Interviewer: Sun Baohong, Dean’s Distinguished Chair Professor of Marketing at CKGSB

Professor Sun Baohong is currently Dean’s Distinguished Chair Professor of Marketing and Head of the Web3+AI Research Center at CKGSB

US-China Technology Frontier Series

This Q&A is part of a series of articles developed by the AI+Web3 Research Center of CKGSB alongside industry experts on the US-China technology frontiers. The series aims to introduce the most cutting-edge theories and practices at home and abroad on the theme of digital technology and internationalization, open up the mindset of students, discuss new business opportunities brought about by the development of science and technology, and identify new opportunities for the technologization and internationalization of Chinese enterprises.

Artificial intelligence (AI) has gradually demonstrated ever-greater capabilities, and has already surpassed individual humans in terms of knowledge accumulation and coverage. And the AI wave is only just beginning. There are a number of key innovation directions open to exploration, including generative AI, convolutional neural networks and robotics.

In addition to its many generative functions, AI also raises questions related to such cutting-edge trends as human social simulation, brain-computer interface technology and AI doppelgängers, as these technologies will not only change the operational approaches of enterprises, but also profoundly affect the structure and development of society.

Chinese technology companies in particular are using AI and other technologies to address global competition and create better enterprise applications. But there is an importance in developing “expert models” that are better suited to enterprise applications and can provide more accurate solutions.

Q. Does AI possess intelligence and to what degree does it have advantages over humans?

A. There is a close relationship between knowledge and intelligence. Knowledge is what we acquire through learning, experience and the accumulation of information and facts. Intelligence, on the other hand, is the in-depth understanding and insight that comes from summarizing, reasoning and analyzing on the basis of this knowledge. Nowadays, AI demonstrates a similar process. In training an AI model, we first need to input a large amount of knowledge and data. When this knowledge accumulates to a certain level, an “emergence” of intelligence is possible.

It wasn’t until November 2022, with the release of GPT-3.5, that OpenAI first observed the emergence of what was characterised as intelligence in AI. This emergence has similarities with the human learning process, as we also gradually generate intelligence after accumulating enough knowledge. GPT-4 was released in March 2023, reaching 1.76 trillion parameters, and was arguably the most advanced AI model at the time. Human cognitive connections total about 100 trillion, so AI is still a little way behind, but is on its way.

Through testing, GPT-4’s intelligence level has reached an IQ of 80-90, however, in terms of knowledge coverage, GPT-4 is far beyond individual humans. The knowledge of individual humans is usually narrow and only covers specific fields, while that of GPT-4 covers almost all the knowledge of human civilization since its development. Moreover, GPT-4 does not need to rest and sleep like human beings, it is able to learn continuously, and its speed of acquiring knowledge is more than a hundred times that of human beings. By 2024, the latest GPT-4o model had reached an IQ of 120.

Now, most of the tasks that GPT-4 can accomplish are almost indistinguishable from human abilities. For example, in terms of visual performance, GPT-4’s ability to recognize species and to recognize faces has surpassed that of humans. In addition, deep AI technology can generate a variety of languages, paintings, artworks and programs that cover almost every aspect of human creativity.

This makes us think: where will humanity go when faced with such rapidly developing AI?

Q: What are some of the main directions in which AI is currently being developed?

A: In the current field of AI, the mainstream technology direction is primarily divided into two, namely: Convolutional Neural Network (CNN) and Generative AI, each with its own strengths. Convolutional Neural Network is mainly used in the fields of image, face and speech recognition, using layers of convolution and pooling operations, image and speech data can be processed and recognized.

Generative AI is based on different types of models that can generate brand new content through complex algorithms and deep learning. It is stochastic and creative, and is capable of “creating something out of nothing,” producing unexpected results and broadening the boundaries of human thinking.

The currently widely used ChatGPT is a key representative of generative AI. It is largely based on the Transformer model, which was first introduced in 2017 and has since been adopted and continuously optimized by OpenAI. In the process, OpenAI collected and processed a large amount of data in areas with low labor costs, such as Kenya, and through training for up to five years, the model eventually began to show intelligence. Once intelligence emerges, it is like opening Pandora’s box, progressing rapidly and uncontrollably.

Generative AI has a wide range of applications, including in areas such as copywriting, report generation, drawing, video production and even writing programs. For example, in the financial industry, generative AI is powerful in data analysis and financial statement processing, capable of completing complex analytical tasks quickly and accurately, with far-reaching impacts on Wall Street and the entire financial industry.

Currently, another area of great interest in the research direction of artificial intelligence is the simulation of human society. This direction aims to explore the formation and development of human society, trying to understand how complex social structures such as interpersonal relationships, social hierarchies, religions, classes, parties, etc. are gradually formed through simulation.

Nowadays, researchers can try to use artificial intelligence to simulate the formation of human societies and to predict the trends of future societies. Such simulations can help people to foresee possible consequences before making major decisions. For example, at the national level, when deciding on major policies, simulations can be used to assess the potential impacts of different decisions. In the past, it was very difficult to conduct simulations on such a large scale; now, with the help of artificial intelligence, this area of research has become more feasible and full of potential.

Robotics is also an important development. Although artificial intelligence has demonstrated strong computational and analytical capabilities, its actual ability to act still needs to be realized through robots. If we think of AI as the brain, robots give this “brain” the ability to act. AI doppelganger technology is also a compelling area. In the future, there may not only be one “you,” but perhaps another “you” in the virtual world. This raises a deeper question: are we really the highest intelligence in the universe? Or are we just manipulated non-playable characters (NPCs), as in The Matrix?

As esoteric as these ideas may sound, advances in artificial intelligence have forced us to rethink how we do things. In addition to Internet and enterprise applications, the application of AI in the biomedical field is also remarkable. For example, the AlphaFold model introduced by Google was able to quickly unravel about 200 million protein structures, solving the problem of protein structure prediction. AlphaFold won the 2024 Nobel Prize in Chemistry.

MIT has developed a drug selection system that utilized AI to screen new antibiotics. The methodology used by AI in the drug selection process is completely different from the traditional methods used by human beings, and the antibiotics selected in the end had remarkable effects. This shows the extraordinary potential and uniqueness of AI in certain areas, beyond human abilities. Although AI sometimes brings unexpected results, its innovative ability and potential value cannot be ignored. In the future, we should further embrace and utilize AI to meet various complex challenges and needs.

Q: Robotics is an important development direction for AI. Can you talk in detail about the opportunities in China?

A. I think robots give AI the ability to actually act, going from just being able to think, to actually able to perform tasks in the physical world. Currently, robots used in traditional industry are commonplace, such as in warehouses, docks and on production lines. However, the future of robotics will be much smarter, especially when it comes to home robots, and I think these will be important developments.

As the global population ages, the demand for robots for the elderly will grow. Future household robots will not only need to have a high level of intelligence, but will also need to be able to adapt to a variety of different environments, which is very different from the way existing industrial robots work in fixed environments. These new robots will no longer be limited to simple, repetitive tasks, but will be able to operate independently in complex and changing scenarios, such as caregiving. For example, Tesla plans to launch home robots priced at less than $20,000, suggesting that future consumer trends may shift from the traditional show of wealth—comparing who has a nice car and a big house—to comparing who has more and smarter robots.

Currently, there are a number of open-source robotics projects that are worth looking at. These projects allow ordinary people to build robots at home. For example, one such project includes robots that can stir-fry and cook at home, and although it does not have a high success rate at the moment, ranging from about 20-30%, it shows us what is possible in the future. By refining and optimizing these projects, it is entirely possible for Chinese companies to further develop technologies to better serve the home market.

Another important development in robotics is the brain-computer interface. Brain-computer interface technology makes it possible for humans to control machines with their thoughts by connecting the brain to a computer. Elon Musk’s Neuralink project is a prime example of this, with a chip implanted in the brain that reads brain waves and helps to enable movement. This technology could help people who are unable to move due to physical impairments to regain mobility through robotic arms or legs. In addition, brain-computer interface technology could help blind people regain their vision. Although this sector is still quite crude, the technology will mature.

Brain-computer interfacing has another valuable application, namely, to provide a “second brain” for human beings. Currently, our brain has limitations in terms of memory and information processing capabilities, and through brain-computer interfaces, human beings are expected to obtain an external “super brain” as a “plug-in” to help people answer complex questions. In the future, AI may be directly connected to the human brain in this way to realize more efficient information processing and decision support.

Q: What are the benefits to developing enterprise-specific expert models for organizations?

A. In China, there is also huge potential for growth in the field of AI in the form of enterprise-specific expert models. With generic models such as OpenAI GPT-4, we have found that while broad in their capabilities, they do not always provide accurate results that meet specific needs. As a result, many organizations have reservations about its practical application. For example, some firms have banned the use of OpenAI’s ChatGPT-4 because they believe that corporate data is valuable and confidential and should not be easily shared with third parties.

I think the value of generalized models in business applications, especially for enterprise-level use, has yet to be proven. In contrast, I prefer enterprise-specific expert models, which focus on specific domains and can provide more targeted and precise solutions. For example, Huawei has developed a system dedicated to weather forecasting, which is designed to meet the needs of enterprises that require special expertise in that particular area. I think the market potential for such expert models is huge.

Multimodal modeling is also an important direction in the latest development of AI. Multimodal modeling means that AI can not only process text, but also understand and generate content in multiple forms, such as video, images, music and so on. This makes AI increasingly human-like, able to access and process information through multiple senses. However, for enterprise applications, I think there is another type of model that may have more potential than multimodal models, namely a collection of expert models, a concept known as “Mixture of Experts” (MoE).

The core idea of the MoE model is to outperform a single large model, such as GPT-4, through the collaborative work of multiple smaller expert models on a specific outcome. Recently, a company called Mystery achieved this goal by using seven small models with 700 million parameters to outperform GPT-3.5, with its 175 billion parameters. Mystery plans to launch a program that consists of a dozen small models, each focusing on a different task area, which is expected to outperform GPT-4 in certain specific tasks.

In 2024, to further enhance the performance of Large Language Models (LLMs), a team from Duke University, Stanford University and Together AI proposed the innovative “Mixture of Agents” (MoA) approach. The system uses multiple layers—the key innovation—each containing a number of individual smart “agents.” Each agent takes input from all agents in the previous layer to generate its response.

This design allows the system to take advantage of the diversity of models, mobilizing a wider range of capabilities than a single model by using multiple agents with different strengths. The multi-layered approach also allows for iterative optimization, where outputs are progressively improved as they pass through multiple stages. MoA opens up the possibility of creating a more powerful and comprehensive AI assistant by combining the strengths of multiple models.

Additionally, it suggests a path to resource optimization, whereby integration using small open source models may reduce reliance on large proprietary systems while producing high performance levels. Organizations of all types may be able to flexibly combine different LLMs to create customized AI systems based on their specific needs. From a research perspective, the MoA framework provides a new avenue for exploring how different AI models complement each other.

This area is of major interest to me because demand for computing power has always been a key limiting factor in the development of AI. Especially in China, the shortage of computing resources limits the training and application of large-scale generalized models. If we can use the limited computing power to train multiple small expert models/intelligent agents and combine them together, this approach has the potential to surpass a single powerful opponent through teamwork.

I’m also interested in longitudinal expert models, especially those that focus on enterprise-specific data. In the US, these models are called “vertical models,” and they focus on a particular industry or domain-specific data set. Every organization has unique data, and this data is its most valuable asset. If an organization can use its own data and industry data to train expert models, these models may outperform large general-purpose models like GPT-4 on specific tasks.

The potential of such expert models lies in their ability to provide more accurate and specialized solutions, which is crucial for many businesses. By combining enterprise-specific data, expert models can generate more precise and targeted results, which not only improves efficiency, but also helps companies gain an edge over their competitors. Therefore, I believe that the application of expert models has a very bright future, especially in markets like China, where the use of limited computing resources to create targeted AI solutions is promising.

Q: Will AI one day replace humans?

A. In reality, AI has been shown to be a powerful aid in many areas. However, as powerful as AI may seem, the results it generates do not always match human expectations. For example, AI tends to give very general or long-winded answers, but these answers often fail to solve specific problems. As a result, AI can only serve as an assistant in some cases, rather than replacing the need for humans altogether.

However, as AI evolves, this assistant role may evolve into a more companionable presence. AI assistants may get to know us better than any of our friends. Not only will they be able to help us in our work, but they may also be able to establish deep connections with us emotionally. This emotional accompaniment and increased understanding gives AI huge potential for use in the future.

Whether it is robotics or brain-computer interface technology, or the development of multimodal models, today’s AI is gradually moving from a single function to diversification and intelligence. In the future, these technologies are bound to profoundly change the way we live and work, and we need to actively embrace these changes and fully utilize the opportunities that AI brings.