Exclusive Interview With Huang Tiejun: In The Era Of Big Models, How Does Artificial Intelligence Connect Technology And Industry?

Since the development of artificial intelligence, it has become the core driving force for a new round of global technological revolution and industrial transformation. Currently, the use of big data to increase computing power and strengthen algorithms to form ultra-large-scale intellectual models has become the core of the new generation of artificial intelligence ecology. It will be a major application infrastructure for the development of artificial intelligence in my country and a key to realizing my country's overall leading strategy for artificial intelligence in 2030. Basic Platform.
The first conference of the Boao Forum for Asia Global Economic Development and Security Forum, with the theme of "Economic Security and Sustainable Development under Great Change", will be held in Changsha, Hunan from October 18th to 20th. It aims to discuss responses to global economic risks and Structural issues will be discussed together on issues such as intelligent manufacturing and technological revolution.

Regarding topics such as the development direction and application prospects of artificial intelligence in the era of large models, a reporter from Beijing Business Daily recently conducted an exclusive interview with Huang Tiejun, professor at the School of Information Science and Technology at Peking University and director of the Department of Computer Science and Technology.

Beijing Business Daily: Large models are called the core of the new generation of artificial intelligence ecosystem. Can you briefly introduce what a large model is? What problems will be solved?

Huang Tiejun: The large model is a basic common model that imparts intelligence to various applications. In general, it is a basic platform that learns the knowledge and rules contained in the data from massive big data, condenses it into a neural network and turns it into a large model, and provides services for various general intelligent tasks.

For example, on the mobile Internet, cloud service vendors can have many service capabilities, but without a carrier like App, it will be difficult for users to obtain various cloud services. From this point of view, App itself is an industrial ecosystem. In fact, large models currently need to solve similar problems.

Large models are public services that are highly radiant and highly technical. All walks of life will have some specific needs in the future, and some companies will need to develop large model conversion and customized interfaces.

Beijing Business Daily: How will large models connect the artificial intelligence technology ecology and industrial ecology? What will be the next application direction of artificial intelligence in the information field?

Huang Tiejun: The understanding and application of artificial intelligence in many industries are still in an exploratory stage, and there is a certain distance between them. How to connect this interface actually requires a group of companies that can transform the capabilities of large models into content needed by various industries.

It is very difficult to predict what the next application in the information field will be. I think in reality, copywriting, information processing and other tasks will be replaced by artificial intelligence, or most of them will be solved by large models of artificial intelligence, which will bring about great application possibilities.

The various applications of search engines are ultimately about the organization, mining and use of information. For example, individuals can collect data and do some information processing through search engines. Now large models solve the problem of collecting massive data. Its data is not the work of any one person or group of people, but collects all the data and reflects it. Come to serve various copywriting information processing applications. The final exit may still be human, but most of the tasks behind it will be completed by artificial intelligence. The possibilities for this application direction are huge.

Beijing Business Daily: How did artificial intelligence technology develop into the era of large models? What are the differences compared to before?

Huang Tiejun: The development of artificial intelligence into large models is determined by the basic laws of the development of artificial intelligence technology. There are two schools of thought on artificial intelligence. One group believes that the scientific mechanisms, theories, mathematics, and algorithms behind artificial intelligence are very important; the other group believes that artificial intelligence is generally a technology, constructing an intelligent system, and then understanding the mechanism of the intelligent system. The latter is the mainstream view of artificial intelligence.

In the process of building artificial intelligence, it was the work of a few scientific researchers at first, and later companies gradually participated. In the future, industry, academia, research, and the whole society will jointly construct a model. Why do this? In fact, the reason is very simple. If the data learned by an artificial intelligence system or model is not complete and timely enough, it will be difficult to believe that its intelligent model is very capable. The so-called large model is to integrate various data resources, the strongest algorithms and computing power possible in society into a public basic platform that everyone can use. This is the direction we must take when constructing artificial intelligence systems. .

In this process, the capabilities of large models become stronger, which in turn will play a role in all aspects of society. Once it plays a role, more people will build it. It is a benign iterative effect.

In fact, the potential of artificial intelligence depends on the data it can obtain. Just like people "read thousands of books and travel thousands of miles", artificial intelligence is also similar. The physical world and even the universe are so vast. If it can be converted into data and information and allowed to be learned by artificial intelligence, the space will be very large.

I dare not say whether the ability of the large model will exceed that of everyone, but at least no one has obtained all the information, and it is impossible to discover the laws behind it. The physical body and life cycle of each of us determine that the data we can obtain is still relatively limited.

Beijing Business Daily: What challenges still need to be overcome in the development of large models? What is the future development path?

Huang Tiejun: There is now a set of technologies and algorithms for training large models, but whether there are better algorithms, the academic community and the industry are still constantly searching and exploring. Currently, it requires a lot of carbon emissions to train an intelligent model. In the future, it may require less carbon emissions to train a model. I think one day, artificial intelligence may cost less to train than a human, and that's another milestone.

Therefore, as artificial intelligence comes into contact with more and more data, and as learning and training efficiency becomes higher and higher, its result will be a large model. In the future, it may be a super large model or a very large model, and it will continue to iterate. The path is already clear.

But there is no basis for where its upper limit is. For now, bigger is better. It may be that after it expands to a certain level, it will no longer be a simple linear increase, or it may be that after a certain level, the growth begins to slow down, but these are still speculations at the moment.

Beijing Business Daily: How to consider some safety and ethical issues during the development process of large models? How to avoid it?

Huang Tiejun: The safety and ethical issues of artificial intelligence will not be solved overnight. For example, information security issues continue to arise with the development of information, and we must solve them after they arise.

In the development process of large models, there are also some inherent risks. For example, the knowledge learned by the model is not in line with ethics and principles. These risks can be controlled in advance; but there are also some risks caused by the continuous advancement of technology. If it comes, then its solution also needs to be solved continuously through technical means. "To untie the bell, you must tie the bell." If we do not develop this technology because of some potential problems, it will not comply with the laws of scientific and technological development.