Compared with the invention, popularization and use of the Internet, artificial intelligence has a greater impact on human society. In the era of artificial intelligence, the most important thing for human beings is imagination.
Xing Li is a professor at the Department of Electronic Engineering, Tsinghua University
More than 30 years ago, Li Xing of the Department of Electronic Engineering of Tsinghua University graduated from the undergraduate and went to the United States to study, and his graduate major was the same as that of undergraduate, which was signal and information processing. During his study abroad and work, he came into contact with the Internet for the first time, and as a user, he deeply felt the great potential of the Internet and developed a certain understanding and thinking about it.
In 1991, Li Xing returned to teach at Tsinghua University and deeply felt that without the Internet, it is impossible to build a world-class university. By chance, in 1993, he changed from an Internet user to an Internet researcher, and became the designer, builder and operator of the China Education and Research Computer Network (CERNET). In the face of the major technical decisions encountered in the early days of CERNET's construction, his understanding and thinking accumulated as an Internet user played a crucial role. Thirty years of reincarnation, now as a user of artificial intelligence, he shares his understanding and thinking about ChatGPT for the reference of the younger generation.
Mind games with ChatGPT
"ChatGPT is about helping people, not replacing them." Professor Li Xing said. ChatGPT has played a big role in helping programmers write programs, and while it still can't completely replace human programming, it can bring us a lot of unexpected inspiration and improve productivity.
In the past, when programmers needed to implement an auxiliary function algorithm or function, because these functions were not the most core, they were not necessarily familiar to programmers, and it often took a long time to retrieve, find and judge. However, after discussing with ChatGPT, it can quickly provide relevant code, language version, operating system and other information, which is convenient for us to make the next step of judgment, selection and modification.
In addition, ChatGPT is also very significant for improving the efficiency of scientific research. Professor Li Xing mentioned that in the past, after he generated an idea, he would often put it aside, think deeply for a period of time, and then spend a few days or even weeks writing code to verify it. However, with the help of ChatGPT, it only takes a few hours to complete the initial proof of concept (PoC, full name of Proof of Concept), and then move forward with the next step based on it, which is as efficient, simple and convenient as a boat on the water.
Of course, it should be noted that although ChatGPT can give answers, many times these answers are not accurate and cannot be adopted directly. So, in this case, you need to continue to interact with ChatGPT and help it give better quality content with guidelines.
ChatGPT's training data contains a wealth of diverse information, covering almost all fields. The key, however, is how we ask useful and highly relevant questions. In other words, the ability to ask questions is very important. Whether you can ask quality questions is very challenging.
Similarly, thanks to ChatGPT's fairly comprehensive (but not necessarily in-depth and accurate) level of knowledge, it tends to give more comprehensive answers than we think, helping us fill in the gaps in our knowledge. So when we generate some innovative ideas, we can have a conversation with ChatGPT, go through iterative intellectual games with it, and ask it what it thinks about new research ideas. ChatGPT also unreservedly lists relevant issues and angles to consider, and this comprehensiveness is an advantage that cannot be matched by individual human individuals.
In addition, no matter how novel, peculiar or deviant the questions asked, ChatGPT always provides responses and answers related to it, thanks to ChatGPT's generalizability to generalize.
According to reports, one of the important factors that OpenAI considers when training ChatGPT is the diversity of data. The first is "linguistic diversity". Although the data were mainly in English, content in Chinese and other national languages was also incorporated into the training as much as possible. That is, if there is a question in the Chinese world that has not been answered, but someone in the English-speaking world has already given the answer, ChatGPT can automatically link the two and flexibly apply knowledge in various languages and cultural contexts when needed.
The second is "diversity of fields". ChatGPT is a polymath rather than a specialist, and it incorporates knowledge from a variety of subject areas, allowing for a balance and connection between disciplines for true integration. This means that sometimes ChatGPT can find solutions to problems in one discipline and provide us with clues to solve them. The magnitude of ChatGPT's knowledge allows it to truly achieve integration and touch, and the knowledge of various disciplines is no longer clearly separated but intertwined. In addition, the training of ChatGPT also takes into account the "diversity of speech" and "diversity of views", and the generalization ability and "thinking" method brought by these diversity are far from the traditional search engine question and answer mode.
The unexplainability of ChatGPT
One current line of thinking is that AI needs to be "explainable." Taking autonomous driving as an example, if the principle of autonomous driving is not explainable, then people cannot analyze the causes of traffic accidents, so it is difficult to improve and enhance the safety of autonomous driving. Therefore, for this type of science and technology, explainability is very important.
However, is explainability still important for technology applications like ChatGPT? Professor Li Xing believes that there must be a question mark on this.
He mentioned that although ChatGPT also has its own understandable scientific theory behind it, its powerful performance is rooted in massive data and powerful computing power. As the saying goes, "a miracle comes out", and it is based on this huge computing power and data support that ChatGPT has demonstrated intelligent behavior that has so far only appeared in humans. Although we know that ChatGPT is obtained through training, the process of training an AI model, that is, the optimization of model parameters, is sometimes more like the "alchemy" process in metaphysics. How to adjust parameters? Why is this tuned? Until now, there is no very precise and complete theory and method, and this randomness and experimentation is the source of ChatGPT's uninterpretability.
Therefore, Professor Li Xing believes that according to the traditional thinking, we should first have a deep understanding of man-made things, and then continue to study and apply. However, ChatGPT breaks this traditional path. With ChatGPT, we can think of it as a product created by God or nature itself, and we first accept, use, and operate it, and then further study and understand it, which is the exact opposite of the process of understanding laws before applying them in past technological inventions.
Based on this, Professor Li Xing believes that the true vitality and soul of general artificial intelligence lies in its unexplainability.
Looking back at the history of science and technology, the invention of the computer is a clear dividing line. Before the advent of computers, machines invented such as telephones, washing machines, refrigerators, radios, etc., were machines with single, clear and fixed functions. However, the computer is completely different, although it is also composed of fixed components such as CPU, memory, hard disk, monitor, keyboard, etc., it can install a variety of different software, write different types of programs, each software has different functions and functions. The use of computers is so rich and diverse, and the results are so diverse, that the people who designed them could not have imagined in the first place. Over time, more functions and ideas will be developed, and the space for computer development will remain vast and infinite.
After the emergence of artificial intelligence, the uncertainty of this function has reached a new level.
In internet games, people can create video games with various mechanisms, types, and rules by programming computers. However, AI can create its own game and decide all the settings to make it run smoothly, making the principles and mechanisms behind it potentially beyond human comprehension.
The unexplainable way to crack
"In the face of unexplainability, we must accept this fact, accept its existence, and live with it." Professor Li Xing said that while accepting, we should strive to study and understand the laws behind artificial intelligence.
He believes that in our existing discipline system and knowledge structure, physics and mathematics have strong explainability, they are composed of extremely accurate mathematical formulas and physical laws, so we can try to describe it with exact formulas or clear models. Although biology is also a discipline that explains the world, its clarity and accuracy are far lower than physics, and the mechanisms and laws behind many biological phenomena are still unable to explain and describe clearly, accurately and rigorously, that is, different disciplines and different things, the degree of explainability is completely different.
The ChatGPT, produced by humans after "refining" from massive data, is more unexplainable, and it is more like an organic life form.
However, Professor Li Xing also proposed another possibility, that is, the scientific system and theoretical tools used to understand artificial intelligence, which have not yet been discovered. In other words, only after the birth of a new theoretical tool can the essence and principle of artificial intelligence be truly revealed. "Just as before the invention of calculus, we lacked a basis to scientifically describe the three laws of mechanics. In all likelihood, the mathematical tools used to study ChatGPT and artificial intelligence have not yet appeared. ”
In summary, Professor Li Xing believes that although the unexplainability of artificial intelligence is a challenge, by exploring new methods and tools, we can strive to understand and reveal its laws. This may involve drawing on the experience of existing disciplines, but may also require the creation of entirely new scientific systems and theoretical tools. Only then can we better understand and apply artificial intelligence.
The distance between artificial intelligence and humans
On the issue of the distance between artificial intelligence and humans, Professor Li Xing believes that the current artificial intelligence is still a usable tool from the perspective of humans, but unlike previous tools, artificial intelligence is based on probability inference, so it has an irrevocable randomness. Therefore, its answer is sometimes correct, and sometimes it is like "a book of serious nonsense", for which we must make our own judgments and decisions, and cannot rely entirely on it.
So, will artificial intelligence develop into an independent intelligent actor similar to humans?
In response to this problem, Professor Li Xing believes that first of all, artificial intelligence must be materialized and have subjectivity. Just as humans are responsible for their actions, AI requires a separate entity to take responsibility. This materialization is not only to give artificial intelligence a shape, but also to establish a close connection between artificial intelligence thinking and reasoning mechanisms and entities, similar to the connection between human bodies and minds.
Second, AI needs to have the ability to reproduce. Reproduction is how living things continue to live, and through evolution and retention of different factors, life becomes more intelligent, flexible and powerful, and at the same time produces creativity. If AI can't reproduce on its own, its capabilities will only decay over time, stagnate, and eventually become obsolete.
Professor Li Xing believes that although the current artificial intelligence has certain creativity and logical reasoning ability, it is only part of human intelligence. Human beings have complete creativity, imagination and subjective consciousness, which is different from the creative ability of artificial intelligence.
The difference between human creativity and artificial intelligence creativity is similar to the difference between a real artist and a skilled craftsman, that is, the difference between Tao and art. The same goes for imagination. Physicist Einstein said in his article "On Science": "Imagination is more important than knowledge, because knowledge is limited, and imagination summarizes everything in the world, promotes progress, and is the source of the evolution of knowledge." At present, although artificial intelligence seems to have "infinite" knowledge, as long as it is knowledge, it has boundaries, and this "infinity" is always a "finite" compared with human imagination.
"As for whether artificial intelligence can eventually simulate creativity and imagination close to humans, we still don't know." Professor Li Xing said.
The zooization of human society
With the advent of ChatGPT, a widespread concern quickly emerged: Will human jobs be replaced by artificial intelligence?
Professor Li Xing believes that the extreme possibility is that the future human society may be similar to a super-large zoo, and most people may enter the "lying flat" state like the lazy animals waiting to be fed in the zoo, spending most of their time on entertainment and leisure, and only need to fill their stomachs at a specific time.
However, he also said that there will definitely be some people who are not satisfied with this state. There will always be people who realize that what really matters and is the process of achieving the goal, not the goal itself. Although goals are still an important prerequisite, for true masters, the most valuable and exciting part is not the result of achieving the goal, but the process of struggle, experience, and self-challenge.
Therefore, even if artificial intelligence develops more powerful, human society may enter a certain "lying flat" state, but there will still be those who are not satisfied with this, and they will continue to pursue their goals and find true value and meaning in the process of struggle.
A great era of challenges and opportunities
"The significance of artificial intelligence and various artificial intelligence technologies is similar to, and perhaps even more important, and wider in impact than the TCP/IP protocol in the Internet." Professor Li Xing believes that compared with the invention, popularization and use of the Internet, artificial intelligence has a more far-reaching impact on human society.
However, the impact of AI has both positive and negative aspects. The shortcomings of artificial intelligence bring us challenges, but at the same time its advantages also bring us unlimited opportunities. For researchers in the Internet field, Professor Li Xing believes that there are at least two major challenges.
The first is the challenge of "distributed". At present, large-scale artificial intelligence computing models such as ChatGPT are centralized, and the nature of the Internet is distributed, which means that artificial intelligence and the design principle of the Internet are exactly the opposite. As an Internet researcher, Professor Li Xing believes that ChatGPT and artificial intelligence should be able to be deployed on a large scale, rather than centralized. Of course, distributed also has its own problems, such as whether the computing efficiency of artificial intelligence can be maintained or even improved after distribution? Is the whole system working well? and other questions need to be further explored.
The second challenge is "trustworthiness." ChatGPT sometimes "carries nonsense in a serious way." According to its principle of randomness, at the moment it can be considered that it is not intentional nonsense. However, thousands of ChatGPT-like services may appear in the future, and it cannot be ruled out that there are cases where misinformation is deliberately provided. Therefore, trustworthiness is very important. Can we provide some level of trustworthiness through the infrastructure of the Internet?
Professor Li Xing mentioned that when IPv6 was first promoted, everyone was looking for IPv6 "killer applications". It was initially thought that online video and the Internet of Things might be the killer applications of IPv6, but it turned out that IPv4 could also meet these needs.
"So what exactly is the killer application of IPv6? When we did IPv4/IPv6 stateless translation transition technology IVI, we proposed that interconnection with IPv4 is the killer application of IPv6. Now I think that robot networking (and even neuron networking) should also be added, that is, the connection between artificial intelligence and robots, which should also be the killer application of IPv6. Professor Li Xing said.
"It's the best of times, it's the worst of times." This sentence is used to describe the era of artificial intelligence, and it is appropriate. Professor Li Xing believes that after the emergence of new things such as artificial intelligence, our original knowledge and skills can be matched, integrated and integrated with new things. Even if there is a contradiction between the two, it does not matter, because something more innovative and interesting may be produced in a frictional collision. This is a gift given to mankind in the era of technological explosion, and it is also an opportunity that we should seize.
This article is based on an interview with Professor Li Xing of the Department of Electronic Engineering of Tsinghua University
Written by: Wang Yajing, ChatGPT also contributed to this article
Source: China Education Network