A shockwave echoed across the artificial intelligence (AI) industry in January as Chinese AI startup DeepSeek launched DeepSeek-R1, an open-source AI model that could match the capabilities of leading American chatbots. At a seminar held at Renmin University of China (RUC) in Beijing on February 23, Beijing Review reporter Peng Jiawei spoke to several Chinese scholars to discuss the potential impacts of the rise of DeepSeek on the tech world, global geopolitics and humanity as a whole. Edited excerpts of the interviewees' perspectives follow: 

From a technical point of view, apart from the fact that it is open source and requires less than $6 million worth of computing power, what does DeepSeek say about where generative AI with the ability to create original content is headed? 

Zhao Xin (professor at the Gaoling School of Artificial Intelligence, RUC): One notable characteristic of DeepSeek is that, if you select the "DeepThink" mode, it always goes through an extended step-by-step thought process before arriving at a conclusion.

The core of this mode of reasoning is enabling AI to think slowly. This is quite counter-intuitive because, naturally, we would expect an AI chatbot to deliver results as quickly as possible. 

To understand why AI is progressing from rapid-fire pre-trained responses to slow reasoning, we must first understand the scaling laws. (Under these laws, the performance of an AI model improves as the size of the training data, model parameters or computational resources increases—Ed.) However, the amount of data we can access to train AI models has limits. In fact, almost all of the public data we can access online has already been used to train existing models.

Given that simply scaling up our data size is no longer sufficient for performance improvements, the industry began seeking new growth paths, one of which is to give AI models more time to think.

This extension of thinking time is achieved through reinforcement learning, a branch of machine learning that forces AI agents to learn through trial and error. To understand this process, just imagine that you are being locked in a dark room by your instructor and given a set of problems with only the final results provided. The only thing you can do is to try different approaches and repeatedly adjust your reasoning until you arrive at the correct answer.

Just as humans employ the slow-thinking modality to tackle complex problems, the act of testing out solutions on its own can unlock new AI capabilities in solving ever-more difficult tasks.

If we look back at history, we see that the three previous industrial revolutions were driven by breakthroughs in basic sciences. Along this line of thinking, the next great leap in generative AI will very likely be its combination with basic sciences. If we were to apply DeepSeek and other slow-thinking models to scientific research, the ongoing Fourth Industrial Revolution would be further advanced. 

How has the rise of DeepSeek reshaped the global AI race? 

Wang Wen (Dean of the Chongyang Institute for Financial Studies, RUC): The emergence of DeepSeek is referred to in the U.S. as a real "Sputnik moment," one that may mark a new AI cold war between the two countries. (In 1957, the Soviet Union launched Sputnik, the first artificial satellite to orbit Earth, an event that started the space race between the Soviet Union and the United States during the Cold War—Ed.)

Indeed, DeepSeek has disrupted and may profoundly restructure the Western-centric balance of power in the global technology landscape. This achievement has challenged the deeply ingrained belief in the U.S.'s presumed AI supremacy and raises unsettling questions.

The American Government has restricted the export of semiconductor production equipment to China and barred Chinese students from entering the U.S. to study STEM (science, technology, engineering and mathematics). Yet it turned out that DeepSeek was able to create a model that rivals the capabilities of leading American chatbots with a team of entirely homegrown talent.

By breaking through the U.S.'s AI hegemony, DeepSeek not only showcases China's technological prowess but also heralds a new technology world order that is less centralized and more inclusive.

The fact that DeepSeek is open source means that other countries that were previously excluded from the AI race can all join the game, not just as competitors but also as collaborators. This approach will help establish a truly multilateral framework for future AI governance.

Cai Yimao (Dean of the School of Integrated Circuits at Peking University): While DeepSeek has brought many opportunities for multiple parties, it has also led us to believe that we can overcome the U.S.' choke points by simply reducing our reliance on computing power.

However, this view is deeply flawed. To always stay ahead of the curve, we must continue to shore up our computing power with solid advances in high-end chip manufacturing. If we abandon this major front, we risk becoming vulnerable and ill-equipped to adapt to the many technological leaps and revolutions that will arrive in this fast-moving field.

We hope to see more exploration by Chinese scientists in the direction of revolutionizing the basic model of computation. Our past chip designs are largely based on the Turing machine model (a hypothetical computational model introduced in 1936 by English mathematician and logician Alan Turing—Ed.) and the Von-Neumann architecture (a computer architecture proposed in 1945 by Hungarian-American mathematician and physicist John von Neumann—Ed.). In the future, however, we are likely to witness the continuous emergence of new methods of computer engineering.

These breakthroughs will help cultivate a homegrown ecosystem that encompasses a wide range of players in the semiconductor sector, from design and production to materials supply and software. The reason for U.S. AI chip maker NVIDIA's global dominance is its ability to build a robust ecosystem around its core processors. Hopes are high that, with the rise of DeepSeek and other home-grown innovations, China can gradually set up its own chip ecosystem.

Shen Yujing (associate research fellow of the Chongyang Institute for Financial Studies, RUC): "Can China beat the U.S. in the AI marathon?" is a question that the tech world has been asking for ages. Before DeepSeek, the prospect of China catching up to the U.S. seemed like a distant daydream. After DeepSeek, it suddenly became a plausible reality.

When it comes to the three pillars of AI development—data, algorithms and computing power—the consensus remains that the U.S. still holds the lead on all fronts. But China is closing the gap as its data growth rate is among the fastest in the world and it is stepping up building new data centers and cutting back on energy consumption in the training of AI models.

The country has also become the world's largest source of AI professionals, with nearly half of the world's top AI researchers having completed their undergraduate studies here.

While we are trying to pull ahead, we should always remember that China and the U.S. have adopted completely different strategies in AI development. The U.S. has relied on the traditional advantages of scale and capital, whereas China, as DeepSeek exemplifies, has prioritized cost-efficiency and openness.

In an AI-driven world, where ever more powerful digital minds are constantly emerging, how can humans maintain a level of dignity and self-worth? 

Xie Yun (chief scientist at TongTech, a Chinese provider of software infrastructure and applications): The foundation of AI is statistics. Simply put, statistical methods enable AI systems to calculate probability, detect patterns and infer conclusions from data. This approach is strictly limited by its data size and cannot make inferences beyond the scope of its sample.

Unlike humans, AI lacks the ability to grasp the fundamental nature of things. That's why scientists often say that the human brain is the deepest black hole in modern science.

The gap between AI reasoning and human insight can be likened to how Johannes Kepler and Isaac Newton differ in their scientific methods. By analyzing a massive amount of data, Kepler managed to deduce a set of laws that describe the motion of the planets. However, he could not have statistically discovered the law of gravitation. It was Newton who formulated this law, which was achieved not through statistical analysis, but through drawing connections between an apple's fall and the wider universe.

One thing we tend to forget about AI is that it is a craft, not a science. Like crafts, AI development is based not on scientific theories, but on experience and empirical learning.

As a saying goes, "You can't get to the moon by climbing successively taller trees." The above two features of AI determine what it can and cannot do, where it is headed and how far it can travel.

Wang: New technologies do not just change the game. They rewrite the rules by triggering structural shifts in certain industries, countries, fields and even the entire world.

One thing we know for sure is that there will be more DeepSeek-like innovations in the future. For most of us, the only thing we can do is to adapt and brace ourselves for the uncertainties that lie ahead. We are standing at a new juncture in time and are all confronted with the task of finding our own worth.

Let us embrace AI, learn from AI and evolve with AI—so that at the very least, we do not become the ones that are left behind.

Copyedited by G.P. Wilson 

Comments to pengjiawei@cicgamericas.com