Throughout history, most of the discontinuous productivity transitions in economic cycles have come from scientific research and innovation. Similarly, in the current epidemic, global economic problems and industrial momentum transformation, scientific research and innovation can help us effectively navigate the economic cycle.
In our time, we have witnessed too many cases of entering the industry from scientific research and innovation, and I think that entrepreneurs who have come out of scientific research and innovation have a core condition, that is, cognitive leadership - the leading judgment of scientific and technological development is a prerequisite for building industry barriers.
As a general technology, artificial intelligence not only requires us to have a prior understanding of the technological path, but more importantly, it can effectively help us break through the boundaries of human cognition and explore the unknown.
If scientific and technological innovation is to become a real engine of the times, it must go through two stages
Looking back at the past 300 years, the core driving force of economic development in each era is technology, and technology is the engine of the times in the true sense. If you subdivide the development context of each technology, you can see that each period has a discontinuous point, and the economic growth curve presents two slopes, that is, when each technological breakthrough is promoted to promote economic development, it faces at least two stages: the first stage is the stage of technology becoming practical, in this stage, the technology from the realization of the breakthrough to the real change in the public life, and the public is gradually familiar with the technology; the second stage is the mass production stage of technology scale, and the cost of technology is reduced enough to land.
There are two reasons for the discontinuity of the growth slope: First, when the technology breaks through to practical value and can be used in the industry, the challenge is still that the industry or the public awareness has not kept up. This takes a period of market education. For example, when alternating current comes out, the public thinks it will electrocute cattle. When the car first came out, the United Kingdom also set up the so-called "Red Flag Act", worried that the car driving on the road, it would scare the horse, so it was required that at 60 yards in front of the car, someone waved a red flag to inform the surrounding people and horses that there was a car coming. What I didn't realize at the time was that in fact, with a car, I wouldn't use a horse anymore. Therefore, the improvement of cognition takes time.
The second stage is the process of reducing the cost of technology commercialization. If production costs become generally higher, the industry cannot undergo disruptive changes. Assuming that playing Go is a productive way, it will take decades to recoup costs with existing AlphaGo/AlphaZero investments. Therefore, when the technology breakthrough, it is necessary to explore ways to reduce production costs under new applications.
This is also why technological breakthroughs are not yet a sign of technology empowering hundreds of industries. It is not until the second stage comes, when entering large-scale production, when the cost is reduced to one-tenth or one percent of the original, that all walks of life will change.
The same goes for artificial intelligence. This wave of artificial intelligence began in 2010. At that time, artificial intelligence was actually "artificially guided intelligence", whether it was traditional generative or discriminative, in fact, it was very dependent on human prior knowledge. One of the biggest problems with human-directed intelligence is that its logical upper limit is human cognition, which brings about the limitations of development.
At present, the development of artificial intelligence depends on changes in two factors. The first is that artificial intelligence is transformed into data-driven, and the algorithm itself has a good "expression ability" for the world; the second is the improvement of computing power, giving artificial intelligence a greater ability to explore and solve the space. At this time, artificial intelligence no longer relies on human prior knowledge, but may transcend human cognition. But it wasn't until around 2014 that it really surpassed people in a lot of verticals.
The foundation of innovation comes from human cognitive innovation in the true sense
From the first stage of scientific and technological innovation to the real commercialization, there is also a key, that is, whether its production cost is too high, resulting in the industry is difficult to apply. Especially in industrial manufacturing, there will be similar problems. For example, if AlphaGo is used in factories, if it takes 70 years to renovate and upgrade the production line to return the cost, no one will use it at all. Therefore, cost reduction is the core of whether it can empower the development of hundreds of industries.
If the first stage is "standing on top of the sky", the second stage is "overwhelming the earth". What are the foundations and conditions for reaching the top of the world? It's about human cognitive innovation in the true sense of the word. Some people have summarized the four paradigms of innovation, two of which have a long history of existence, one is Aristotle's deduction of reasoning, and the Western theory of philosophical science, including Euclidean geometry and Newton's laws, are derived in this way, that is, from a fundamental origin to a boundary that can no longer be advanced, by constantly exploring the possibility of a farther boundary, so as to go further. It wasn't until a thousand years later that Bacon came up with another paradigm of innovation—empirical induction. This also means that it took thousands of years for humans to break through the cognitive limits.
Why did it take a thousand years for a new paradigm of innovation to emerge? The core reason is that human beings did not know their ignorance before, always felt that they could use a unified method to deduce the boundaries of a certain thing, and expected to be able to pursue GUT (Great Unification Theory), and only after repeated failures did they find that empirical induction is only a valuable innovative abstraction under specific space-time conditions. Until 2007, some scholars proposed that the third paradigm is computer simulation, that is, through an analytic and basic state, let the computer simulate. Today's weather forecasts and even the cosmic three-body problem belong to the paradigm of computer simulations. The fourth paradigm is big data-intensive innovation. This can be said to be a computer induction.
The core ability to explore the unknown world is changing
However, the real scientific research and innovation breakthroughs of mankind often come from the brain holes of geniuses or the conjectures of geniuses. Scientific breakthroughs in human history are basically based on various thought experiments, and even those that are unreliable. This kind of sudden thought experiment was finally verified and called the brain hole of genius. But why hasn't the brain hole of genius become a paradigm of scientific innovation in human history? Because the paradigm is predictable, but the flash of genius is unpredictable, but the major scientific breakthroughs that people remember today came from that path.
Interestingly, Turing asked the question in 1950: Do machines think? 70 years after this question is raised, we can ask another question: Can machines make conjectures? Actually, the answer is yes. Because after this change, the pattern of human cognition of the world has changed dramatically, because the conjecture of machines is scalable and replicable.
Still take Go as an example. The complexity of Go is 3 to the 361st power, that is, black and white plus no total of 361 squares, that is, 10 to the 170th power. Although its complexity is not actually the highest we have solved, it is enough to prove that it is a conjecture. The 170th power of 10 does not seem like a very large space, but it is not easy to calculate the 170th power of 10. For example, if the number of atomic lattices in the universe is 10 to the 70th power, if each atom is a supercomputer, calculating 100 million billion times per second, from the Big Bang to today, the 10th to the 170th power cannot be counted. Therefore, with the existing human cognition, it is impossible to get a definite solution, in a sense, AlphaGo/AlphaZero is not solving at all, but guessing a solution that seems to be correct.
Today we are faced with the problem: if computers and artificial intelligence can easily guess Newton's laws, and Newton may not be born until 200 years later, then what should we not use Newton's laws? This means that how we should find the boundaries of the new conjecture; in addition, the new conjecture must also have path dependence and not be too advanced.
For example, we travel back in time to Aristotle's time and talk to him about Newton's laws, tell him how everything works, what laws mechanics are, and he may be able to understand them step by step. But if you tell Aristotle a conjecture, a conjecture about relativity, light bends, and he doesn't even know what light is. Therefore, it is not possible to be too advanced in cognition.
Liu Cixin's novel "Shiyun" once wrote that an alien high-end civilization came to Earth, believing that the Earth civilization was too weak and the bit rate of information transmission was too low, so it was necessary to exterminate human civilization. Later, when I saw the poems in human history, I thought it was very interesting. He came up with an artificial intelligence "poetry cloud".
Well, the problem is, you can write all the possible poems, including the best ones, but you can't figure out what good poetry is.
In fact, artificial intelligence can give good conjectures today, but sadly, humans have no way to identify which conjectures are correct, truly meaningful, and of great value at this time. So we still need to learn from the history of human development, test the boundaries of conjectures in application, and use this boundary to standardize the conjectures we use. Therefore, its use must be completely accelerated in the innovation chain in order to really use the "Newton's law" that we guessed.
This is a change in the core capabilities of our generation's exploration of the unknown world. In the future, the more disciplines without laws, the more they will be subverted, because these disciplines are areas where machines can make a great difference.
Artificial intelligence has the opportunity to become the label of our time
We've always thought we had a good grip on the world, but the reality is that we find boundaries in our applications. Now many people think that artificial intelligence is driven by big data, but this is not entirely the case.
The primary core of the development of artificial intelligence lies in the large computing power behind artificial intelligence. We see that in some applications, in fact, it is not completely data-driven, we ourselves are pushing an AI device, at the bottom of this AI to do conjecture mode. The reason why it is called an AI device is that it is compared to the particle collider, artificial intelligence and particle collider were invented in 1956, and both use randomness to solve the problem of exploring the world.
In the past decade, the demand for computing power of the best ARTIFICIAL algorithms has increased by 1 million times, many of which are useless for data, and many long-tail applications are not big data problems.
The second core is "overwhelming," a term that described internet marketing 10 years ago. The internet's greatest contribution over the past 20 years has been to solve the long-tail demand matching. In fact, offline applications, including industrial and urban governance of artificial intelligence systems to solve the completion of this long-tail chain, thereby forming a value closed loop, because the current bottleneck is human efficiency.
Each person encounters about 600 objects a day, and every three things form a detection element, such as when people are speaking, people, hand-held microphones, and stages can be used to identify the behavior of speech. If a person encounters 600 items a day, there are more than 34 million combinations, and if there are more than 34 million AI models, it is possible to form a real physical world of the Internet and search engines, so that mechanical manufacturing, unmanned driving and other applications are benefiting from it. But even we have only produced 22,000 commercial AI models in different application scenarios, which is still far from more than 34 million. The same is true for industrial applications, for example, the application of defect detection of overhead cables of AI high-speed rail, these defects have high frequencies, there are also very low frequencies, but the problem is that if artificial intelligence can only solve high frequency defects, still have to rely on manual solutions to very low frequencies, or even low frequency defects, the efficiency has not improved, so for the long tail problem, there is no possibility of solving the head problem to solve all the problems, there is no so-called two-eight law.
What exactly is the era? Behind the times is technology. Steam, electricity, and even the Stone and Iron Ages to the Information Age are all technologies. Technology is the driving force behind changing the political, economic and cultural culture, bringing about leaps and bounds in productivity.
But there are also some technologies that have not become the label of the times, such as the cloning technology of the past, not because it is not unimportant, but because it has not changed the price of the factors of production. The reason why artificial intelligence has the opportunity to become the label of our time is because artificial intelligence can reduce the price of factors of production in our time.
Predicting Machines: The Simple Economics of Artificial Intelligence mentions that when the price of a certain base product drops dramatically, the whole world changes. Now the production efficiency of artificial intelligence models has increased by 300 times in three years, and when artificial intelligence really goes to the ground, it can truly become the infrastructure of an era.
Author: Xu Li (The author is the co-founder, chairman and CEO of SenseTime, and this article is the author's speech at the Antai School of Economics and Management of Shanghai Jiao Tong University)
Editor: Chu Shuting
Image source: Shanghai Jiao Tong University
![From "Genius Girl" to "Elder of the Four Dynasties" Cai Xuetong: The fourth place is not my end[fig]](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACwAAAAAAQABAAACAkQBADs=)