"Fortune" (Chinese version): There was an exchange three years ago. At that time, you said that artificial intelligence would usher in explosive growth if it exceeded an industrial red line. Do you think artificial intelligence has now surpassed that industrial red line?

Xu Li: We have actually benefited from the dividends of two waves of artificial intelligence development. The first wave is what you just mentioned. The technological breakthrough five years ago proved for the first time that AI can exceed the "industrial red line" in some fields, which can usually be considered to surpass human accuracy. But at that time we also faced another dilemma, that is, as the technology gradually penetrates into the industry, we will find that the cost of producing artificial intelligence itself is very high.

To give a simple example, if we assume "playing Go" as a production process. We can think of AlphaGo as exceeding the industrial red line, which is the human level. But if the purpose of investing in AlphaGo is just to make him play chess as his main occupation, then its investment may be higher than the total investment of nearly 300 professional chess players in China. Therefore, the first wave of dividends usually only brings the first adoption of AI technology in certain specific industries. Large-scale application also depends on reducing the production cost of AI itself.

The second dividend of the development of artificial intelligence is that with the continuous investment in underlying AI infrastructure, the ability of neural network technology to train general models is gradually breaking through. At present, automated, large-scale, and intensive mass production of AI models is gradually being realized.

When SenseTime was founded, it began to invest in and train general-purpose large models. Simply put, general large models have the ability to draw parallels. After establishing general capabilities, in subdivided scenarios in different industry fields, high-quality model algorithms can be iterated on only relying on small samples, truly breaking through labor-intensive investment and meeting the needs of long-tail applications. This is like analogizing human abilities. Human perception and cognitive abilities are very universal. To understand some new problems, we only need small data.

Fortune (Chinese version): I saw you participated in a public event some time ago. You said that true disruptive innovation does not come from the traditional innovation paradigm, but from the brainstorming or thought experiments of geniuses. Do you believe that geniuses’ flashes of inspiration are accidental or inevitable?

Xu Li: First of all, many disruptive scientific innovations and major breakthroughs in human history were accomplished through very accidental discoveries or accidental thought experiments. Most of them originated from "genius's guesses", and this kind of completion is unpredictable.

So what's the reason behind it?

Major innovations are anti-consensus. As long as they are anti-consensus, it is impossible to plan, so they cannot be standardized by paradigms. It is impossible for humans to use fixed methodology to plan something that the general public has no consensus on and predict what will happen in the future. The reason for this phenomenon lies in the limitations of human understanding of the unknown world.

But in today's era of artificial intelligence, machines can also make guesses without relying on human cognition. It is expected to help us discover the nature of scientific laws earlier and explore and discover the unknown faster. This is the innovative new paradigm brought by artificial intelligence.

I mentioned last time that assuming a machine can guess Newton's law, will humans use this law? This is a problem we face. Of course, it is a difficult problem in itself for humans to distinguish valid guesses from machines.

I have read a novel by Liu Cixin. A very powerful agent wants to understand the writing of poetry by humans. In the end, we found that even if we could exhaust all the possible poems, we still couldn’t know which poem was good. Therefore, it was also full of challenges to identify the pros and cons of machine guesses without fully understanding them.

"Fortune" (Chinese version): My understanding is that this is the real difference between intelligence and wisdom.

Xu Li: For example, we have many machine conjectures about some future possibilities and technological breakthroughs. Some of these conjectures can greatly promote technological progress in the future, but some of them cannot. So how to screen the validity and reliability of these conjectures will be a question left to the future.

"Fortune" (Chinese version): So our topic today is to discuss the next step of artificial intelligence in the future. In the foreseeable future, what do you think is the next step of artificial intelligence?

Xu Li: I just mentioned the two major dividends of artificial intelligence in recent years. It is with this understanding that SenseTime has invested nearly 10 billion in total to build the AI ​​infrastructure necessary for the large-scale production of artificial intelligence algorithms. We call it a large artificial intelligence device.

The reason why it is named a large device is analogous to the particle collider of physics, and it is hoped that artificial intelligence will help to transform an innovative paradigm. Through the "artificial intelligence particle collider", massive data and even a huge solution space will be disassembled and collided, using a certain degree of randomness to break the boundaries of human cognition and application.

I think that if some industries have many unknown fields, such as earth sciences, life sciences, pharmaceuticals, atmospheric science research, etc., machine guessing in these fields may give us some unexpected surprises.

"Fortune" (Chinese version): Is this random experiment a matter of chance as you just mentioned?

Xu Li: To give an example, recently we have seen many breakthroughs in protein structure analysis by artificial intelligence. It may take a long time for humans to complete such a job, but machines can do it in a short time.

The current hypothesis is that the sequence of amino acids is the only factor that determines protein structure. But then we can also do many random impact experiments similar to particle colliders. For example, put more input factors into the experiment to see if there is any improvement in the predicted results. When predictions become more accurate, we can use such results to re-understand the core elements that determine protein structure. This is accelerating scientific research through random collisions. The results of R&D will accelerate progress in biopharmaceuticals and other sciences.

Of course, sometimes related breakthroughs may not be explained in this era. It may be necessary to wait until a "Newton" appears a hundred years after the machine guesses Newton's laws to truly explain the reasons behind it.

Fortune (Chinese version): Is this what SenseTime is doing as well? For example, the combination of proteins you just mentioned.

Xu Li: SenseTime’s large AI device provides an underlying capability platform. As mentioned before, as does a particle collider. In this process with a certain degree of randomness, there is a very large space for exploration. The demand for infrastructure, especially computing power, is actually growing exponentially. Data shows that in the past ten years, the computing power requirements of the best artificial intelligence algorithms have increased by more than 1 million times.

This is contrary to the perception of many people. Everyone thinks that the more sophisticated the algorithm, the less it should calculate. In fact, this is not the case. The better the algorithm, the more it can be calculated. It is actually verifying different possibilities and exploring the boundaries of applications.

Fortune (Chinese version): It’s the anti-consensus you just mentioned.

Xu Li: Yes, anti-consensus.

"Fortune" (Chinese version): If we zoom in a little closer, the metaverse may be the hottest topic in the technology circle in the past year. What role does artificial intelligence play in the metaverse? How do you understand the Metaverse?

Xu Li: I think the Metaverse has always existed in various daily applications, including the games we originally talked about, which are actually some forms of the Metaverse.

But the core reason why the metaverse is so popular recently is actually the same reason why artificial intelligence is attracting everyone's attention, because it is gradually getting closer to the real world.

If the virtual world is very unreal, just like you play games every day, and the content you play is completely disconnected from the real world, you will not feel that it is a universe. Only when it is more like the real world, when everything in the real world can be reflected and connected to the metaverse, and can thus interact, you will feel that it gives me a second possibility, thus bringing about a second, third or even more universes.

In this process, the role of artificial intelligence is to connect the real world and the virtual world, allowing users to live in different universes. In the real world, everyone has their own identity, and in the virtual world, you can also interact with the surrounding environment in real time through other characters.

"Fortune" (Chinese version): Can the conversion between the real and virtual world be completed through computing power?

Xu Li: Yes. Computing power can help project the real world into the virtual world. Usually the first step is to collect relevant information, which has no real semantic meaning and only completes the so-called spatial digitization. However, through breakthroughs in AI and computing power, everything is gradually being given a meaning that humans can understand. It is not only the digitization of space, but also the structuring of elements and interactive processes, and then corresponds to the virtual world. Allowing people to not only access, but also use, modify and even interact with real-world relevant content.

So my understanding of the real metaverse is that people can interact with the real world in it, and can also project various experiences in the virtual world into the real world. Otherwise, its meaning is just a game, or just an APP.

"Fortune" (Chinese version): So if we go one step further, first I want to discuss a hypothesis and premise with you. Do you think we have reached a moment when we should discuss the ethical issues of artificial intelligence? First of all, do you agree with this hypothesis?

Xu Li: I think it is worth discussing. The innovation of artificial intelligence comes from the subversion of past perceptions, so ethical governance issues in industry applications have become increasingly important. At the same time, one of the paths of scientific development is deduction, that is, giving a starting point and constantly pushing it to the boundary. Since there are some core principles for this ethical governance today, we can completely evolve forward through the principles to explore the boundaries between our ethics and machines.

Fortune (Chinese version): You just mentioned principles. If this principle were formulated for artificial intelligence, how would you formulate it?

Xu Li: SenseTime has conducted a relatively comprehensive study on the framework of ethical governance of artificial intelligence around the world.

We found that most ethical governance frameworks can be classified into three major categories: the first category is people-centered, whether it is human fairness, identity issues, dignity issues, confirmation of sovereignty issues, data security issues, etc., all belong to the people-centered category; the second category is called technology controllable, including transparent calculation, explainability, technical security boundaries, etc.; the third point is sustainable development, seeking green, long-term development.

I believe that the future ethical governance of artificial intelligence will actually seek a balance among these three pillars.

Of course, what we have always emphasized is to uphold the ethics of "development" in artificial intelligence. The easiest way to meet all the requirements of the governance framework is to make no breakthroughs in anything, but then everyone will forget what our original goal is. Our original intention is to use artificial intelligence to promote the development of society. This is the goal. Taking mathematical optimization as an example, our optimization function is artificial intelligence to promote social progress. There are three constraints that need to be met. But many times, if we only talk about constraints and ignore the goal of the optimization function, then the optimization will be unsuccessful.

Fortune (Chinese version): If you put it this way, do you think artificial intelligence will exacerbate inequality or equality between people? What is the definition of this wonderful world? How does artificial intelligence reach this so-called beautiful world?

Xu Li: Talking about the fairness of artificial intelligence is basically talking about the deviation of data samples that may bring about the deviation of artificial intelligence algorithms. This can be improved with the evolution of technology. And I think fairness on a larger level comes from the possibility that a few people can benefit from the convenience brought by artificial intelligence, which will lead to inequality.

As for my understanding, the core factor for artificial intelligence as a general technology to change from quantitative to qualitative is whether it can truly reduce the cost of its production factors. If a certain technology is particularly sophisticated and requires a very high cost of investment to complete, then only a small number of people, leading companies, and leading scenarios can use such technology, which will inevitably lead to inequality.

Looking back at history, whether it is the Iron Age, the Steam Age, the Electric Age, or even the Information Age, all eras are named after technology. The technologies that name this era have one core thing in common, that is, they all significantly reduced the prices of production factors at that time, thus significantly improving production efficiency.

If artificial intelligence technology is to bring universality, universal benefit, and fairness, it must be that it reduces the cost of the production factors of our era hundreds of times. For example, in a true sense, it makes educational resources, medical resources, etc. that are originally inaccessible to us better, cheaper, and more equitably open to the whole society.

If you are interested in this topic, please click on the Wealth Plus applet below to express your views.