Roundtable Discussion 3: The Definition of Artificial General Intelligence (AGI) Urgently Needs to b

Pan Qingzhong：NextJoin the roundtable discussionIn the process, weThere are many problems, but time is limited.Hongyi, Professor Shen just mentioned that the gap between China and the US was not large at first, but it has widened somewhat since then.He asked this question.I'd like to ask, do you agree with what he said?Or do you have any new ideas?

Zhou HongyiI completely agree with all his ideas about the bright future of China's economy.

Pan QingzhongYou didn't answer my question, you only responded to his.

Zhou HongyiActually, it's normal to have different viewpoints and debate between large-scale models. The "thinking chain" approach emphasizes reflection, correction, summarization, and induction. Personally, I don't think the gap between large-scale models from China and the US is that significant. I know many people will criticize me for saying this, because the difficulty of large-scale models is unlike lithography machines or chips, which face many obstacles of original technological innovation. The algorithms for large-scale models are public, the papers are openly discussed, and many models are open source. We should be grateful for the open-source culture. Moreover, in the US, including GPT and OpenAI's Sora, which hasn't been released yet, many domestic companies like Kuaishou and Douyin have also released seemingly good products in the past six months. I think the gap is at most about a year. Because large-scale models are still composed of software and algorithms. But where does the real gap lie? My previous view that it was mainly in computing power may be outdated. As I mentioned earlier, the US is currently building large-scale models, like Facebook's 100,000-card cluster. They're training a 10,000-card cluster of Lambda3, and it's making a lot of mistakes every day. A 100,000-card cluster might have a card fail every half hour, requiring data adjustments. Huawei's cards are still very good, but there's still a gap compared to them. However, the recent emergence of reinforcement learning in GPT-o1 has changed this view. Reinforcement learning no longer aims to train a model with infinitely large parameters. For example, our original large model learned a lot of gossip and knew all the gossip questions on the internet, but its reasoning ability was very poor. A child who doesn't read gossip every day but studies hard can also have very strong mathematical reasoning ability, but if you ask him what happened recently, he won't know. This doesn't affect the model.

Pan QingzhongThe gap is not large, and we may even have advantages in many aspects.

Zhou HongyiThe second gap is data. There are still some issues with the quality and quantity of data on the Chinese internet. However, if we apply the large-scale model to industries now, we don't need to demand as much from the messy data on the internet. We can demand professional data, data on the real know-how of many companies. We are relatively optimistic about the development of the country's large-scale model.

Pan QingzhongAI presents many challenges, especially for Chinese entrepreneurs. If you are an entrepreneur, how should Chinese entrepreneurs face the challenges of AI? These challenges include employment, improving efficiency, structural adjustment, and corporate restructuring.

Zhou HongyiI'm a tech person, a product manager. Entrepreneurs still need to develop an understanding of AI. You have to believe this thing is real. You have to believe that if you don't use it, AI might not eliminate you, but your competitors using AI capabilities will. Entrepreneurs don't need to be too anxious. Find the bottlenecks and difficulties in your internal processes or in the external product and service experience, such as those that require a lot of manpower or have a particularly poor user experience. Make breakthroughs in these single points by using open-source large models to create professional large models, solving problems one by one. Make sure that the large model doesn't become a toy like a chatbot, but truly improves the labor productivity of the enterprise in a certain link. This is how it becomes a tool for new quality productivity, accumulating small victories into a big victory. When an enterprise has more large models and solves more problems, it may not necessarily change the industry, but it itself realizes what we call digital transformation and intelligent transformation. There is a prerequisite for using large models. Why digital transformation and intelligent transformation? Digitalization is the foundation of large models. Without digitalization and data accumulation in an enterprise, data cannot be transformed into information or knowledge. Developing your own large model is like a tree without a source, water without roots, or a skilled cook without rice. Entrepreneurs are overthinking things, and there's nothing I can change. What I can change is whether the company can popularize AI within its ranks, enabling all employees to use AI to improve their skills, and whether the company's specialized large-scale models can truly improve labor productivity. I think that's a more pragmatic approach.

Pan QingzhongThat's very well said, thank you very much.Hongyi talked about the promotion and transformation of enterprise models.I'd like to ask Dr. Xiaochuan, you just mentioned that the current model competition is basically a battle of a hundred models or a battle of a thousand models. From an industry perspective, when will the industry reach a certain level?Is it stable?Or has the industry developed to a certain stage?

Wang XiaochuanFirstly, there are currently about 300 model-making companies in total. However, the number of leading companies aiming to benchmark against the US is probably only in the single digits, less than 10. As Mr. Zhou mentioned, different industries don't necessarily use such a large model. Like creating people, there are vocational school graduates, junior college graduates, PhDs, and academicians—different levels serve different purposes. I haven't considered it from a stable perspective because technology is still rapidly iterating. What I hope to see more is that companies actually use the model to improve their services. There's no need to rush; it's only the second year. Within five years, it will be as ubiquitous as chips, used by various companies. What we're creating isn't the old calculator, nor is it a piece of code. We're creating digital employees. What kind of employees does your company need? The more knowledge-intensive, communication-intensive, and analytical employees are, the better they can perform their work. Looking at a five-year timeframe, China will definitely see sufficient development and generate significant productivity.

Pan QingzhongThe concept of digital employees.Teacher Shen, you gave a very good talk about the digital economy right from the start.How can China maintain the development of its digital economy in the future?There is definitely competition. Could you tell me about it?

Shen JianguangA crucial point to consider is why Europe has fallen behind in this global competition. Regardless, China is ahead of Europe. I'm not very technically savvy, and this isn't about the technological gap between China and the US. My chart shows the market capitalization of Chinese tech companies. Look at how many of our tech companies Apple's companies are equivalent to! It's constantly being applied; why has it grown for ten years? The US stock market sees substantial annual profit increases; it has already applied AI technology. The question here is how to further develop the digital economy? I think two points are crucial: how to cultivate an innovative atmosphere and environment throughout society? One that allows for trial and error, allowing entrepreneurs to continuously experiment and try new inventions.

Take Elon Musk, for example. When he started making new energy vehicles, he broke through and disrupted the traditional automotive industry. Later, he went on to build spacecraft—it was completely unstructured. China may pursue a more structured approach to innovation, or capital may be able to enter the market. However, the problem is that many innovative fields, like Musk's example, truly need an environment where they can experiment without boundaries. Only then can innovation occur. That's the first point.

Secondly, as everyone has seen in the big picture, innovation requires substantial capital support. How can we ensure that capital plays a positive role? Why was the gap between China and Europe so small in 2018? I study Europe.

Pan QingzhongAre you referring to a gap in technology or something else?

Shen JianguangI'm not an expert on technology, but I'd like to talk about market capitalization. From a social application perspective, why are all these original innovations concentrated in the US? When we were catching up, we could rely on a national system, but now, in a competitive environment, when you need original innovation, our capital market is crucial. How can our venture capital firms, PE and VC firms, play their roles? In 2018, global PE and VC firms played a huge role in the capital market. The founder of JD.com was an absolute innovator in e-commerce logistics; no other model existed globally. Most of the capital market was skeptical, only one or two firms invested, and it succeeded. In areas like capital innovation, many aspects are unpredictable. In this situation, what kind of capital market should adapt to it? A tolerant attitude and environment from society towards innovation, entrepreneurs, and capital. The capital market is vital. That's why I say this wave of stock market growth has boosted confidence, allowing companies to list, whether overseas or domestically, providing a continuous stream of fresh capital to drive the development of these innovative technology companies.

Pan QingzhongThe core point is that after capital plays its role, China's digital economy can still flourish with the support of technology.

Shen JianguangYes. Including what the others just mentioned about how expensive computing power is, it's all in the tens or hundreds of billions of dollars.

Pan QingzhongEconomist's thoughts, thank you.Next, Academician Sun, you have been engaged in artificial intelligence for more than 40 years, and you have finally seen the light at the end of the tunnel. You have started to study a niche field. AI has been talked about since the 1980s and 1990s, but there have been no application scenarios or breakthroughs.In this process of comparison, what are your thoughts on scientific research in China and the United States, since scientific research is a long-term endeavor? What are your expectations and prospects for the future?

Sun MaosongThis is a difficult question to answer. In China, the general approach is basically to follow the lead of others.

Pan QingzhongCurrently or previously?

Sun MaosongFor decades, we've been more accustomed to following others. Of course, there are objective reasons; others are already ahead, and we have to catch up. But even though our R&D investment has been enormous in recent years, comparable to the US, we're still in the follower stage, only leading in a few areas. In artificial intelligence, for example, we've almost always been a follower, sometimes far behind, sometimes close, but the most innovative ideas don't emerge from our ranks. Even if they do, we can't implement them. I think this is the most profound lesson. As some have mentioned, companies like OpenAI—and not just those, but also Microsoft, Google, Facebook, and OpenAI—were all ahead at the time. Americans are good at seizing historical opportunities. History presents a window of opportunity, and they seize it immediately. What starts as a very small beginning, like Google's founders who were just two students, quickly grew rapidly. It even gives me the feeling that if a company doesn't take off within three years, it might never take off; once it does, it becomes a behemoth.

Pan QingzhongThat's survivor bias.

Shen JianguangIts soil allows those that survive to thrive.

Sun MaosongTo summarize, in terms of scientific research, including both teachers and students, the focus should be on cultivating students. The driving forces of interest, curiosity, and scientific ideals are too weak, which hinders innovation from zero to one. People generally view academic pursuits and scholarship from a worldly, utilitarian perspective, making it difficult to achieve breakthroughs and increasing the risk. We fundamentally lack the necessary preparation, which is the most profound problem facing Chinese education. We are also relatively good at following; AI education is doing well globally, and we still have a group of people who can keep up, which is already quite remarkable. However, to truly lead in uncharted territory, we are currently not ready.

Pan QingzhongThank you. In the long run, education is still the key; we need to go from following to leading. This year, computer science is reportedly the most popular major, which is certainly a positive sign. Having a clear direction is important, but it requires a long-term perspective. What you said about achieving more innovation from scratch is crucial.Mr. Zhou, how far are we from achieving general artificial intelligence?Last time I heard you give a speech that specifically addressed this point. How far do we still have to go to reach this goal?What are AGI's thoughts on this?Because this is also the direction of the future.

Zhou HongyiThere is no standard answer to this question because the definition of AGI is inconsistent. What is AGI? Is it something that is far more powerful than the combined abilities of all humans in the world? Do professional doctoral students emphasize AGI more in their specialized exams? If you use this definition, then this GPT-o1 has already surpassed the math, physics, and chemistry exams taken by American doctoral students.

Pan QingzhongGPT-o1 still needs further development.

Zhou HongyiLater, OpenAI gave a definition of AGI, which ultimately means it's no longer about developing large models, but about how large models can be combined with intelligent agent frameworks, workflow software, and current AI systems. Because ultimately, human capabilities don't rely on individual strength. For example, if five people work together, each person's ability is limited; through teamwork and collaboration, we can create a more powerful result than five people working individually. Altman originally said around 2027, but recently he wrote an article saying it's accelerated, with only a few thousand days left. I calculated that it's still about a decade or so. Let's not get hung up on this now, as if we have to wait for AGI to arrive before we can talk about intelligence. The age of artificial intelligence is already here. There are applications in medicine, classical poetry, security, and search. Through everyone's joint efforts, I think it's more important to integrate current AI technologies into suitable scenarios, making them truly usable, and to help society develop a correct understanding of AI. Otherwise, if we keep promoting AGI and give the impression that it's just for making short videos, creating erotic images, or writing short essays, many entrepreneurs will think these are just toys created by the internet industry—toys, not tools. My view is that with the right methodology, even before we've reached the AGI era, I believe AI has already made huge breakthroughs compared to the applications of the earlier, less sophisticated AI. It can already understand models, understand human knowledge, and do many things.

Pan QingzhongWe need to find application scenarios and frequently talk about the positive role of artificial intelligence.Dr. Ogawa, we've talked a lot today, but we haven't yet discussed what kind of policy support and protection artificial intelligence needs.We'd like to hear your thoughts on aspects such as governance structure.

Wang XiaochuanBecause I come from the internet industry, the previous generation enjoyed a relatively free capital environment and policy regulatory environment, so I accumulated more knowledge in this area, though not much. This time, representing Baichuan, we're using the term "creating doctors." When people hear about creating doctors, they feel it's technically very difficult, wondering if it's even possible. Including AGI, it's a bit like the story of Lord Ye who loved dragons but was terrified of them. If doctors can be created, they'll win Nobel Prizes, surpassing humans. Using creating doctors as a benchmark is already a very advanced deployment technically, even more difficult than self-driving cars. Creating doctors involves dealing with the national health commission system, the medical insurance bureau system, and the hospital system. When you try to implement it within these systems, you realize that medicine is a very conservative industry, a matter of life and death. This requires new pilot tests to verify that it's better than humans. In the early stages, we felt various anxieties, a kind of insecurity stemming from humanity itself, the feeling that they can't be replaced. When I talked about AlphaGo in 2016, saying that machines could defeat Lee Sedol, Yu Bin collapsed, and the technology industry wept. When AI arrives and is replaced, it will bring a painful backlash. The obstacles could come from the doctors themselves, or from officials; success is a performance indicator, failure means losing one's job. Whether it's due to human nature or the resistance brought about by policy directives, we believe that while we are still in the conceptual stage, we need to raise awareness and address the practical difficulties that arise. Firstly, it's not an ethical issue; privacy issues are discussed extensively, but we must get started. If we want to create an environment conducive to capital investment and experimentation, like in the internet age, this requires more communication with the government. Currently, it's proceeding in an orderly manner, and we hope this forum will foster more discussion, bringing the issues to light for resolution.

Zhou HongyiYour words prove my point. I think it's easier to create a doctor than an AGI. He's also working on a large, professional model. If we keep having this expectation that OpenAI can lead us astray, thinking that a large model can do everything—treat patients, write poetry, answer riddles, solve business problems, and even create artificial intelligence—that's illogical.

Does the whole world use one big model to solve problems? AGI is a pseudo-concept. If it is defined as professional people recognition, then Jiu Ge's poetry writing level is already higher than that of modern people. The concept of AGI should be classified from a professional perspective.

Wang Xiaochuan: In the US, they're watching OpenAI closely, but there's also SIA, Musk, and OpenAI. OpenAI is the most dangerous company in this field, not because it's disruptive to humanity, but because it might be overshadowed by Obama's emphasis on the existence of its own market and definitions within finance. The brilliance of OpenAI might be preventing us from seeing what other excellent American companies are doing, and it could potentially lead us astray. Like you said, don't let OpenAI lead you astray; follow Mr. Zhou's lead, not OpenAI's.

Zhou HongyiFinally, let me give you an example. Tesla's FSD is about to make a breakthrough. Huawei's technology roadmap in China is to achieve end-to-end autonomous driving with a large model. There are many things that cannot be done, but by training a large model with 1 billion kilometers of driving experience, which is richer than the driving experience of any single person on Earth, you could say that it is a professional driver. Would you say that it counts as AGI? I think this question today is very good. The concept of AGI needs to be redefined.

Pan QingzhongThank you. The discussion was very in-depth. Professor Shen, could you briefly describe AI in a sentence or two? Artificial intelligence can improve efficiency; how exactly does it improve efficiency?

Shen JianguangFor example, our intelligent customer service can help provide better service and improve efficiency. Of course, an even better example is programming; with the assistance of AI, programming can significantly improve efficiency. These are some of the applications our company is using, including unmanned transportation using artificial intelligence, drones, and warehousing. These are already in use.

Pan Qingzhong：Thank you so much.Finally, Professor Sun, I originally wanted to ask you a question about AI and artificial intelligence. You've written so many poems, more than 40,000 of them, including those by Emperor Qianlong.Artificial intelligence can have some impact on Chinese culture. You can answer or not answer, choose something you want to say or summarize it in a few words.

Sun MaosongThe culture is too complex to discuss. What I'll say at the end is that building AGI on large-scale models is a problem, but in China, we might be able to catch up and surpass the US. This is likely due to what I'll discuss later: integrating AI into various disciplines, making it a "AI for Science" approach. The Americans started a little earlier, but we have the data. If Xiaochuan's company were to do this, they definitely wouldn't have a problem; this is a significant opportunity. Including if we turn financial data into a comprehensive engineering project, it wouldn't be an exaggeration to say we could produce a Nobel Prize in Economics every two years.

Pan QingzhongThe next one is Professor Shen.

Shen JianguangAs Mr. Zhou just mentioned, we should apply what we learn to our academic disciplines and into industrial applications.China has an advantage, and our large domestic model manufacturers are more than capable of handling this.However, this does not mean that all fields and all disciplines are suitable for large models. Therefore, it is necessary to find a place that is suitable for the characteristics of large models to do these things. If a suitable place is found, it must also require a large model.This still requires arduous intellectual work.Large-scale models are not created overnight, nor are they a magic bullet; they certainly require a lot of hard work and are not easy to achieve.If you keep going, you will surely overcome all obstacles.

Thank you all very much. Let's give a warm round of applause to our four distinguished guests for their wonderful speeches and remarks. Thank you very much.