Where Does China Stand in the AI Wave?
China’s top policy experts discuss the US-China gap, open vs. closed, and societal implications.
This piece was authored by “Bit Wise,” an anonymous ChinaTalk contributor focusing on China’s AI policy.
Debates and discussions by Western public intellectuals on AI governance are closely followed in China. Whenever prominent figures like Sam Altman, Joshua Bengio, or Stuart Russell give interviews, multiple Chinese media outlets swiftly translate and analyze their remarks.
English-speaking audiences, however, seldom engage with the AI governance perspectives offered by Chinese public intellectuals.
In this article, ChinaTalk presents the highlights and a full translation of a panel discussion on AI (archived here) that took place six weeks ago in Beijing. Hosted by the non-profit organization “The Intellectual” 知识分子 — whose public WeChat account serves as a platform for discussions on scientific issues and their governance implications — the panelists delved into a wide range of topics, including:
the state of China’s AI industry, discussing the biggest bottlenecks, potential advantages in AI applications, and the role of the government in supporting domestic AI development;
the technical aspects of AI, such as whether Sora understands physics, the reliance on the Transformer architecture, and how far we are from true AGI;
and the societal implications — which jobs will be replaced by AI first, whether open- or closed-source is better for AI safety, and if AI developers should dedicate more resources to AI safety.
The panelists are all real heavyweights
They all attended the Second International Dialogue on AI Safety in Beijing (also in March 2024), where they engaged with prominent Western AI scientists and drafted a Consensus Statement on Red Lines in Artificial Intelligence.
Xue Lan 薛澜 is China’s ultimate tech-policy guru. With a background in engineering and public policy, he frequently advises the government and serves as a Counselor of the State Council. Xue is the director of Tsinghua University’s Institute for AI International Governance, Dean of Schwarzman College, and one of seven members on the National New Generation AI Governance Expert Committee.
Zhang Hongjiang 张宏江 is a founding chairman of the Beijing Academy of AI (BAAI), one of China’s most advanced AI research institutes. Trained in multimedia computing, he joined HP Labs in Silicon Valley in 1995. Four years later, he returned to China and participated in the establishment of Microsoft’s China business, where he later became a Chief Technology Officer (CTO). In the early 2010s, he founded his own company. In 2017, he joined Source Code Capital, a Beijing-based VC firm founded by a former vice president of Sequoia Capital China, which has made notable investments in Bytedance, Meituan 美团, and Li Auto 理想, among others.
Li Hang 李航 is head of research at Bytedance, where he leads both basic research and product development across fields like search, recommendation, chat, and more. He has decades of experience in China’s AI industry, having previously worked at Microsoft Research Asia and Huawei’s Noah’s Ark Lab. He studied in Japan in the 1990s and worked at NEC Corporation.
And Zhou Zhonghe 周忠和, who moderated the discussion, is Editor-in-Chief at The Intellectual. An evolutionary biologist by training, he is an academician at the Chinese Academy of Sciences, China’s preeminent research institution — which also functions as a consultancy for science and technology policy. Since March 2023, Zhou has been a member of the Chinese People’s Political Consultative Conference.
Highlights
Open source
Whether frontier models should be open-sourced has become a major point of disagreement in AI safety debates globally. These debates will have direct legislative influence. For example, a draft expert proposal for China’s AI law published in March 2024 would exempt some open-source models from many legal requirements. (Note that this is not an “official” draft law yet — it’s just an informal expert proposal.)
Xue Lan: Open source should be encouraged. Large companies can find profit models at the commercial application and product layer, and compete on that layer. Judging from the practice in various research fields, this model — open source in the basic research stage, closed source in the productization stage — seems to be a very effective approach in promoting human progress. …
The discussion between open and closed source actually involves weighing the pros and cons. Some people may worry that open source will give extremist groups or individuals the opportunity to misuse the technology. Regardless of open source or closed source, as long as you intend to do evil, you will always find a way. Various technologies we have today may cause damage to human society if abused. Biotechnology is a good example: it also has the risk of misuse. Therefore, what is more important is how to establish a system to prevent and stop any individual or organization from abusing technology to harm society.
Li Hang: I strongly agree with a view expressed by Harry Shum at an AI conference last year: whether or not to open source depends on a company’s business strategy and position. … The best company will definitely not open source. The second-best will not open source if it wants to compete with the first place. The third and fourth might choose to open source to gain some competitive advantage. … Among AI companies, OpenAI and Anthropic are currently not open source. Meta and Amazon are open source.
Zhang Hongjiang: In both open and closed source, security and safety issues are always inevitable. But with open-source models, it’s easier for us to verify and review. In the future, any AI model released should pass safety certifications.
AI safety funding
Zhang Hongjiang: At a closed-door AI meeting, I once heard a point of view that really surprised me, but I believe that the data is correct: 95% of R&D expenses for nuclear powerplant equipment go into safety. This is a revelation for the AI field. Should we also invest more resources in AI safety? If 95% of nuclear-power R&D is invested in safety, shouldn’t AI also invest 10% or 15%, because this technology may also lead to human extinction?
The future of AI: large vs. small models
Will general large models make smaller models for specific use cases redundant? Li Hang and Zhang Hongjiang had a long back-and-forth, some short extracts from which are quoted below. This kind of discussion can have a profound impact on whether Chinese developers want to compete with OpenAI and Anthropic for leadership in frontier models, or rather focus on applications, such as industrial AI.
Li Hang: Certainly, China still has many opportunities in the application and commercialization of AI. We can see that in the Internet and mobile Internet era: Chinese companies have performed very well, especially in the past ten years. …
Zhang Hongjiang: I have a different view. This time may be really different from before. The current goal in AI is AGI, which emphasizes generality or universality. … Each time OpenAI releases a new version or adds a new feature, some companies go bankrupt. …
We must not have the mentality that, just because we succeeded before and have caught up in terms of applications, the same must happen with new technologies. This time may be really different. …
Li Hang: I understand your point of view. But I think in practical applications, large models are definitely not always able to completely replace what small models do on a specific vertical class; maybe GPT-4 scores 90 points, and a small model also scores 90 points. In certain areas, customized small models may have an advantage because they would be more cost-effective and commercially viable.
Zhang Hongjiang: I disagree. I can give you an example right away. After ChatGPT came out, all kinds of translation software died.
Talent, data, compute — what’s the bottleneck?
Due to US export controls, many consider compute the key constraint on China’s AI development. The panelists in this debate, however, identified talent as the more important concern.
An audience member asked whether China could centrally integrate existing domestic computing resources to support major AI breakthroughs.
Xue Lan: I think this is difficult to realize. If we go back to a few decades ago, under the nationwide system and planned economy, it might still be possible. I think today it would be very difficult.
That’s a sober statement, given that the new National Data Administration is actually working on the “national integrated computing network.” It is supposed to work akin to a power grid: users anywhere can access computing resources. Xue seems to be skeptical of the promise.
But overall, the panelists did not believe compute to be the major bottleneck:
Li Hang: Although computing power currently encounters a bottleneck, this is a relatively short-term problem.
Neither did they see data as a large obstacle:
Li Hang: There are currently relatively many high-quality English data resources, while there are relatively few Chinese, especially on the Internet. But through machine translation or some other future technological tool, we can achieve interoperability between the data in the two languages. So I don’t think the gap in language data is going to be a particularly big obstacle.
Zhang Hongjiang: We have many ways to make up for data.
Instead, they focused on talent and ecosystem development:
Li Hang: In the long run, talent cultivation is the most critical. … I think undergraduate education is very important. In the United States, undergraduate students in machine learning at top universities have very difficult assignments and even have to stay up late to complete them. US undergraduate education has done a good job of cultivating some basic skills in the computer field, but domestic education needs to be strengthened in this regard.
It is also important to integrate university research with industry. … Short-term problems, such as data problems, are relatively easy to solve — but talent cultivation requires the joint efforts of the entire society.
Zhang Hongjiang: Stanford University annually publishes an AI Index report. This report counts the publication of AI papers globally. Guess where MIT ranks among the top-ten institutions?
Zhou Zhonghe: Isn’t it ranked first?
Zhang Hongjiang: If MIT ranked first, I would not ask the question. In fact, MIT ranks tenth. The top nine are all Chinese institutions. This shows that we must have a lot of talent in the industry. We simply need to turn the quantity of published articles into quality, move from follower status to breakthroughs and leadership.
I think it’s important to develop children’s thinking skills, not just specific knowledge. American schools offer logic and critical thinking courses to fourteen-year-old students. This course teaches children how to think, rather than a specific professional knowledge. From any professional perspective, logic and critical thinking skills are very important if you want to engage in research.
Xue Lan: Another very important point is how to create an ecosystem that allows Chinese companies and research institutions, including universities and others, to organically integrate with one another. This has always been a problem that China needs to solve. This is especially important for China’s AI field. As Mr. Zhang said just now, we publish a lot of papers and have a lot of patents. What we lack most is an ecosystem.
How a weak toB market hurts China
Zhang Hongjiang: Regarding going overseas, what everyone is paying attention to is when consumer-oriented companies (to C) are expanding overseas. But in fact, enterprise-oriented companies (to B) are going overseas more frequently, and this is actually very reasonable. …
In the past twenty-five years since I returned to China, I have always been very envious of the overseas toB software market. … If we compare the software market and Internet application market in China and the United States — especially in the fields of software tools and cloud services (SaaS) — we will find that the US enterprise market (to B) is far more mature than China. …
The core of toB software is to provide enterprises with tools to improve efficiency, such as Office, SAP, Feishu 飞书, etc., which are all tools to improve productivity. When companies use these tools extensively, it means they are pursuing efficiency improvements. Many of our companies have not yet made full use of these productivity tools, which shows that our production efficiency is not high enough and that we still rely on manpower to complete tasks.
Sora, scaling laws, and AGI
Xue Lan: Sora … does not just show an understanding of the text, but also of the operating laws of the physical world. …
Zhang Hongjiang: Sora has really shocked me. Models like GPT 3.5, GPT 4.0, and Claude are mainly good at understanding language. But Sora has demonstrated the ability to understand not only language, but also the physical world.
Zhou Zhonghe: Just now, Teacher Zhang expressed the belief that the videos generated by Sora are in line with the laws of physics, and that Sora can understand common sense. But there are also different voices, such as Turing Award winner Yann LeCun, who say that training with large amounts of text data alone cannot reach the level of human intelligence. … You didn’t seem to agree with this view just now?
Zhang Hongjiang: I may offend someone. I think those who say this route will hit limits soon are actually not real practitioners and have not participated in it. If you are doing this, you will believe in Scaling Law, just like those of us on the front line.
If you are interested in this, you can check out the recent interview by Yang Zhilin 杨植麟, an assistant professor at Tsinghua University and CEO of Moonshot AI. He talked about Scaling Law. I think his views on Transformers and the future are very insightful. We are still far away from the limit of Transformers. The current problem may be that there is not enough data. We should find ways to continue to expand the data scale. …
Li Hang: I partially agree with Teacher Zhang. I believe in the potential of Scaling Law and think we haven’t seen the limits yet. My views, however, are not entirely the same. I had a face-to-face chat with Mr. Yann LeCun during a meeting in Hong Kong last year. He felt that the current large models lacked world models. What I understand is that, if we can combine visual, language, and other multi-modal information, we will be closer to the world model. As I just said, when video is generated, semantic information and 3D information can also be generated, which is closer to humans. I think there is still a lot of room in this area, and we will go down the road of Scaling Law. …
I think the existing large models are in the right direction and are an important way to AGI. But there may also be some aspects of human intelligence that cannot be realized with large models. … I just hope that in the future, we can better imitate the human brain and achieve AGI that is closer to humans.
Which jobs will be replaced by AI, and how quickly?
Xue Lan: It depends on two aspects. The first is how quickly AI itself improves. The second is what changes in the social systems, which provide very strong protection to traditional industries.
Li Hang: All industries, including programming and AI, will experience bifurcation, with the best and most creative talent playing a more important role.
Zhang Hongjiang: Workers of average skill level may be replaced, but the top does not need to worry because they are in the minority. As for assistant jobs, such as legal assistants or analysts, they will be replaced to a large extent.
What about blue-collar jobs?
Zhang Hongjiang: Why I am so excited about Sora lies in its preliminary understanding of the physical world. When a system can understand the physical world, it can direct robots to perform tasks. The progress of AI in hardware may be slower than that in software, but with the advancement of robotics technology, such as the improvement of dexterous hands and mechanical abilities, it is only a matter of time before AI will have an impact on blue-collar jobs.
Full Translation
Note: this is a machine-assisted translation that was manually checked for obvious errors only. Some mistakes and inaccuracies likely remain. Read the original Chinese here (archived here).
Does Sora begin to understand the laws of physics and common sense?
Zhou Zhonge: Our first question today is about Sora. What are the aspects of Sora launched by Open AI that make you really excited, or what is the most noteworthy aspect of Sora? What impressed you the most about Sora?
Xue Lan: I was shocked when I saw Sora. Because it is not just doing textual communication, but also has a certain degree of imagination. Sora can generate a series of coherent dynamic pictures based on a short textual description. This does not just show an understanding of the text, but also of the operating laws of the physical world.
When we discussed AI in the past, we always thought that imagination was unique to humans. But now, Sora shows that AI can also have such abilities. This may be what sets Sora apart.
Zhang Hongjiang: First, I want to talk about how Sora shocked me. The release of Sora was actually a demonstration, they released demonstration videos but did not release the model itself. However, from the 40 demonstration videos, we can see the huge progress of AI technology, which is very exciting.
I noticed several highlights: first, it generates high-resolution video, which has not been achieved in past video generation; Second, the length of the video reaches 60 seconds. In the past, Runway, which had a better performance in generating videos, could only do it for a few seconds after two years of work. Third, and the most shocking thing is that in one of the scenes, a car Off-road jeeps are racing on rugged mountain roads. In the past, this kind of shot required a device to follow behind because the road was bumpy and difficult, but the resulting video was very realistic. The logic of vehicle driving is also excellent. It always drives on the right and turns very naturally.
Zhou Zhonghe: I have a layman’s question. When we ordinary people look at these videos, they may think that they are just high-quality images. Aren’t its imaginations and logic also instilled by humans?
Zhang Hongjiang: No, humans did not explicitly tell it to do this.
In the traditional computer graphics paradigm, creating videos usually requires building detailed physical models. For example, if we want to make a video of a vehicle driving, we need to first create a 3D vehicle model and a scene model containing roads and other environmental elements. The scene environment model is like a virtual world, and the vehicle model is an object moving within this world. These models need to precisely define dynamic behaviors and interactions with the environment, explicitly encoding various physical laws and motion rules to ensure that the vehicle’s performance when turning or driving conforms to real-world situations.
When using a large model like Sora, we do not directly tell the AI-specific laws of physics. On the contrary, the AI learned it on its own by analyzing a large amount of video data. We didn’t tell it that in most places in the world, cars drive on the right. It was also not told that if it did not follow the route and turn, the car would hit the hill. We didn’t tell it any of these clear rules.
Sora has really shocked me. Models like GPT 3.5, GPT 4.0, and Claude are mainly good at understanding language. But Sora has demonstrated the ability to understand not only language, but also the physical world.
Zhou Zhonghe: You say that it has understood the laws of physics and common sense, but I always feel like it is just imitating.
Zhang Hongjiang: It’s imitation, but remember what [Richard] Feynman once famously said: “I have no way of understanding things I can’t create.” Now that we can create it, doesn’t it mean that we understand it?
Xue Lan: Speaking of intelligence, we must first define human intelligence. We have cognitive abilities, reasoning abilities, creativity, and probably other aspects like emotional intelligence. If we look at these, in some aspects, such as cognitive and reasoning capabilities, AI may have surpassed ordinary people, and may even surpass the most amazing people. But in other dimensions, humans will perhaps still maintain an advantage.
You think that AI is just imitating. But isn’t the process by which humans acquire these abilities essentially the same? Both are constantly receiving external information and gradually forming cognitive and thinking patterns. There is no fundamental difference in methods between the two.
Zhou Zhonghe: Li Hang, do you feel the same way?
Li Hang: In some respects yes, in some no. The “PixelDance” system which we developed ourselves surpassed “Runway” and the other best systems at the time. But when Sora was released in February, PixelDance was surpassed by Sora. My first feeling was that we needed to do things quickly because the competition is very fierce now.
In addition, judging from Sora’s technical reports and related papers, although it has some degree of technical innovation, it is not revolutionary. The main innovation is to change the Unet architecture of the diffusion model to the Transformer architecture, which allows the model to process more data and better learn physical phenomena. From a scientific perspective, I think large AI models currently have a common feature, that is, they are all based on the Transformer architecture, and the most basic things are the same.
Zhou Zhonghe: So there haven’t been any big breakthroughs in basic scientific principles in recent years?
Li Hang: Yes. There has been some progress, but ever since the Transformer architecture emerged in 2017, we have seen constant convergence/dwindling. Transformer was originally developed for natural language processing, but now, even computer vision is starting to move toward using the same architecture.
Zhou Zhonghe: So, do you have the same feeling as Mr. Zhang, and think that the AI model is now beginning to understand common sense and physical laws?
Li Hang: Yes, I agree with Mr. Zhang. But I think in the future there is still a lot of room. Sora is not 3D yet. 3D generative technology would allow us to see what objects look like from different angles. For example, when we see a person from the front, we have the ability to imagine what the back of his head looks like. 3D generation can help us see the back of a person’s head.
Currently, 3D generative technology is still in its infancy. From the papers presented at last year’s International Conference on Computer Vision, one of the top conferences in the field, you can see the results of current technology. They are actually all relatively simple. For example, looking at an object, a table or a chair, from another angle.
Furthermore, current generative models do not have object semantics. When we watch a video, we can recognize that this is a car and that is a road, but in Sora’s model, it does not understand these accurately, it just processes pixels and data. In the future, if we can further develop this technology so that it can not only process 3D spatial relationships, but also understand the semantic information of objects, then we will be closer to creating human-like intelligence. In this regard, we still have a lot of work to do.
Is the Transformer architecture the best path? Are we too path-dependent?
Zhang Hongjiang: The Transformer architecture mentioned by Teacher Li just now is indeed very critical. Since 2017, the path of Transformer has pointed out a way for those of us who make AI or make large models. In the past seven years, the industry has recognized it as a straight path to AGI for the development of large models.
Technological development is path dependent. This means that once we determine the right direction, all attention and resources will be focused. Back seven years ago, Google researchers invented the Transformer architecture, but it was OpenAI that really recognized its potential and invested fully in it. Although Google has also launched many models, none has been able to surpass the achievements of GPT3.5 for a long time.
Nowadays, everyone who makes large models focuses on the Transformer architecture, including Sora. In the past, video generation could not be achieved using the Unet framework, but switching to the Transformer architecture achieved a breakthrough. In the future, whether it is GPT4.5, GPT5, Claude, or Gemini, the path of Transformer will continue.
Xue Lan: Now that we’ve mentioned this, let me ask a question. Will we ignore other potentially better paths because of path dependence? The Transformer architecture was really important in 2017, but there were other paths at the time. Are we sure now that this is the best path? Is it possible that we are missing other potentially excellent paths?
Zhang Hongjiang: Your question just gave me an opportunity. What I want to say is that this is like the invention of electricity. When alternating current became mainstream, direct current was eventually only used for batteries. This path dependency is very critical.
Li Hang: What I would like to add is that as scientists, we are always looking for better and disruptive technologies to change existing Transformers. But everyone has made a lot of effort, and I also know some research. Although some models look good on a small scale, none of them can surpass Transformer once scaled up. The current conclusion is that Transformer is indeed very powerful, other models, but so far, have not been able to successfully subvert Transformer.
Zhang Hongjiang: This brings up another topic, which is the Scaling Law of large models. The Transformer architecture has become mainstream, and we have verified its power. The next step is to make it bigger and bigger and feed it more and more data. Most practitioners in this field believe that the potential of this architecture has not yet reached its limit.
Therefore, you can see why everyone is now eager to invest in chips and increase computing power. When we are pursuing computing power, we are actually competing for the resources of the data center (IDC). In the process of seizing IDC, we had to fight for power supply. This series of actions is actually because everyone agrees with one point of view: Scaling Law - it is believed that as the scale of the model increases, the performance and capabilities will also improve accordingly.
Zhou Zhonghe: Just now, Teacher Zhang expressed the belief that the videos generated by Sora are in line with the laws of physics, and that Sora can understand common sense. But there are also different voices, such as Turing Award winner Yann LeCun, who say that training with large amounts of text data alone cannot reach the level of human intelligence. Others feel that this is a dead end and that it is reaching its limit. You didn’t seem to agree with this view just now?
Zhang Hongjiang: I may offend someone. I think those who say this route will hit limits soon are actually not real practitioners and have not participated in it. If you are doing this, you will believe in Scaling Law, just like those of us on the front line.
If you are interested in this, you can check out the recent interview by Yang Zhilin 杨植麟, an assistant professor at Tsinghua University and CEO of Moonshot AI. He talked about Scaling Law. I think his views on Transformers and the future are very insightful. We are still far away from the limit of Transformers. The current problem may be that there is not enough data. We should find ways to continue to expand the data scale.
Zhou Zhonghe: Others mentioned electricity and other resources.
Zhang Hongjiang: Scaling Law will bring about many related things. For example, in the chip industry, Nvidia’s stock has skyrocketed, and so has IDC. Data has become very valuable now. These are ecosystem factors that drive the development of the entire industry. Therefore, I think the limit is still far away. I respect scholars like Yang Likun very much, but I don’t quite agree with him on this point of view. He believes that data-driven systems do not have real reasoning and learning capabilities and cannot reach AGI. That’s his definition, but by my definition, we’re moving in that direction.
Zhou Zhonghe: Teacher Li Hang, do you have the same views as Teacher Zhang?
Li Hang: I partially agree with Teacher Zhang. I believe in the potential of Scaling Law and think we haven’t seen the limits yet. My views, however, are not entirely the same. I had a face-to-face chat with Mr. Yann LeCun during a meeting in Hong Kong last year. He felt that the current large models lacked world models. What I understand is that, if we can combine visual, language, and other multi-modal information, we will be closer to the world model. As I just said, when video is generated, semantic information and 3D information can also be generated, which is closer to humans. I think there is still a lot of room in this area, and we will go down the road of Scaling Law.
Job displacement
Zhou Zhonghe: As AI is developing so rapidly, its impact on our lives and work is of great concern to everyone. For instance in autonomous driving or the Internet industry, but there are even people who say that the impact on white-collar jobs will be the greatest. Mr. Xue, which field do you think will be hardest hit in the short term, white-collar workers or blue-collar workers?
Xue Lan: I think all industries involved in the acquisition, processing, and dissemination of data or information will be hugely impacted. This includes not only the media industry, such as film and television, but also education, art, and even medical services. It also includes clerical/secretarial work and assistant jobs in the legal profession. All fields related to information processing — including information acquisition, processing, and dissemination — may gradually be impacted.
Zhou Zhonghe: How quickly will these changes happen? Five years or ten years?
Xue Lan: It depends on two aspects. The first is how quickly AI itself improves. The second is that changes in the social systems, which provide very strong protection to traditional industries. Therefore, it is a process of technological progress on the one hand, and institutional changes on the other. These two aspects need to be coordinated to move forward.
Zhou Zhonghe: You mentioned that education will also be affected. If certain industries are expected to cease to exist in the next five to ten years, why do we still study them in universities now? This is an urgent issue because change is coming so fast and perhaps we haven’t discussed it enough.
Li Hang: It is hard to predict, but there are two obvious trends. First, new jobs will emerge, such as data labeling. Data is very important. The mathematician Terence Tao 陶哲轩 and others are beginning to label data related to proving mathematical theorems, and use the new programming language Lean to describe this process. He uses his own experience to write out the proof process, and then lets the AI learn it. I predict that AI will be able to do proofs of mathematical theorems in the future.
Secondly, all industries, including programming and AI, will experience bifurcation, with the best and most creative talent playing a more important role.
Zhou Zhonghe: So for ordinary people, which industries will see the greatest impact?
Li Hang: Changes in the hardware field may be slower because hardware iteration takes time. As for software, development will be faster. But I think in the long term, the shape of software development will also change. Some simple procedural AI can be written.
Zhou Zhonghe: Can machines completely replace human creativity?
Li Hang: As Mr. Xue said just now, I think there are three aspects in which artificial intelligence is currently difficult to replace humans: emotion, creativity, and free will. Emotions are human instincts. To realize emotions on a machine is equivalent to creating a living person. The same goes for creativity, and free will. These three are not pure intelligence, but life phenomena.
If something is a task and it can be evaluated on how well it is completed, AI can basically complete it. John von Neumann once said this, which means tell me a task, no matter how complex it is, define it clearly, and I can build you a machine specifically to complete your task.
Nowadays, everyone often talks about AGI and general artificial intelligence. “General purpose” is indeed very revolutionary. This tool is not specialized in doing a certain thing like von Neumann said. It becomes very versatile and can accomplish many, many tasks. But on the other hand, as long as these tasks can be defined, data can be marked, and evaluation criteria can be established, it seems that AI can basically do it. As mentioned earlier, several characteristics of human beings are life phenomena, not tasks, and AI cannot handle them. But AI can do many tasks better than humans. It is possible that AI will surpass humans in most jobs in the future.
Zhang Hongjiang: I agree with Teacher Xue’s point of view. Human intelligence has multiple dimensions, and machines may surpass humans in many dimensions. For ordinary people, what they are most worried about may be their jobs.
I do think white-collar jobs may be hit faster. I was talking about AI with Cui Jian 崔健 last week and mentioned that in the future, 97% of people may not work and only 3% will have a job. This is not my opinion, but that of the author of “A Brief History of the World.” Workers of average skill level may be replaced, but the top does not need to worry because they are in the minority.
As for assistant jobs, such as legal assistants or analysts, they will be replaced to a large extent.
Now some AI tools can help us work more efficiently. For example, many people have to manually arrange meetings through Microsoft Outlook. In the future, AI may do this work, and the efficiency will be much improved.
In addition, we cannot stay at the stage of improving efficiency. 30 years ago, chess master Kasparov was defeated by Deep Blue. Kasparov said at the time that machines will be of great help to us in the field of chess in the future. However, today, 30 years later, what we see is that machines no longer need human help at chess.
Zhou Zhonghe: So, what about blue-collar jobs?
Zhang Hongjiang: Why I am so excited about Sora lies in its preliminary understanding of the physical world. When a system can understand the physical world, it can direct robots to perform tasks. The progress of AI in hardware may be slower than that in software, but with the advancement of robotics technology, such as the improvement of dexterous hands and mechanical abilities, it is only a matter of time before AI will have an impact on blue-collar jobs.-
Zhou Zhonghe: What impact does artificial intelligence have on programmers?
Zhang Hongjiang: The work of software designers may be replaced by automated tools, such as Microsoft’s GitHub Copilot and Google’s corresponding products. These tools can already do a lot of things. Driven by large models, they can complete the writing of many commonly used programs, which can at least improve the efficiency of developers. Other repetitive tasks can also be easily replaced.
Zhou Zhonghe: Will the industry that studies artificial intelligence still need so many people in the future?
Li Hang: More people are needed, but it will also be divided into two levels. For example, in the work of data annotation, there is a huge difference from simple common sense annotation to the professional annotation I just mentioned.
Zhou Zhonghe: Mr. Xue, for our country’s natural science researchers, especially those who are in a leading position in the fields of technology development and basic research, how will the development of artificial intelligence affect their work?
Xue Lan: I think there will definitely be a significant impact, and scientific research will also be polarized. But we need to distinguish between normal scientific research and scientific revolution. If we look at Thomas Kuhn’s scientific revolution theory, scientific research can be divided into normal scientific research and scientific revolution. In normal scientific research, we have a clear understanding of the basic paradigm of a certain field and are solving some unresolved problems, which is like solving a puzzle. For example, current research on protein structure is one such area.
In these fields, artificial intelligence may replace many jobs. The jobs of scientific researchers working under the existing paradigm may be replaced by artificial intelligence. But for those researchers who can discover new problems, create crises, and challenge existing paradigms, their work is difficult to replace with artificial intelligence. Such researchers will always be in demand.
How far are we from true AGI?
Zhou Zhonghe: A very hot topic right now is how far we are from true general artificial intelligence (AGI), and what the definition or measure of reaching AGI is. In addition, the AI revolution led by OpenAI is widely recognized, but what about other competitors? For instance, some people say that Claude surpassed ChatGPT, but is this statement an exaggeration?
Li Hang: There is currently no strict definition of what AGI is.
Generality is a distinguishing feature of large models. They can do many different things, and have large room for development. However, it is still different from humans in terms of creativity, emotion, and free will. The performance of the model in these aspects can only approximate human intelligence, rather than fully realizing human intelligence.
AGI, as Jensen Huang said, may be realized within five years, but it is also based on a specific definition. I think that once we can clearly define a task and evaluate how well it is accomplished, AI usually does a good job. However, humans can understand the world and can daydream and imagine. If we want to catch up with humans in these aspects, I think artificial intelligence will not be able to do it within ten or twenty years.
Zhang Hongjiang: I would like to argue on the three points mentioned by Teacher Li Hang. I agree with two of them — but on the emotion side, I think emotion can be thought of as a rewarding function for humans. If we can learn the human reward function and provide enough data, then artificial intelligence may have the potential to simulate emotions.
This is a high-dimensional complex function that can be computationally complex, which is where large models have the advantage. For example, in the AI for Science project created by Teacher E Weinan 鄂维南, in material design, simple material structures can be calculated through differential equations, but complex molecular materials are almost impossible to calculate. It is better to use large models for simulation. In areas such as materials design, weather prediction, and life sciences, when problems are too complex to be described by mathematical equations, that’s exactly where large models come into play.
Zhou Zhonghe: Some people say that the large model is just an isolated puzzle piece of AGI, and there are still many puzzle pieces that have not been found. Do any teachers agree with this view?
Zhang Hongjiang: This must be what a philosopher said.
Zhou Zhonghe: This is something you don’t agree with, right? Teacher Li, do you think the current direction of large models is a sufficient condition for leading to AGI?
Li Hang: I think the existing large models are in the right direction and are an important way to AGI. But there may also be some aspects of human intelligence that cannot be realized with large models.
Zhou Zhonghe: Mr. Li, you once mentioned that deep learning will be mainstream for some time in the future, but in the long run, we should still get inspiration from human brain computing.
Li Hang: My point of view does not conflict with the current large model route. I just hope that in the future, we can better imitate the human brain and achieve AGI that is closer to humans.
There are three levels of computing: functional, algorithmic, and physical implementation. The existing large models based on Transformer mainly imitate humans at the functional level, but at the algorithm level, the structures of AI and the human brain are still completely different, so we can learn more from the human brain at the functional level.
Xue Lan: I am going against the grain of both of them. Path dependence does exist, but we may have to go back to the path we missed many years later. For example, in the past, our technology and funds were invested in fuel-powered vehicles. In fact, electric vehicles were discovered at the beginning of the last century. There was a time when they could still compete with gasoline vehicles, but they were later ignored due to various reasons. Now, many years later, electric vehicles are developing very well again, so it is not ruled out that the missed roads will have good development in the future.
Zhou Zhonghe: Mr. Xue, considering the current development trend of artificial intelligence, what type of talent do you think we need most? Are we more inclined to technical talent majoring in computer science, or do we need more interdisciplinary experts, such as those in brain science, life science, and other fields?
Xue Lan: I think the talent we need must first be creative, and creativity will still be crucial in the future. In addition, the traditional education model that emphasizes professional segmentation may need to change. We need to rethink education from a new dimension.
The knowledge and skills we have learned in the past can be divided into two categories: one is necessary to survive in society, and the other is learning that improves our cognitive abilities. I estimate that with the development of artificial intelligence, the first type of abilities may be replaced by systems, and there will be less need to learn these things.
Instead, what we need to think about is how to improve the cognitive abilities and creativity of educated people through education. A completely new approach to education may be needed, completely different from the existing education system. In fact, artificial intelligence should have the greatest impact on the education system, but we are the slowest to move.
Zhou Zhonghe: Both Mr. Li and Mr. Zhang work in a company. What kind of talent do you hope to recruit? What kind of talent do you expect the future education system to produce for you?
Li Hang: In the field of artificial intelligence, I think undergraduate education is very important. In the United States, undergraduate students in machine learning at top universities have very difficult assignments and even have to stay up late to complete them. US undergraduate education has done a good job of cultivating some basic skills in the computer field, but domestic education needs to be strengthened in this regard. From the perspective of pure computer and artificial intelligence talent training, I see and feel that there is a certain gap.
Zhou Zhonghe: After artificial intelligence became popular, many people suggested starting to learn artificial intelligence in middle school.
Zhang Hongjiang: As a father of two children, I may know a little more about education. I think it’s important to develop children’s thinking skills, not just specific knowledge.
American schools offer logic and critical thinking courses to fourteen-year-old students. This course teaches children how to think, rather than a specific professional knowledge. From any professional perspective, logic and critical thinking skills are very important if you want to engage in research.
Future talent most need the ability to think logically and critically. Education such as instilling knowledge actually cultivates employability, and it is the employability of the past and will not work in the future. From the perspective of scientific research talent. Looking back, before the professionalization of science and hundreds of years after the Renaissance, there were no professional scientists. What was needed was their ability to think and observe.
Xue Lan: I think the more we are in this situation, humanistic qualities will become very important.
Open source or closed source — which model is safer for humanity?
Zhoug Zhonghe: Recently, news emerged of Elon Musk’s plans to sue OpenAI because it allegedly violates its non-profit purpose. We know that the non-profit model has its advantages, but a lack of funding can also affect the development of technology.
Xue Lan: The development of modern science and technology hopes to advance on the road of open science, especially in the field of artificial intelligence. Open source can promote the exchange and development of technology. Those who oppose open source worry about security risks, but some people think that closed source may actually be more dangerous. These debates have always existed. I think we can ask two other experts in the field of artificial intelligence to explain how we should judge this issue?
Li Hang: I strongly agree with a view expressed by Harry Shum (沈向洋) at an AI conference last year: whether or not to open source depends on a company’s business strategy and position; industry leaders may choose not to open source. The best company will definitely not open source. The second-best will not open source if it wants to compete with the first place. The third and fourth might choose to open source to gain some competitive advantage.
I think this view makes sense, at least from a historical perspective. Among AI companies, OpenAI and Anthropic are currently not open source. Meta and Amazon are open source.
Zhou Zhonghe: Apart from the commercial aspects, we may be even more concerned with the impact of open sourcing on technological development and security issues. What do you think about these aspects?
Li Hang: Actually, from a business perspective, Meta and other companies choose open source not because of such other considerations, but because doing so can bring certain commercial benefits. Harry Shum believes these companies open source to gain commercial advantages, and I agree. This involves a more fundamental question: whether the research and development of AI should be oriented towards the market economy.
Zhou Zhonghe: Mr. Xue, do you agree that AI companies should be profit-oriented?
Xue Lan: Our current research on artificial intelligence is actually equivalent to basic research in science and technology. In science and technology, basic research is often open source. However, the reward mechanism for AI research is slightly different. In academia, being the first to publish a paper is recognized as having priority. This is the so-called reward system, which has already been generally acknowledged by everyone. Among large AI companies, however, the real competition lies in the final application and products.
From this perspective, Open source should be encouraged. Large companies can find profit models at the commercial application and product layer, and compete on that layer. Judging from the practice in various research fields, this model — open source in the basic research stage, closed source in the productization stage — seems to be a very effective approach in promoting human progress.
Zhang Hongjiang: I very much agree with teacher Xue’s point of view. Open source is an important tool in the exploration phase. It encourages everyone to discuss, communicate, and evaluate together. Only like that can we truly promote progress in a field. From this point of view, I strongly agree with open source. Open Source can bring together people who really do the research. In today’s AI field, there are still many problems that need to be solved in terms of methodology and architecture. Open source is a very good communication carrier.
Zhou Zhonghe: You just mentioned safety/security issues. So how should we solve safety/security risks that open source may bring?
Zhang Hongjiang: In both open and closed source, security and safety issues are always inevitable. But with open-source models, it’s easier for us to verify and review. In the future, any AI model released should pass safety certifications. Furthermore, if AI is a technology that can change humanity, we need to invest more resources in AI safety research.
At a closed-door AI meeting, I once heard a point of view that really surprised me, but I believe that the data is correct: 95% of R&D expenses for nuclear powerplant equipment go into safety. This is a revelation for the AI field. Should we also invest more resources in AI safety? If 95% of nuclear-power R&D is invested in safety, shouldn’t AI also invest 10% or 15%, because this technology may also lead to human extinction?
Zhou Zhonghe: Mr. Xue, you often actively discuss the issue of AI governance in China and on the international stage. Safety/security is obviously a big concern. Given that AI has already reached a relatively high level of intelligence, can we consider open source some kind of safeguard?
Xue Lan: Yes, the discussion between open and closed source actually involves weighing the pros and cons. Some people may worry that open source will give extremist groups or individuals the opportunity to misuse the technology. Regardless of open source or closed source, as long as you intend to do evil, you will always find a way. Various technologies we have today may cause damage to human society if abused. Biotechnology is a good example: it also has the risk of misuse. Therefore, what is more important is how to establish a system to prevent and stop any individual or organization from abusing technology to harm society. We need to think more about how to control these risks through regulatory mechanisms.
Suppose from a business perspective of an enterprise, strategies like closed source need to be adopted. In this situation, we need to establish a regulatory mechanism. This regulatory system will regulate and constrain the company’s closed-source research to ensure its legal compliance. Therefore, I believe that the needs of enterprises and the regulatory system are mutually reinforcing and complementary.
What is China’s standing in the current wave of artificial intelligence?
Zhou Zhonghe: China has published a large number of academic articles about AI and is also in a relatively leading position internationally. But to get more specific, what are China’s advantages and disadvantages in AI? For example, some people believe that China lacks high-quality and large-scale Chinese-language data sets. Does this perhaps impact our AI development?
Li Hang: Certainly, China still has many opportunities in the application and commercialization of AI. We can see that in the Internet and mobile Internet era: Chinese companies have performed very well, especially in the past ten years. We have very distinctive features in the application layer of commercialization of the Internet. For example, in many practical use-cases of large models, it is not necessary to pursue large-scale models like GPT-4. Even relatively small models can plan an important role in specific practical scenarios.
From my personal point of view, the future development opportunities for AI are concentrated in four major areas. First, there is the field of white-collar work assistants in a broad sense that Mr. Zhang mentioned before; there is a lot of application space here. Second, the development of robots is also very critical, they can help workers and ordinary people complete generic work tasks. Third, AI for science — e.g. the use of AI to promote scientific research — has potential in mathematics, physics, chemistry, and other fields. Fourth, in the entertainment industry, whether it is video production, movies, games or virtual reality, there are many opportunities for the development of AI.
As for the data issue, there are currently relatively many high-quality English data resources, while there are relatively few Chinese, especially on the Internet. But through machine translation or some other future technological tool, we can achieve interoperability between the data in the two languages. So I don’t think the gap in language data is going to be a particularly big obstacle.
However, if we compare the development of AI to an arms race, then we can say that OpenAI is really a step ahead in AGI, and that everyone else is trying to catch up. As for China, both the business community and academia need to increase our efforts in all aspects to narrow the gap.
As we just discussed, there are many start-ups in China that are doing well. One could say that Chinese companies have mastered the scientific principles of large AI models, because these scientific principles are basically public — everyone can learn them. But now the question is more about engineering and product development. If these can be called technology, we have actually not yet fully mastered these technologies. No matter whether Chinese companies or research institutes, they still need to work hard to master these technologies as soon as possible before they can produce things of the same level as GPT-4 or Sora.
Zhang Hongjiang: I would like to add something. The current wave of deep-learning-driven AI is not at all a new phenomenon. Today’s situation only strengthens this kind of view. This wave is a combination of algorithms, computing power and data. To a large extent, algorithms reflect the actual strength of talent, or what we just call talent. So we could also say it is a combination of talent, hardware, and data. Through these three points, we can clearly understand where we stand.
We have many ways to make up for data. The next step is just talent. Stanford University annually publishes an AI Index report. This report counts the publication of AI papers globally. Guess where MIT ranks among the top-ten institutions?
Zhou Zhonghe: Isn’t it ranked first?
Zhang Hongjiang: If MIT ranked first, I would not ask the question. In fact, MIT ranks tenth. The top nine are all Chinese institutions. This shows that we must have a lot of talent in the industry. We simply need to turn the quantity of published articles into quality, move from follower status to breakthroughs and leadership. There are many more trains of thought we can continue to think along these three points.
Regarding what Li Hang just said, that China has advantages in application, I have a different view. This time may be really different from before. The current goal in AI is AGI, which emphasizes generality. It used to be about single-point capabilities, but now it is about general intelligence.
This is quite different from before. Each time OpenAI releases a new version or adds a new feature, some companies go bankrupt. Things you used to think only a company can do can now be covered by adding a feature to a large model. After Sora came out, Runway and other video generation companies became very nervous because they are just one tool, whereas Sora is part of a large multi-modal model.
If large models continue to develop in this direction, their functionality will become increasingly comprehensive. When a few words can generate a new application in the GPT store, when the functions of large models are all-encompassing, isn’t that similar to the fear that Microsoft’s operating system brought to other software companies in the 1990s? And you cannot even use antitrust methods to fight against it, because the large model itself is so strong.
If you hope to rely on a 70-point dedicated smaller model with slightly lower performance in specific use cases, you will always be afraid that someone will make a 90-point large model that will just smoothly cover all of these functions and sweep away these small models.
So why is it that even though more than a year has passed since the advent of ChatGPT, we have not seen many emerging companies in the field of so-called native AI applications? On the one hand, one has figured it out, and on the other hand, the performance of large models is not yet good enough to satisfy users immediately no matter in which field it is used.
Because users have very high expectations for applications. For example, when speech recognition technology was immature thirty years ago, Apple made a handwriting and speech recognition product called Newton. But the problem was that at the time speech recognition technology was not up to standard, and even handwriting recognition was not up to standard yet because neural networks only had three layers. In the end, users did not approve of it and the market for this product plummeted. When a technology can not support a product, or can only support it 60%, but the user expectation is 90%, the product is basically equivalent to 0%.
We must not have the mentality that, just because we succeeded before and have caught up in terms of applications, the same must happen with new technologies. This time may be really different — that’s what I want to say.
Li Hang: I would like to add one point. I mostly agree with Mr. Zhang’s views, but let’s look at the search engines we are discussing. There are giants like Google, but of course there are also other search engines. Even if Google dominates the market, other companies still have opportunities.
Zhang Hongjiang: But Google occupies 90% of market share, while Microsoft has only gained 3% of market share after more than twenty years of hard work — that’s the difference.
Li Hang: I understand your point of view. But I think in practical applications, large models are definitely not always able to completely replace what small models do on a specific vertical class; maybe GPT-4 scores 90 points, and a small model also scores 90 points. In certain areas, customized small models may have an advantage because they would be more cost-effective and commercially viable.
Zhang Hongjiang: I disagree. I can give you an example right away. After ChatGPT came out, all kinds of translation software died.
Li Hang: But I learned that in the actual use cases of some companies in the US, the cost of GPT-4 is actually very high due to domain adaptation issues, or there are many examples of it not being very successful.
Zhang Hongjiang: Then that’s the problem of GPT-4: it hasn’t reached 90 points yet.
Li Hang: No matter if we look at it from an economic perspective or from an engineering implementation perspective, although large models may perform well in general fields, market demand is not always concentrated in general fields. There will always be certain niches where a particular small model or technology may be more advantageous overall. I don’t really believe that one single big model can do everything.
But I also agree that Google dominates the market with a whopping 90% market share. In this case, we can see that no one has been able to shake Google’s position yet, even though other companies are also working hard. GPT-4 also has a very stable market position now. And if users become dependent on it, it will be difficult to switch to other services. This stickiness makes it difficult to replace companies leading in general technologies. So in that sense, I agree that super-large models have broad coverage and strong market influence. But there are still opportunities for other companies.
Zhou Zhonghe: Looks like this question still deserves more discussion. Finally, let’s ask Mr. Xue: from the government’s perspective, how to better promote China’s AI industry?
Xue Lan: As for the government’s AI policy, one aspect is to promote development, and the other aspect is to regulate risks. So this is actually two wheels spinning at the same time. I think everyone has thought about government investment now. But I think Another very important point is how to create an ecosystem that allows Chinese companies and research institutions, including universities and others, to organically integrate with one another. This has always been a problem that China needs to solve.
This is especially important for China’s AI field. As Mr. Zhang said just now, we publish a lot of papers and have a lot of patents. What we lack most is an ecosystem, and this cannot be researched, it needs leaders to create favorable conditions so that all kinds of institutions and talent can play their role in such an environment. Just like the kind of supportive environment that Beijing Academy of AI has established and developed. …
At the global level, we need to establish a global risk prevention and control mechanism. This requires the joint effort of governments, enterprises, and research institutions. At the same time, we also need to realize that the development of AI is not only a technical issue, but also a social issue. It involves ethics, law, employment, and many other aspects. We need deep research and discussion in these aspects to ensure a healthy development of AI.
Regarding risk management and regulation, I would like to add one point. This is a multi-layered issue. First, as soon as possible, we should reach a consensus at the global level and formulate a global risk prevention and control mechanism. This is not only the responsibility of the government, but also requires the participation of enterprises and research institutions globally. Although everyone is working hard in this direction, it is not easy to actually implement these measures. It requires joint efforts from everyone.
Audience Q&A
Audience question 1: I am a 2021 graduate of Schwarzman College of Tsinghua University. I am now working in the industry ecosystem at Zhipu AI. Thank you very much, teachers. Your sharing for nearly two hours was very inspiring. I have two questions. The first question is that recently many Chinese artificial intelligence start-ups, especially those engaged in C-end applications, have chosen to go overseas to conduct business. It may be due to two considerations: first, hoping to use the most advanced overseas models, and second, the overseas payment market is relatively more mature. At the same time, most domestic companies focus on the B-side, helping local companies reduce costs and increase efficiency. What do you think of this development divide and trend? What impact will it have on the scientific research and industrialization process of artificial intelligence in China?
Second question, Sam Altman of OpenAI recently released a plan in which he plans to spend US$7 trillion to promote innovation and change in the semiconductor industry. What do teachers think about the impact of such large-scale financing as a private enterprise on the entire industry or academic research? And how might China’s ecology be inspired by it, or how can we find another path to match such an ecosystem?
Zhang Hongjiang: Answer the first question. Regarding going overseas, what everyone is paying attention to is when consumer-oriented companies (to C) are expanding overseas. But in fact, enterprise-oriented companies (to B) are going overseas more frequently, and this is actually very reasonable.
If we compare the software market and Internet application market in China and the United States — especially in the fields of software tools and cloud services (SaaS) — we will find that the US enterprise market (to B) is far more mature than China. Therefore, China’s enterprise-level software overseas strategy is not only reasonable, but also very necessary. After decades of development, China’s software industry still has a large gap in the toB field compared with the toC field. If I were a corporate decision-maker, especially in toB business, I would be more inclined to go overseas, because the market there is more receptive to paid services.
The core of toB software is to provide enterprises with tools to improve efficiency, such as Office, SAP, Feishu 飞书, etc., which are all tools to improve productivity. When companies use these tools extensively, it means they are pursuing efficiency improvements. Many of our companies have not yet made full use of these productivity tools, which shows that our production efficiency is not high enough and that we still rely on manpower to complete tasks. So, I hope everyone will see this not just as a business model, but as a key issue about whether we can improve production efficiency. Low production efficiency is obviously a very big problem for the industry.
In the past twenty-five years since I returned to China, I have always been very envious of the overseas toB software market. I hope that there will be a breakthrough in this area in the future so that we can truly increase per capita productivity — because the increase in per capita income essentially depends on the growth of per capita productivity. Otherwise, we will continue to face problems with per capita efficiency. That’s my take on the first question.
Li Hang: Let me answer the second question. The development of artificial intelligence does require significant investment. We just discussed that basic research should be open source and public, which is the source of the development of AI technology. Currently, AI technology, especially large-scale models, is more focused on engineering implementation. We have observed that, at least in some areas, the innovation capabilities of industry have surpassed those of basic research institutions. This is a new phenomenon, unprecedented in the history of human scientific research.
I believe that open source efforts will continue, as many academics are doing it, although their work may lag in some aspects behind leaders like Open AI. For example, projects such as Lambda are open source, but we must also realize that even open source from commercial companies has its own special nature and may have other considerations behind it. From a technology advancement perspective, open source may still face challenges because it requires a lot of development work and, more importantly, technological innovation.
I also agree that if it is a closed-source model, the government should carry out a certain degree of supervision, and companies should also take responsibility and do the right thing. This is very important. AI research is generally conducted in a large-scale environment, and the development of AI requires a large amount of investment. This is the reality we are currently facing. As we discussed before, in order to move the technology forward, we need significant funding to scale the models.
Zhang Hongjiang: I think the number 7 trillion is not necessarily directly related to closed source and open source. If you subscribe to the Scaling Law and have thought about how close we are to achieving AGI and how many resources are needed, you might come up with numbers like this. We may need to invest that much money, or even more, not just for one business, but there may be other businesses that need to invest a similar amount of money. This is not a slap on the head, but a reasonable prediction based on Scaling Law. As model parameters increase, we need more powerful computing power for training and inference, more computer rooms, more IDC facilities and power supply. If current models were scaled up 1,000 times, such an investment might be necessary. In order to build this system and achieve this goal, corresponding investment is required. I think this is his thinking, and from Sam’s perspective, this guy often thinks about problems ten years in the future.
Audience Question 2: I graduated with a PhD from the Department of Precision Instruments of Tsinghua University. I am currently doing technology industry research on artificial intelligence at the China Academy of Information and Communications Technology. My question is, first of all, we mentioned that the development of artificial intelligence relies on three core elements: algorithms, computing power and data. At present, algorithm research may face the problem of dispersed research power. In terms of computing power, it is affected by the US ban on the export of high-end AI chips. In terms of data, high-quality Chinese datasets may not be sufficient. Faced with these challenges, how should we weigh the priorities of these three and which one should be developed in a limited manner?
Secondly, if computing power is currently the main factor restricting the development of advanced large-scale AI models like Sora in China, can we use national power to integrate existing domestic high-end computing resources, such as NVIDIA chips, to support the research and development of domestic scientific researchers to achieve breakthroughs?
Li Hang: The first question, as I mentioned before, in the long run, talent cultivation is the most critical. Although computing power currently encounters a bottleneck, this is a relatively short-term problem. In the long run, the development of artificial intelligence is inseparable from the cultivation of talent.
Personally, I think undergraduate education is very important. It is also important to integrate university research with industry to promote research in the field of artificial intelligence. For long-term development, the most important thing is talent. Short-term problems, such as data problems, are relatively easy to solve — but talent cultivation requires the joint efforts of the entire society.
Zhou Zhonghe: Speaking of talent, I thought of a question. We usually focus on the number of people and the number of published papers, just like the ranking mentioned by Teacher Zhang before. But in China’s science and technology field, what we need now are truly innovative and top-notch talent. This is not just a matter of quantity, right?
Zhang Hongjiang: That’s right. Especially in the future, those who make up the numbers are likely to be replaced by AI. What we need is the kind of people who can do real innovation.
Xue Lan: Let me talk about the second question. I think this is an assumption that is difficult to realize. If we go back to a few decades ago, under the nationwide system and planned economy, it might still be possible. I think today it would be very difficult — so if this premise does not exist, there is no need for us to discuss the later parts.