The Ecosystem and Technological Trends of LLM in China and the U.S.
Source:https://mp.weixin.qq.com/s/pIOm2QZbuE6AvgW_ucdWBw
On April 12, 2024, Tencent Research Institute, in collaboration with the Qianhai International Affairs Research Institute and Qingteng, jointly launched the AI & Society high-end seminar on artificial intelligence + social development. The theme of the first session was "The Entrepreneurial Ecosystem of the Era of Large Models," and the seminar was held in Qianhai, Shenzhen.
The conference invited industry leaders and renowned scholars in the field of artificial intelligence and social governance from both domestic and international backgrounds. The discussions mainly focused on the topic of building an entrepreneurial ecosystem for large models. Through expert keynote speeches and roundtable discussions, the topics of the ecosystem and technological trends of large models in China and the United States, as well as AI investment trends and application directions, were discussed.
The following is the transcript of the roundtable discussion on "The Ecosystem And Technological Trends of Large Models in China And The United States," as confirmed by the guests:
Speakers:
Si Xiao (司晓， Roundtable Moderator), Vice President of Tencent Group and Director of Tencent Research Institute
Zheng Yongnian （郑永年）, Professor at The Chinese University of Hong Kong (Shenzhen) and Director of Qianhai International Affairs Research Institute
Yao Xing（姚星）, Founder and CEO of XVERSE
Dai Guohao（戴国浩）, Associate Professor at Shanghai Jiao Tong University and Co-Founder & Chief Scientist of Wuwen Xinqiong
Si Xiao:
Next, we will start the first roundtable discussion. I believe everyone found the insights shared by the three guests earlier quite enriching. We want to continue with some in-depth discussions in this roundtable session. You'll notice that our panel is quite diverse, and from the speeches of the experts earlier, you can sense that the design of this forum is well thought out.
Firstly, we have the first question for the three guests. Both Professor Zheng and Mr. Yao mentioned the comparison between large models in China and the United States. I believe that from the perspective of experts and frontline entrepreneurs, there may be some differences. I would like the three of you to discuss from your respective perspectives the main differences in the development of large models in China and the United States, and what factors have led to these differences in influencing the development of the large model ecosystem. Looking ahead, what similarities and differences might there be in the competition and industrialization paths of large models between China and the United States?
Zheng Yongnian:
I wouldn't call myself an expert; as a humanities person, I don't understand technology much. But I saw a video last time discussing the differences between China and the United States, saying that American entrepreneurs start businesses to realize ideas, while Chinese entrepreneurs start businesses to become bosses. I think this situation exists. Why has chip development in the United States been so fast? Why doesn't the U.S. government regulate this? It might stem from a sense of cultural mission. Look at many of Musk's behaviors; they are hard for us to understand. When Musk created Tesla, Tesla was just a tool for him to pursue another goal - to conquer the universe. Such thinking is rare here. Our civilization is very practical; we learn fast and can do things better.
From this perspective, it is likely that the future of originality will still be led by the United States. This outcome is not because Chinese people are unintelligent; on the contrary, Chinese people are very smart. If you look at the teams leading chip development in the United States, they are mostly Chinese. To cultivate the ability for original technological innovation from "0 to 1," we need institutional reforms. I have also talked about how to overcome the "middle technology trap." This is a systematic project involving technological invention, basic scientific research, applied technology, enterprises, markets, and government at various levels. If we view originality as "0-1" and applied technology as "1-10," we find that anything that can achieve "0-1" in originality can also reach "8-10" in applied technology, meaning it can be the best. But if it can't achieve "0-1," it's like China is stuck at "4-7 or 8," and can't progress further, leading to this situation. In terms of technology, the speeches of the guests have inspired me on how to apply technology, which we can do better. How to shift towards applying technology, with some original elements, is the key to future competition between China and the United States.
Si Xiao:
Professor Zheng's words reminded me of my own experience. In 2015, when I was a visiting scholar at Stanford, I rented a room similar to a youth hostel environment, where everyone did their own thing during the day and interacted in the lobby at night. One thing that struck me was that people from outside the United States who came to Silicon Valley to start businesses firmly believed they could change the world. You might wonder, how could they think that? How can they start changing the world with nothing? But they truly believed deep down that they could change the world, and this culture differs significantly from what we have domestically.
Earlier, Mr. Yao also mentioned the moon landing project, which in a way is also a belief in technology. We all know that OpenAI was initially ridiculed for sticking to the Transformer path, with many thinking it was not feasible. But later, it was also based on a belief in technology, taking this thing to the extreme. How do you view American innovation and entrepreneurial culture? And what aspects could China possibly learn from?
Yao Xing:
Professor Zheng mentioned earlier that in China, from a policy perspective, when it comes to making cars, they first make the brakes and then think about the car. This approach considers the issue comprehensively from a political and social development standpoint. I feel that we tend to focus on safety factors first, while they just dive into making the car. The same goes for autonomous driving technology. During my recent visit to the United States, Tesla's FSD V12 for autonomous driving hadn't been released yet. After experiencing it, I learned that they plan to remove the steering wheel in the future. I asked, "Why would you do that? Isn't it dangerous?" They said it's something Musk wants to do; he wants to push the boundaries to the extreme. However, from a Chinese perspective, it seems quite unreliable to me. Removing the steering wheel is seen as pushing the limits to the extreme, which puts a lot of pressure on the company. They themselves feel unsure and believe there are many bugs. From a legal policy standpoint, the responsibility used to lie with the user, but now it shifts to the car manufacturer. If there's no steering wheel and an accident occurs, the manufacturer will be held responsible. This kind of thinking represents a certain mindset—do you prioritize having brakes first or the car itself? It's a matter of habit, and changing it will take time.
Moreover, in terms of the environment, we tend to focus more on returns, right? How's the return on investment (ROI)? They, on the other hand, consider more long-term aspects. It's hard to say who's right or wrong. From a rational standpoint, doing foolish things without considering returns can be scary, but if you don't take risks, you might never reach the moon. When I was in the U.S. talking to them, I asked if they thought this was a bubble. They also thought it might be a bubble, similar to the situation in 2000. However, their attitude is like, "So what if it's a bubble?" The losses might just be a few hundred billion dollars. With the U.S. economy, including the stock market, having seen such growth, they believe there's an opportunity to try it out. Due to our cautious and rigorous approach, we often miss out on explosive opportunities. Following and adapting is crucial for us—it's like feeling our way across the river. I think this is a cultural habit that will take a while to change. Each approach has its own advantages, and I believe we are adopting one of the safest methods.
However, we should also recognize some of our strengths. When I visited NVIDIA the other day, I mentioned that the failure rate of our H800 is probably around one in a thousand. They were quite surprised and asked, "How is that possible?" They mentioned that what they see at OpenAI is in the single digits. I explained that it's because we use the best equipment and NVIDIA's technology. They asked why that was the case, and I said it was because of the diligence of Chinese people. Being on call 24/7 and promptly addressing any issues—if someone goes on vacation, the failure rate might increase. I believe that the diligence, cultural aspects, and service mindset of Chinese people are hard to replicate. So, in the competition between China and the U.S., I think there will be a period of comparison and competition.
Si Xiao:
After the two guests' discussions, I remembered a popular online media figure, Liu Xingliang, who once explained why Europe lacks internet companies. One point he made, almost like a joke, was that the U.S. leads in innovation, China leads in application, and the EU leads in legislation. This is why they don't have major internet companies, but this view may be biased and could also be related to the lack of a single language market in Europe.
Zheng Yongnian:
The issue of "returns" is crucial and a constraint for us. Looking at it practically, businesses need to generate returns; there must be ROI, or how can we operate? The problem for us up to now is that we haven't solved the issue of having a venture capital system like the one in the United States. When we talk about venture capital, it's a different concept from what the U.S. has. Venture capital involves "playing" with money, so do we not have money? We do have money, but we lack the space to "play" with it. This is why we see a significant outflow of wealth and capital leaving; it's because we haven't provided enough room for them here. We need to address this financial issue and ensure that finance and the real economy are not in opposition. When we talk about building a strong financial nation, it's all about financial regulation and control policies without leaving room for financial development. This fundamental issue hasn't been resolved. After World War II, the U.S. used a lot of money to build its manufacturing and industries. If the U.S. didn't have the venture capital system, its industries today might not be what we see. So, I think this is an ideological issue. Perhaps during the early days of reform and opening up, we lacked money, but today, it's not about lacking money; it's a systemic issue.
Si Xiao:
Zheng Laoshi just mentioned that our forum must have an investor's perspective for it to be complete. Professor Dai, the project you're working on, you mentioned in a previous interview, is something like a trading platform for computing power, like the "Taobao of computing power," right? Maybe I think it's more like the "Tmall of computing power" because the barrier to entry on the supply side is actually quite high. Could you share your thoughts on this topic from your perspective?
Dai Guohao:
Earlier, we were debating whether "Taobao of computing power" or "Pinduoduo of computing power" is a better term to use. Of course, it's all in good fun (laughs). I want to talk about two things. Today, I learned a lot from the sharing of Professor Zheng and President Yao. At lunch today, we were discussing the overall social environments in China and the U.S. Chinese people, which are too intense. You could say that we started late, and nurturing original innovation takes time to develop. You need to survive first before you can thrive. Particularly in cities like Shenzhen, I feel a strong entrepreneurial culture. I arrived last night at 11:50, and while the driver was picking me up, we were chatting on the way. He mentioned how many areas in Shenzhen are developing rapidly; it used to be Bao'an, but now it's Qianhai. In other cities, when looking back at their development history, they might talk about what it was like thirty years ago, but in Shenzhen, they talk about what it was like five or ten years ago. I could distinctly feel how our hardworking individuals have built a city amidst the currents of the times.
Why did Professor Zheng mention that it's always Americans who do the "0 to 1" innovation? I think it's mainly because of the high pressure to survive. When we look at the differences in the B2B ecosystems between China and the U.S., the U.S. tends to do better in B2B business. This is because they have a very clear sense of boundaries, and each party plays its own role.
But Chinese people might think, if I do my own thing well, can I also do a bit of his thing? Can I earn more money? I used to travel to the U.S. a lot recently and found that the highways in Silicon Valley start getting congested around three or four in the afternoon. When do you see traffic jams on the highways in Shenzhen? So, I think it all boils down to the fact that everyone's survival pressure is different. China isn't lacking in originality; we have the Four Great Inventions and various technologies that emerged during the Tang and Song dynasties. When the economic scale is sufficient and leading, we naturally have the energy to do "0 to 1" innovation. This takes a certain amount of time and social development to address.
We still need originality. First and foremost, I am a university professor. My students have two directions: some are system-oriented, and they may not publish as many articles as students focused on algorithms. Sometimes, they say, "Professor, I'm in a rush. I want to publish more articles. My neighbour might publish several first-author articles in a year, but for me, working on the code alone is already quite laborious. It takes even more time to turn it into a paper for publication. What should I do?" I tell my students not to rush because students are the most innovative group at the source. This is the time with the lowest cost of trial and error because once you enter society, you face survival pressure, making it challenging to do things like the "0 to 1" that Professor Zheng mentioned, as the cost of trial and error is too high.
In our current scientific and technological innovation industry, we excel at taking technology from one to ten, improving performance. As President Yao just mentioned, we have managed to reduce error rates and failure rates from one percent to less than a thousandth. These are all things at which our industry excels. So, what I tell my students is that you need to have ideals, you really need to have some beliefs, like entrepreneurs in Silicon Valley. Making mistakes is okay because we have provided you with a good platform. I firmly believe that in the future, with economic development and everyone's joint efforts, perhaps in ten or twenty years, more original things can emerge from Chinese companies. By doing one to ten well, we can create more space for future 0 to 1 innovation.
Si Xiao:
The three guests spoke excellently. Let me summarize a bit. It seems that some early intervention in systems and regulations may not be good, but if the joke we just mentioned is valid, the real negative example is the EU, right? They have been overly aggressive in regulation and too conservative, leading to a lack of industry development. A typical example is the establishment of GDPR, a threshold for personal data protection. What was the original intention? Surely, it was to keep American companies out, impose high compliance costs, and easily impose fines. However, they didn't consider that the same regulations would apply to local entrepreneurs and small businesses may not even afford legal fees. So, considerations of systems may be quite comprehensive. However, not everything can be attributed to strict regulations or systems that cause issues. For example, U.S. companies have a so-called geek or technical belief, but they are a minority. Companies like Google or Apple, although technology-driven, don't seem as tech-enthusiastic as OpenAI or Musk. Applying technical beliefs to them may not be suitable. As President Yao mentioned, if every company becomes a technical believer, it could be quite risky. The bigger base still needs to follow ROI and adhere to shareholder return responsibilities; otherwise, the investment market may struggle to operate effectively. But perhaps its strength lies in tolerance for mistakes or the existence of madmen. This may be a strong point in the U.S.
Apart from that, I also want to ask Teacher Zheng, as you mentioned earlier, to compare nuclear weapons to artificial intelligence. One is just there for show, can't really be used, it's for deterrence, so it can achieve a balance of terror between countries. But artificial intelligence, obviously, is a conventional weapon, it can be used, and it's a productivity tool, and no one would say productivity can't be used; of course, nuclear weapons and artificial intelligence have another connection point, that is, AI weapons cannot be militarized, it's just banned in that direction, but it's definitely a conventional tool for use. So, what kind of impact will the next development of artificial intelligence have on China and the U.S., and even on the world's multipolarity? Can you elaborate more on this?
Zheng Yongnian:
In the field of large-scale artificial intelligence models, the United States currently holds a monopoly position. In fact, the threshold is getting higher and higher, making it difficult for small countries to develop artificial intelligence - large models require a lot of investment, which most countries cannot afford. Faced with the dominance of the United States, China's development is crucial. I have talked to European friends, and they also hope that China can rise because, currently, there is no country that can "counterbalance" the United States. Only when China rises can the situation possibly change. China should not be too conservative in practical development. Firstly, our regulatory technology is top-notch in the world; in fact, using our current regulatory technology for development may be safer than in the U.S. Secondly, our applications are very powerful. Additionally, we have talents in artificial intelligence, such as those trained in top universities like Tsinghua, but we lack sufficient platforms, leading to talent outflow to the U.S. So, from a global political perspective, China can excel in the field of artificial intelligence, even better than the U.S. The current issues to address are finances and openness - how to solve the return issue and the data issue. Once these two are resolved, we can quickly catch up. I don't believe that China lacks originality; it's just that we have somewhat tied our own hands and feet.
I was talking to my American friend, and I said that China's regulatory technology will eventually be needed by the United States someday. If there's no such regulation, AI could destroy society - Musk can use AI, and even lunatics can. If AI is combined with nuclear weapons, it becomes even more powerful. So, we need to understand that China's regulatory technology is crucial. The U.S. is like the "spear," and China is like the "shield," so in reality, both countries complement each other very well. China should learn from the U.S., and the U.S. should learn from China. Mutual learning is essential for a peaceful world; otherwise, the world will definitely be imbalanced. In our world today, after the advent of the internet, there's a stark division in intelligence. A small group of people, like Musk, who control AI, are becoming smarter, while the majority are becoming less so. People like Musk, who controls AI, are seen as "shepherds," and now AI is the "sheepdog," and most of us are just "sheep." I worry that one day, the "sheepdog" will surpass the "shepherd," and then AI might rule the world, which is a threatening prospect. We are no longer in an "AI & Society" scenario; we are now in an "AI Society." Even though many say we should think before developing a technology, once a technology is created, no one will abandon it unless there's a consensus to destroy it - which is unrealistic. We can only move forward.
Si Xiao:
I completely agree with you about removing the "and" in "AI & Society." I think we are currently in the process, and when we naturally drop that "and," that will be the true artificial intelligence society. Just the other day, I came across a video where Li Feifei and Geoffery Hinton were discussing. Li Feifei mentioned that Stanford should recruit students who are proficient in using ChatGPT and large models. In a way, those coming out of Stanford should also be adept at using ChatGPT because they need to command the virtual armies of AI, essentially representing various AI roles. Earlier, we discussed innovation, and I would like to continue by asking both Yao and Dai, as entrepreneurs, what specific constraints China faces when developing large models, such as computing power; besides diligence, what other advantages do we have? Though hard work can compensate for a lack of skill, do we have any other advantages?
Yao Xing:
Computing power is something everyone can see because large models follow a formula - roughly your parameter quantity multiplied by your token count, then multiplied by how many operations, let's say ten. Currently, the challenge is around 10 to the power of 25, which can be calculated backwards. If you need to train it in three months or a hundred days, how many cards would you approximately need? This is a rule for reverse calculation. If you multiply billions of parameters by trillions of tokens or use trillions of parameters multiplied by tens of trillions of tokens, you're around 10 to the power of 25, then multiplied by ten floating-point operations. How many times can an H100 or H800 card compute per second? It's about 700T for H100 and over 200T for A100, which is probably more than a third of H100. But clearly, we are stuck in terms of computing power. Our capabilities are still quite far behind. If the time needed is stretched from three months to ten times that, becoming three years for each iteration, it's like crossing a river by feeling the stones; you'll never quite reach it. It's like being in a different physical space where a day for them is equivalent to 72 hours, while you only have 24 hours. How can you compete with that? You can't keep up. This is why computing power is the biggest bottleneck because the physical computation rules are like this. Unless you find a more efficient algorithm, saying, "I don't need that much data," unfortunately, everyone is searching for these methods, but they are not easily found. Physically, if it takes three years instead of three months to do something, you're essentially out of the competition. Let alone three years, even iterating once a year puts you out of the competition. So, I believe the biggest hindrance so far lies at this level. We might be able to catch up a bit now, but if the next challenge is, for instance, 10 to the power of 26 or 27, and we don't have the resources now, we might never reach that point. This, I think, is the most significant limitation. Of course, the advantage is hard work, which indeed can be helpful. Seven times 24 hours is much better than five times eight.
But I think another benefit is that, from the perspective of the Chinese community's wisdom, I believe that today being able to grasp these things is not only for us mainland Chinese but also for the entire Chinese community. It's about how to create an environment where China and the U.S. can learn from each other and exchange ideas. I think this might be a future necessity, something that needs to be resolved in the long term. The relationship between the two should not be inherently adversarial. If we look at it from a genuine perspective, it should be about mutual learning. Clearly, we have some areas where our systems are better. For example, when I went to the U.S. this time, there were many incidents of zero-cost shopping, smashing windows, and cars being smashed as soon as they came out, a lot of incidents. Can we really say that the U.S. is a safe society? In such a powerful country, I didn't feel safe at all; it felt very unsafe, with cars being smashed in broad daylight. So, indeed, there are many areas where they could learn from us. Of course, if we can resolve these issues well, I believe that with our intelligence and diligence, with some constraints removed, we could potentially create a harmonious world.
Dai Guohao:
I completely agree with Mr. Yao, the core issue is computing power. Large models have shown us a way to infinitely scale, which is what everyone refers to as the Scaling Law. The general understanding of the Scaling Law is that the more parameters you have, the possibility of fitting into a larger space increases, making the model more powerful. What we at Wuwen Xinquan are doing is ensuring that the deployment cost of a specific model is kept low so it doesn't require a high threshold of computing power or financial investment to scale up the model in different scenarios quickly.
Why did we spend a long time discussing whether to focus on Taobao or Pinduoduo?
It's because our logic is to reduce costs while ensuring the quality of the model. Through joint optimization from algorithms to hardware, accelerating at every level from application to algorithm, model, including operators down to the underlying hardware. This requires a high level of team capability and considerable expertise in products, algorithms, systems, and hardware. Our team background is the most suitable domestically to get this done right. Recently, when I was in Silicon Valley chatting with many friends, we discovered that the most in-demand positions are not people training large models or engineers working on large model algorithms; it's actually Prompt engineering. As the two teachers just mentioned, this is essentially using AI or commanding AI, and I completely agree with this point. Why? Because I think this year might be a turning point; some even say this year could be the year of large-scale model reasoning. We've observed that under the backdrop of the Scaling Law, beyond just increasing the number of parameters to make the model stronger, for a specific scale of a model like a 7 billion parameter model, the gap in capabilities between open-source and closed-source models is narrowing continuously. Instead, there will be more and more people building applications on top of these models. A 7 billion parameter model is fundamentally a finite space and a finite space always has an upper limit. With time, the capabilities of a specific scale model will approach this limit. And on some relatively medium-sized models, through various tuning efforts, they might slowly reach that level.
What do we do after the model's capabilities plateau? We deploy, reason, continuously deploy, and apply. So, Wuwen Xinquan, including many of my own research endeavours, aims to keep costs low during deployment. It's like using a steam engine to generate power. When the steam engine was invented, people weren't replaced; instead, more industries using steam engines and generators emerged, allowing people to harness the tools better. This year could be a significant turning point. We can truly start using models in various scenarios, such as specific-scale models and scenarios like mobile applications or AIGC applications. We also hope that through our efforts, we can at least reduce these costs and help everyone better utilize large models.
Si Xiao:
I want to ask the two experts with technical backgrounds again. We're currently talking about computational power being a bottleneck, mainly in the training phase. But will there be bottlenecks in the inference phase in the future? From the perspective of chip substitutability, the substitutability in the training phase may be very poor; everyone can't replace NVIDIA. However, in the inference phase, there are many chips available domestically. So, how do you see the possibility of bottlenecks in chips in the inference phase in the future, or is this phase okay?
Yao Xing:
I think the training phase is more apparent, but there are also issues in the inference phase. Inference is heavily related to VRAM, and the size of VRAM determines the size of parameters. Larger models like A800 may even require several cards; why several? Because the volume is increasing. Currently, in terms of technology, there are relatively few domestic efforts in this area. Pure computational speed itself has advantages for inference. Today, we're all just shifting the focus to the training phase. For example, when we create a universal large model, this model will already be super AI in the future, and there will be no need to train it further. But the inference phase serves society and people. If you want to reduce the cost of single-frequency items to serve, you need to focus on inference. If your inference is inferior to others, the ROI will also be poor. So today, I think it's the entire chain. Looking at the pipeline of large models themselves, from the chip perspective, there's a significant gap in the entire chain. We should objectively assess this; otherwise, our progress with large models will be challenging, and our training time will be much longer than others.
Dai Guohao:
I completely agree. Whether in training or inference, fundamentally, it's all about computation, using digital chips to perform calculations on algorithms. Training has its algorithms, and inference has its own. When looking at the Scaling Law, we're examining how close we are to the upper limit of capabilities. But ultimately, the inference is determined by costs. To create a steam engine, I could incur significant costs, but when it comes to installing this steam engine in every corner of society, the cost must be low. In the future, as long as the energy supply remains constant, the cost will be electricity, where electricity equals the energy cost of moving each digit multiplied by the number of digits moved; this is the cost of storage.
Additionally, there's the cost of computation. In digital circuits, an algorithm is ultimately expressed as the number of times 0s and 1s flip in the circuit and the energy consumption per flip. The number of flips is determined by the algorithm, so we need to explore more efficient algorithms. The energy consumption per flip depends on the process. Fourteen nanometers are not as efficient as seven nanometers, and seven nanometers are not as efficient as three nanometers. Today, the gap in inference is not nonexistent; it's just that people aren't paying much attention to it; everyone is focused on training. The difference in training lies in the fact that OpenAI started five years ago, while domestic manufacturers may not have started as early.
Inference is a whole different ball game. Today, we can see its future scenarios and can strategize earlier. Take the number of flips, for example; it's something we can optimize algorithmically. For instance, we can do some model compression or train smaller models. A 3B model, compared to a 7B model, can have significantly fewer flips. Although the gap between seven nanometers and three nanometers still exists, this coefficient upfront can become smaller, allowing for earlier entry into the race and earlier strategic positioning in these domains. This is crucial for our future competitiveness and ability to go head-to-head. We still need to overcome some computational bottlenecks and process issues to excel in some areas, but at the same time, we should also consider whether, in the presence of certain hardware process gaps, there are opportunities to leapfrog in other dimensions.
Si Xiao:
Alright, I have another question for the three experts. In the era of large models, you'll notice that the scale and size of startups are decreasing. Some have mentioned that in the AI era, startups might need to think about globalization from Day One. How do you all view this issue? When the fundamental premise changes in the era of AI, what differences would there be in a company's organizational structure compared to the previous era?
Yao Xing:
The decrease in numbers is definitely related to productivity and tools. If we view AI as an AI OS, with more advanced OS systems emerging, it's clear that there's no longer a need to start from the very bottom to build things. Each generation of OS advancements actually shrinks the company's workforce because productivity is increasing. It's similar to the Industrial Revolution; with the advent of machines and steam engines, the workforce decreased significantly. The same goes for the Information Age. So, the reduction in company size is likely due to a significant increase in productivity. As Dai mentioned earlier, why do we need to hire prompt personnel? The paradigm has shifted; we no longer need to build many things from the ground up. Based on models, using the prompt method, we can create many applications. So, users and questioners are more focused, while those building these applications are less so. This is normal. When we use steam engines, do we still study the principles of steam engines after all these years? Many things no longer require in-depth study; that's how society progresses. Just like when we write code in high-level languages, do we need to understand assembly language? We don't need to understand assembly language or single-chip principles. This is quite normal, so the trend of companies shrinking is an inevitable part of societal development. As for starting globally from Day One, it's been a hot topic in China recently. I think it's related to the environment and the investment environment. Small companies in China struggle to secure funding, so they always look to more active, open, and less restrictive places for capital. This is one reason why they aim for globalization from Day One. We've also seen many Chinese companies moving to the US to operate there.
Another aspect is commercial returns. China faces many commercial challenges, including issues with intellectual property rights. Everyone is competing in the same space. Americans compete, too, but they choose different paths. They highly respect intellectual property rights; if they create something, others feel hesitant to imitate or follow suit. We must always strive to create something new. In China, everyone rushes into the same field, hoping to come out on top. It's quite cutthroat. So why go global from Day One? It's because it's tough here. Perhaps other places offer better capital and business opportunities, which align with the overall trend. In China, it's hard to outcompete with each other, but outside, they aim to outcompete with others. That's probably how it goes.
Si Xiao:
I'd like to ask Teacher Zheng to share his thoughts. I feel what you've discussed may not be limited to changes in the organizational structure of companies or society.
Zheng Yongnian:
Earlier, everyone was talking about "juggling." First of all, when we compare ourselves to the U.S., you might ask, "Is Musk juggling?" He is, indeed. But the direction of our juggling is off. We're truly caught up in "internal juggling" while they're engaged in "external juggling." From the days of Taobao to Pinduoduo, and now with online live streaming, the entire society is being dragged down—consumption is downgrading, and society is sinking. Musk represents external juggling. Why do they talk about open-source? It's a cultural thing, a sense of mission. Especially in entrepreneurship, I've observed a phenomenon where many startups are incubated in China but end up setting up companies in Southeast Asia because they can't find room to grow domestically. This is a significant loss for us, and this situation is becoming increasingly common, with a greater impact on society. As a university professor, I'm deeply concerned. In the era of ChatGPT, what's the reason universities exist? ChatGPT can provide people with a vast amount of knowledge in a short time, directly challenging the value of "professors." To a large extent, universities have degenerated into a group with the loudest voice of vested interests, their legitimacy being self-justified rather than accepted by society.
For thousands of years, until the pre-internet era, it was the young who learned from the old. But with the advent of the internet, human society has completely flipped; now, the elderly need to learn from the young. So, why do I advocate for reform, including our capital flow? The AI world is getting younger and younger. We shouldn't keep pouring resources into academicians; we should allocate more resources to the younger generation. After the era of Liang Qichao, where youth is strong, China is strong, and the youth is strong, China is strong, this is the current era. We need to shift the focus to the youth. Currently, the difference between us and the U.S. lies in the fact that they are truly an "innovative" society, while we are a "conservative" society. Why can't we produce a Musk? Because Musk wouldn't survive in China. Yesterday, our principal told me that it's challenging for schools to do something. When we try something new, out of 100 people, 99 will oppose you, and only you will be thinking about innovation. So, I'm very anxious; we need openness, and there's hope if we open up.
Dai Guohao:
As a young teacher, I'm very grateful to Dean Zheng. Young teachers indeed need more resources because they can accompany students in coding and writing papers, and at least I still have the energy to stay up late with my students to meet deadlines.
Returning to the previous question, it's not just about young people needing more communication and information exchange with their elders; international relations need the same. Today, where is the most concentrated source of AI information? It's in the cafes of Silicon Valley. I hope for a more open and inclusive environment in the future, whether it's between China and the U.S. or between teachers and students.
Si Xiao:
Even though we're running overtime, the forum is being held in Shenzhen. Could each of you provide some suggestions for the development of artificial intelligence in Shenzhen or the Greater Bay Area? Or perhaps highlight areas that need improvement? I believe entrepreneurs should have some insights.
Zheng Yongnian:
In February this year, The New York Times published an article by columnist Ross Douthat titled "Only America Can Save the Future," mentioning the concept of a "new Faustian" civilization, a new America spreading from Houston to San Francisco. So, I've been thinking, where will the emergence of China's "new Faustian" civilization be? I believe China's "new Faustian" civilization is rising in the Pearl River Delta and the Yangtze River Delta regions, especially in the Greater Bay Area, with three new pillars—basic scientific research, applied technology transformation, and finance, which are no less than California in the U.S. So, I say, this is the land of hope for the new civilization in the AI era, and I hope we can fulfil our responsibilities, achieve deep integration, seize the opportunities of the times, and realize Chinese-style modernization and high-quality development.
Yao Xing:
In fact, I have quite a lot of exchanges with Beijing and Shanghai as well. As an entrepreneur, I feel there is quite a lot of hope. Currently, the AI industry in Shenzhen is certainly behind Beijing and Shanghai in China, so let's not compare ourselves to the Bay Area for now. However, the Greater Bay Area is young and vibrant, and Hong Kong has a dense pool of talent. There are many excellent schools, but many mainland students find it challenging to study there. Both the Hong Kong University of Science and Technology and the Chinese University of Hong Kong have many outstanding AI teachers who have nurtured many students. But why haven't we emerged in this wave of AI development? Let me make a suggestion today. I think perhaps the original Guangdong Province or the Bay Area focused more on foreign trade-oriented economies, on places dealing with imports and exports, and on high-end manufacturing. In terms of understanding artificial intelligence, perhaps the top government agencies need to keep up with the times and understand more to know how to build a talent hub for AI. This includes support in computing power, funding, and industry, which is still lagging behind Beijing and Shanghai today. I think maybe the government needs to enhance communication with figures like Dean Zheng at the policy level to hear these voices in order to bring about better changes.
Dai Guohao:
Let me share my experience. After turning 18, I did my undergraduate, PhD, and postdoc at Tsinghua University, so I spent 12 years in Beijing. At the end of 2022, I went to Shanghai, where I spent 2 years. Today, participating in this forum in Shenzhen, I can see that these three places are currently the most advanced in AI development in China. It's deeply felt that each place has distinct characteristics or features, and as Yao just mentioned, there may be differences in some aspects, but I believe that the presence of these unique characteristics will lead to significant differences in everyone's future development. What AI is competing in right now, in my opinion, is talent. I believe that in the future, especially as Dean Zheng mentioned, there will be many opportunities for us to speak up. We must prioritize talent, which is something that is highly valued in Beijing and Shanghai at the moment. Computing power or other means of production, I believe Shenzhen can excel, but talent is something that cannot be compensated for economically. It's a profound matter that requires joint efforts from the government, academia, and industry. So, we are also very eager to see more talent coming to Shenzhen in the future to help Shenzhen do this well. So, a small suggestion: I also hope that in the future, Dean Zheng can represent us young people and speak up more. Thank you.
Si Xiao:
I believe everyone has gained a lot from this round of dialogue. Once again, let's thank all the guests with warm applause.