Exclusive Interview with Professor Yu Yang from the School of Artificial Intelligence, Nanjing University: Artificial Intelligence Is a Path to the New Configuration of Human Society in the Future
Artificial Intelligence has become a crucial driving force behind the new round of technological revolution and industrial transformation, and large models are like a new power system for information technology. Artificial intelligence technologies represented by "ChatGPT" have set off a new upsurge worldwide, triggering a new wave of AI applications. Yu Yang, Professor at the School of Artificial Intelligence of Nanjing University and Founder of Nanjing QuestAI Technology Co., Ltd., recently accepted an exclusive interview with Nanjing Innovation Investment Group, sharing his views on the technology behind GPT, potential new applications that may emerge, and investment opportunities in the AI field at the current stage.
Yu Yang is a Professor at the School of Artificial Intelligence of Nanjing University and Founder of Nanjing QuestAI Technology Co., Ltd. He mainly engages in research on reinforcement learning, and his work has won 5 international paper awards and 3 international algorithm competition championships. He was selected into the National Young Talent Program, named one of IEEE's "Top 10 Rising Stars in Artificial Intelligence", and awarded the CCF-IEEE Young Scientist Award and the first Asia-Pacific Data Mining "Young Achievement Award". He was also invited to deliver a "Young Highlights Talk" at the International Joint Conference on Artificial Intelligence (IJCAI) 2018.
On November 30 last year, OpenAI, a U.S.-based AI R&D company, launched ChatGPT, an AI chatbot program, which opened a window for interaction and dialogue between humans and robots. It is regarded by all sectors as an important milestone for humanity to enter the AI era. Could you briefly sort out the relationship between artificial intelligence, machine learning, and large models?
Artificial intelligence (AI) is a broad field. According to the definition by John McCarthy, the "father of artificial intelligence", AI is the science and engineering of making intelligent machines. McCarthy also emphasized that AI is related to understanding human intelligence, but it is not necessary to realize intelligence in a biological way.
The classic content of AI includes search and reasoning, knowledge representation, and machine learning. In terms of the number of participants, published papers, and application implementation, machine learning has now become the largest sub-field of AI in scale. The rise of machine learning stems from people's demand for summarizing knowledge from data.
In the process of machine learning, data serves as the input raw material. After processing by algorithms, a model capable of analyzing and predicting new data is finally generated. For example, in face recognition scenarios, the model receives image input and outputs corresponding recognition results. As for the so-called "large models", there is no unified definition standard, but they generally refer to models that are trained on extensive data, have a large number of parameters, and can handle multiple types of tasks. ChatGPT is a typical example: it is trained using large-scale internet text data to support its execution of diverse tasks, demonstrating the wide application capabilities of large models.
What makes ChatGPT so impressive?
The reason ChatGPT is highly praised lies in two key innovations. First is the breadth of its training. ChatGPT is trained using unprecedentedly large-scale internet text data, covering almost all available information on the internet. This endows the model with a massive knowledge base and diverse understanding capabilities. Of course, behind the successful training on such large-scale data, appropriate algorithm and model design, as well as strong engineering capabilities, are indispensable.
Second, OpenAI has made a breakthrough discovery in the application model of large-scale language models. This discovery can be regarded as a major advancement compared to traditional machine learning applications. In the traditional paradigm, the training and application of a model are consistent—for instance, a model trained for face recognition will mainly be used for recognizing faces in images. However, OpenAI found that by training a model to predict the next word in a text sequence, a powerful tool can be created. This tool can not only answer questions but also generate text with specific styles or content according to instructions. This discovery has greatly expanded the application scope of language models, making their applicability far beyond traditional boundaries and bringing people an impressive experience.
The core of GPT lies in that machines have mastered language—they have become machines that understand and are good at expression, while also possessing common sense about the world. They have mastered events that have occurred around the world, knowledge, common sense, and even simple reasoning. This is an extraordinary breakthrough. Specifically in terms of product forms, what super applications may emerge?
In my understanding, the core role of large language models is to innovate human-computer interaction capabilities and methods. This transformation is not limited to achieving a more natural dialogue experience; it also changes the fundamental way we interact with technology. With the help of large language models, machines are no longer just tools that simply execute fixed commands or respond to preset instructions. Instead, they can understand complex language expressions, and even vague human intentions and emotions.
I cannot predict exactly what super applications will emerge based on large language models, but it is expected that such applications will definitely allow people to communicate with machines in their most natural and intuitive way—through language. Regardless of their technical background, age, or cultural background, people will be able to communicate effectively with machines without learning complex interfaces or command languages. This will enable machines to better integrate into our society and daily lives.
In the previous AI era, some AI products and commercially successful companies have emerged. What will be different in the GPT era?
No matter how vigorous the wave of large language models is, the success of products and businesses will always depend on meeting people's needs. For previously clear needs, the integration of large language models can significantly improve efficiency. For example, Microsoft's Copilot series is enhancing existing applications such as Github programming, Word document processing, and PPT design. The application of large language models in this type of demand is likely to achieve rapid popularization.
Large language models may also help develop new needs—such as chatbots for emotional comfort. However, developing new needs always requires a relatively long acceptance cycle.
In the field of reinforcement learning that you focus on, what new cutting-edge progress has been made recently?
Reinforcement learning is a branch of AI technology for building intelligent decision-making systems. The main recent progress in reinforcement learning, in my view, is that it can move beyond the world of games and be truly applied to solve decision-making problems in various industries. Based on these advancements, Nanjing QuestAI Technology Co., Ltd. has become a leading enterprise in the application of reinforcement learning. It focuses on industry and manufacturing, providing core technologies for cutting-edge decision-making and control required by advanced manufacturing.
Since 2016, the previous round of AI commercialization has encountered some problems—AI companies have had to undertake many tedious and fragmented customized projects. How can large models avoid the problems encountered in the previous round of AI commercialization?
Customization is about meeting the personalized needs of different customers. The fundamental reason why many companies have to invest a lot of time and resources in customized projects during the process of commercializing AI technology is, in my opinion, the lack of versatility of the AI technology applied. Such technologies can usually only solve certain links in the application chain, rather than handling all links. Especially in terms of understanding and modeling customer needs, it has been difficult to handle these tasks through AI technology in the past. As a result, manual work is required to adapt to different needs, increasing the pressure of customized delivery.
Once AI technology itself can handle demand understanding and business modeling, the efficiency of its commercialization will be significantly improved. In the application of reinforcement learning, Nanjing QuestAI Technology Co., Ltd. is precisely using AI technology to handle business modeling. The enhanced ability of large language models to understand customers' language descriptions can also alleviate the pressure of manual adaptation to a certain extent. Therefore, we will witness the rapid popularization of large language models.
What is the difference between large models and small models? Where are the opportunities for small models? Large companies have inherent advantages in capital reserves, talent resources, application scenarios, and data accumulation. Do small companies have opportunities? If so, where are the opportunities?
The "large" we usually refer to is an intuitive description of "foundation models". When a model needs to "absorb" a large amount of data, its parameter scale usually also needs to be large—thus giving rise to "large models". For large models, our main expectation is that they are highly versatile. However, versatility and personalization often conflict, and this conflict creates room for industry enterprises to provide differentiated services.
In addition, OpenAI had only 200 employees in 2021 and now has fewer than 800 employees—not a "large company" in terms of headcount. When disruptive technologies emerge, large companies that respond slowly may face more inherent disadvantages. Small companies with core technologies will always have opportunities.
AI large models are currently a hot topic. If we want to make investment layouts for the new generation of artificial intelligence, what are the relatively certain investment opportunities at the current stage?
In my view, artificial intelligence is not just a technology, but a path to the new configuration of human society in the future. There may be many uncertainties in specific businesses and projects, but this path leading to the future should be quite certain.
AI chips, AI servers, and storage are all important components of AI infrastructure. The intensified competition in AI hardware has led to a sharp rise in chip prices, making chips the largest and most strategically significant growth point for chip manufacturers. Driven by multiple factors such as the booming AI market, premium prices of NVIDIA graphics cards, insufficient production capacity, and chip bans, the construction and investment of domestic AI infrastructure seem to have ushered in a new opportunity. How can domestic AI infrastructure enterprises seize this opportunity to catch up? Moreover, what potential investment opportunities are there in the infrastructure of the AI era?
Based on the information I have seen so far, domestic AI hardware enterprises are making every effort to meet the market's computing power demand. Beyond process and production capacity, one point that the general public may not easily understand is that the competitive shortcoming of domestic AI hardware mainly lies in the software ecosystem. The recent ban is a rare opportunity to improve the software ecosystem.
Another very important "infrastructure" is basic scientific research. AI is a rapidly iterating field—the best technology is always the next one. The ability in cutting-edge basic research determines the capability for sustainable development and self-disruption.
Thank you for Professor Yu's sharing. It is hoped that general artificial intelligence can free people from low-level and tedious tasks and open a new era full of creativity.
Source: Yu Lingyu (First Investment Department)
Reviewer: Xue Yao
Publisher: You Yi
