers like hardware (chips, architecture), large models, and algorithmic optimization. If we want to make true breakthroughs in the U.S.-China AI competition, the key lies in higher-level innovation—at the level of theory and philosophy.

If we merely follow the well-worn American path—computing power, algorithms, deployment—we will always remain followers.

Future Frontiers: The Humanities

The most difficult problems of the future lie precisely in the complex social systems that the humanities and social sciences focus on—such as population, policy, civilizational evolution, and value systems. These issues are currently impossible to model or experiment on. If academia continues to rely solely on verbal explanations and retrospective analysis—essentially playing armchair quarterback after the fact—its predictive power will remain close to zero.

However, today, large-scale simulation experiments and agent-based modeling are beginning to make it possible for the humanities to evolve into an experimental science for the first time. The true frontier of AI lies in using simulation and modeling to bring civilization, society, economy, and policy into the realm of testable, verifiable science—not just optimizing image, speech, and dialogue systems.

Our True Focus

We have now largely completed the initial construction of our philosophy of intelligence, theoretical framework, and core models. The focus is shifting toward algorithm optimization and engineering deployment.

Our next priorities are:

• Rapid scaling

• Full engineering and commercialization of the model

• Deep integration with industries and real-world applications

• Establishing a factory for AGI agents (general-purpose intelligent agents)

Our relationship with large models is not one of opposition, but of symbiosis. Large models are like the human subconscious—they provide the foundation for perception and memory. What we aim to build on top of that is the cognitive and decision-making architecture for general-purpose intelligent agents.

Summary and Appeal

Artificial intelligence is neither a myth nor synonymous with existential threats. It is a tool vital for humanity's future evolution.

Misguided narratives and cognitive deficiencies constitute the real chokepoints.

I appeal to media, government institutions, and researchers to prioritize foundational logic and original innovation in AI, considering carefully: what kind of AI does China genuinely want?