The sword of the Yi Gongzi - from a business point of view, look at the world
In the 90s, SGI (Silicon Valley Graphics) developed high-performance computers, and it was rich. On the day Lee joined the company, the company happened to issue a regular "small" benefit, and all employees in the company had a $1,500 Tag Heuer watch and an expensive lambskin jacket for each employee.
Kai-Fu Lee has been at Apple for six years and has never seen such a trench pomp and circumstance.
But in the past two years alone, the SGI has gone from luxury to bankruptcy. Kai-Fu Lee, who witnessed this ups and downs, summed up a lesson: scientific and technological innovation must not be separated from user needs.
When Kai-Fu Lee decided to change jobs, Intel and Microsoft both threw olive branches, and the work content was surprisingly the same - go to China, build a research institute, and recruit China's young generation and vigorous high-caliber students.
Jobs also called. The former boss had just returned to Apple from exile and was ready to do something. "Kaifu, why are you going to Microsoft? Why go to China? Hurry up and give me back the apple! ”
Lee refused. During his time at the SGI, he acted as a bridge between the company and China, and it was clear that China's rapid development, especially the emergence of the Internet, convinced Lee that the Internet would change China.
In 1998, he returned to China to establish Microsoft Research China, later known as Microsoft Research Asia. This is the most incredible year for the power of the Internet, Google, Sohu, JD.com, Tencent, and Sina were born one after another.
Chinese engaged in computer science in the United States have been visited by Kai-Fu Lee. After being acquired by General Electric (GE), the focus of Zhang Yaqin's focus shifted from pure scientific research to turning technology into products. This "market" experience makes Lee pay special attention to it.
To engage in research, we must first build culture, and everyone is equal before science. Lee Kai-fu demanded that he should not be called "president" or "dean", but must be called "Kaifu". Zhang Yaqin was the first to make bad and used the name of the restaurant "KFC" that was very popular in Beijing at that time to call Lee Kaifu, because the latter's name was abbreviated as KF. Lee fought back, calling Zhang Yaqin "Toothpick" (YQ).
The nascent Microsoft Research Asia is in a storm, and the government is also very supportive, and joining can solve the Beijing hukou. In the direction of basic research, combination of production and research, and training of young scientists, Microsoft Research Asia has sown a seed for China.
The story that followed was familiar to everyone. More famous than the research content is the "people" of Microsoft Research Asia, Zhang Yaqin, Shen Xiangyang, Wang Jian, Wang Haifeng, Lin Bin, Yu Dong, Deng Dafu, Jia Jiaya, Sun Jian, Yu Kai, Tang Xiaoou... This long list of "graduates", some went to Baidu, Ali, Tencent, the headline Xiaomi, and some went to the sea to start small giants of entrepreneurship and technology, incubating scientific research into specific solutions, supporting half of China's technology business.
01
Torrential currents. Twenty-four years later, things are very different. China's AI technology and applications have also reached the first echelon in the world.
In 2022, Kai-Fu Lee told Chinese media, "China will lead the global AI development together with the United States, and in the next two decades, AI and automation as platform technologies, and the intersection of advanced computing, life sciences, new energy and other new technologies will bring disruptive industrial changes, and China is expected to lead in this process."
But before the words fell, Microsoft's milky ChatGPT was born, making a common topic hot again: how big is the AI gap between the United States and China? Fears abounded.
Recognize the gap and measure it objectively.
It may be generous to admit that in many aspects of technology and business, the United States is still leading the world and is an example for us to learn from and catch up. But the vast market and rich scenarios are a springboard for Chinese AI. Especially in traditional industries with huge scale, there is a deep "industry moat". The world's factories are in a critical period of digitalization and intelligent upgrading and transformation, and platform technologies such as AI and automation are taking turns to promote their iteration to a new paradigm of growth driven by technological innovation.
Kai-Fu Lee also predicted that in the future, the AI competition between China and the United States is not necessarily a zero-sum game, especially China's advantages in big data and AI, automation and intelligence are expected to lead the United States and maintain the status of the "world factory".
If you extend the time, from Microsoft Research Asia to ChatGPT, the two different development paths of AI in China and the United States are clearer.
Kaiming He of Microsoft Research Asia and his mentor Jian Sun developed the deep residual network ResNet. DeepMind's AlphaGo Zero leverages ResNet's research.
Later, Sun Jian left the institute and joined Megvii. And Tang Xiaoou, another mentor of Ho Kaiming, also embraced China's local AI market and started SenseTime after leaving Microsoft Research Asia.
Megvii and SenseTime have become the "four tigers" of China's computer vision AI. The CV field is a typical "scene-technology-product" rotation, and then drive a representative of the next generation of technology leap, in the same period, not only the four tigers, but also Tencent Youtu and other technology companies layout. The bigger aspect I think of is that the research of face detection has emerged in the laboratory decades ago, through generations of research scholars, but this generation of scientists finally has the opportunity to move the hard work of generations of scholars out of the laboratory, through mobile phone unlocking, mobile payment, into the life of every ordinary person.
AI top talents are like this, doing research is the best, and then jumping out to join China's vigorous AI career, supporting half of China's technology industry. This proves at least two things, first, Chinese can also do basic research and can also do a good job of general artificial intelligence. Second, there is no difference between doing basic research and doing scenario solutions. It is precisely the two that complement each other and jointly promote the development of the AI industry.
Many people say that the biggest difference between AI between China and the United States is that the United States focuses on basic research and China focuses on solutions. In fact, not only artificial intelligence, but all scientific and technological developments in this century have derived different paths on both sides of the Pacific.
In the wave of the Internet, the United States is relatively less enthusiastic about e-commerce, and the penetration rate of online consumption has not been up. In contrast, almost all Internet companies in China have done e-commerce, and China's e-commerce penetration rate is the largest in the world, once higher than the 2nd to 11th combined.
The same goes for mobile internet. US "mobile payments" are not active. However, with a better network environment and a regulatory system that encourages innovation, China leapfrogged its development and directly skipped the credit card era and entered the era of digital payments.
Another example is unmanned driving. The United States focuses on the intelligence of vehicles. China's advantage is better infrastructure, road conditions, network and transportation planning, so it chose the route of road coordination.
There are also obvious differences in the industrial Internet. The industrial characteristics of the US economy are concentrated in the upstream and downstream of the smile curve, with a high proportion of technology, the Internet, and finance, coupled with expensive manpower, and strong willingness to pay. The characteristics of China are that the industry is concentrated in the middle of the smile curve, as the world's factory, rich scenes, complete industrial chain, coupled with policy support, university concentration, industry-university-research docking is very convenient, and technical verification is better landed. As a result, the United States focuses on basic research, mostly starting from technology, while China's advantage lies in many scenarios, more needs, and more scenarios, often the landing of scene backwards technology.
02
The year Kai-Fu Lee founded Microsoft Research China. In the opposite direction, Ren went to the United States and marveled at Bell Labs. The Americans are doing this thing well, combining production and research, large enterprises run research institutes, commercial products make profits, long-term research and innovation, and walking on two legs can be more stable.
Today, Ren Zhengfei's sigh reflects another dimension - the attitude towards basic research is not a matter of vision, but of strength.
To use an inappropriate metaphor, when Microsoft came to China to build a research institute, when Ren Zhengfei envied Bell Labs in the United States, Microsoft and other American giants were already rich young masters, money was not a problem, too much money was a problem. To throw money at basic science, the benefits are obvious - both to seize the high ground of science and technology, but also to dilute the bad image of monopolists.
And China's private enterprises are just on the road of hard struggle. The habit of careful calculation cannot be changed, often starting from market demand product demand, and then slowly investing in scientists and basic research, and then combining market demand to drive basic research landing.
I have talked to many friends who are engaged in AI research in companies, and their starting point is often simple and realistic. For example, Tencent's early development of AI technology was due to the product demand of QQ space - many users used personal computers to take selfies of QQ avatars. The QQ team thought that making a technology could achieve the center of the avatar.
After solving this technical problem, a small team gradually settled out, jumped out of the product, and specialized in the image technology itself. Incubated face detection, portrait expression, intelligent P map and other technologies, and then used back to the product, hatched every day P map, portrait beauty technology, was applied to the national product national K song. This small team, later known as Tencent Youtu.
In 2015, it was also because of the demand - financial reform to encourage small and micro financial inclusion. At that time, WeBank was established, the first bank in China without offline branches, and the image technology accumulated by Tencent was used for online remote account opening verification and risk control. The industrial technical value of AI is beginning to appear.
Then enter the industrial Internet, such as using image AI to do quality inspection. The manufacturing industry is not tall enough, but using AI to do quality inspection is really not simple. Take Fuchi Hi-Tech as an example, its product is the bracket of the mobile phone camera, which is irregular in shape and only the size of a finger, but there are as many as seventy or eighty points that need to be detected on it. There are two difficulties, one is that the false detection rate must meet the standard, otherwise the missing defects need to be additionally arranged for manual detection, which is equivalent to not improving human efficiency; The second is to race against time, because mobile phone products are upgraded every year, so the new generation of quality inspection programs must be made before the production of new generation products, otherwise it will not land.
Tencent uses the relevant capabilities of large models such as mixed elements in industrial scenarios, and exports capabilities through the Tencent Cloud TI platform to facilitate customers to do their own data training, and more than 50 users in eight industries have used this capability. Fuchi Hi-Tech calculated that the efficiency of the AI in this project is 20 times that of the original manual work, and under the condition that the machine continues to be fully loaded, it can save the company tens of millions of yuan in costs a year.
Ma Huateng has always said that it is necessary to follow the technology research and development strategy of paying equal attention to research and application. Tencent's AI lab later developed into a matrix, Tencent Youtu, AI Lab, WechatAI and RoboticsX RoboticsX Robot Lab, which can not only come up with ultra-large-scale (trillion) pre-training models "mixed element large models", but also Tencent Cloud Intelligence, which combines cloud and AI, takes root in factories with technology and runs out of the industrial quality inspection training platform TI-AOI.
Compared with the United States, Chinese science and technology has a distinct feature, starting from the needs of the scene - going to the bottom to do technology - and then returning to product innovation, so as to roll forward like a wheel.
In the transparent factory of Baowu Steel's "1580 production line", blue-collar workers become white-collar workers, and there is no need to go to the noisy and hot production line, as long as they are in a quiet, air-conditioned conference room. The digital twin technology used in it is derived from the game scene that Tencent is good at.
Similar to Tencent. Alibaba's cloud services, which first started in e-commerce, have been helping small businesses sell goods online 20 years ago. Meituan's drone technology has been humming to the head, and many years of investment in drones is because it is closely related to the main business of "food delivery". Jingdong mainly promotes intelligent supply chain, which is also derived from the technical foundation of JD.com's self-built logistics for 20 years.
The Industrial Internet is a much larger market than the Consumer Internet. Especially in China, the world's second largest economy, the world's factory, rich scenes. Moreover, it is undergoing a transformation from industrial layout data labor intensive to efficiency improvement. At present, various Chinese technology companies have taken out their own resources to rush in, but the technical success of each of them is highly related to their main business.
In contrast, the last hit in the US AI industry was DeepMind's Alpha series, which first made the technology to win the human Go championship, which is explosive enough. And the application scenario, commercialization is not urgent, slowly groping, several years later this technology is used to solve the problem of protein folding structure, participate in the research and development of new drugs, only then can the hero have a place.
Similarly, the current inductive generation stage of ChatGPT's information is still relatively rudimentary, and it is still necessary to continue to practice, polish and iterate in application scenarios to really come in handy in the laboratory.
03
In other words, whether it was Microsoft China Research 25 years ago or now it is trying to support ChatGPT, Microsoft has its own calculations.
On the bright side, it is to cultivate Chinese young talents, combine industry and finance, and pursue innovation... But there is still some embarrassment to say, Lee Kai-fu shared in his autobiography: At that time, the Chinese market called Microsoft Micro$oft, because it was too profitable, a set of software sold so expensive, and monopolized. But the work of founding the China Research Institute has greatly improved Microsoft's image.
I have to say that Microsoft is really "selfish", but it is this "selfishness" that promotes the progress of human technology, and the biggest charm of the market economy is to encourage enterprises to freely pursue their own interests, technical pursuit and business desire, driving social progress. The virtuous circle behind scientific and technological innovation comes from the cycle of demand - profit - continuous investment.
The emergence of ChatGPT also has new enlightenment, and the breakthrough of AI requirements for technology, data, and scenarios requires continuous investment from large enterprises.
OpenAI's talented engineers are precious, but Microsoft behind them is also indispensable. Engaging in AI large models is an extremely expensive thing, the data contains nearly 100 billion parameters, 10,000 NVIDIA A100 chips are the threshold of computing power, a complete model training cost is close to 100 million yuan, and the results are full of uncertainty.
Microsoft first invested $1 billion, then $10 billion, all the way from internal resources, to servers to equipment, cloud computing resources, and even moved the cheese of Microsoft's internal technology department.
Several aspects of Microsoft's contribution come from the nature of its platform enterprise. ChatGPT uses cloud resources, and if leased at market prices, OpenAI would have gone bankrupt, but Microsoft has a cloud business and can be very flexible. Moreover, platform-based enterprises have data for ChatGPT training, users, businesses, and scenarios, which can allow the technologies encountered in AI development to get experimental opportunities.
There is one thing I want to emphasize. ChatGPT is absolutely good for human AI, but it is not necessarily a sure good thing for Microsoft, a commercial enterprise.
Because innovation is unforeseeable. The core technology of the ChatGPT model is Transformer. But who knows if the next model to beat Transformer will pop up tomorrow?
Back then, AlphaGo won the human Go championship, what a sight Google was. But as soon as ChatGPT came out today, didn't it also become the object of ridicule. Further on, the grand celebration of human AI was IBM's Deep Blue defeating the chess championship. Later, IBM was overly obsessed with technology and fell behind.
Therefore, technological innovation such as AI is suitable for platform enterprises located in the middle of society. Down, entrepreneurial enterprises, no money, no resources, no scenes, can't do it. Up, government industry funds have money and resources, but they are often far from commercial scenarios. In the final analysis, high-precision technological innovation is nine deaths, to encourage large enterprises to take risks, to build a good model of production, education and research, to rely on market incentives.
When it comes to the difference between the AI industry between China and the United States, in fact, China's foundation is very good. The number of AI papers is the first very early, and the quality of the papers is also very good. With the top 10% of cited papers, China jumped to the top spot in 2019. Nikkei Shimbun conducted a statistical analysis with the assistance of Elsevier, a large Dutch academic information company, and the top 10 companies with the number of papers in 2021, four Chinese companies in the top 10, Tencent, Alibaba, Huawei, and State Grid ranked 5th, 6th, 7th, and 9th, respectively. In the field of artificial intelligence, Tencent has filed 10,630 patent applications in the past five years, ranking first in the global Internet industry.
Again, in the face of the AI gap between China and the United States, we must not only acknowledge the gap, but also not be presumptuous. Find your strengths, find your problems.
What China should do is to make good use of its own characteristics, rich industrial chain scenarios, industry-university-research integration, and active investment of Internet leading enterprises, and make good use of the market cycle.
Just as Minister of Science and Technology Wang Zhigang recently mentioned when talking about ChatGPT, he hopes to combine both scientific research and technology traction, as well as through scenario driving and user needs, so that AI can contribute to China's economic and social development and China's scientific and technological development.
This year's theme in China is to grasp the economy, and senior leaders have recently stressed the need to support the private economy. Indeed, technological innovation competes with who has a better business environment. To encourage innovation, we must tolerate trial and error, have reasonable market incentives, and platform enterprises dare to invest, so that innovative technologies such as AI can create a bloody road.