Community Blog Alibaba Researcher Wu Hanqing: What Kind of Intelligent System Does the World Need?

Alibaba Researcher Wu Hanqing: What Kind of Intelligent System Does the World Need?

In this post, Wu Hanqing, a researcher from Alibaba, shares his thoughts on artificial intelligence and its impact on our society.


Wu Hanqing, affectionately known as "Xiao Hei" or "Dao Ge", is a researcher at Alibaba. At the age of 15, he was admitted to the Special Class for the Gifted Young of Xi'an Jiaotong University. After graduation, he started working with Alibaba and became Alibaba's youngest senior technical expert at the age of 23. At 32, he was named as the 2017 MIT Technology Review's Top Innovators Under 35 (TR35). In spite of immense online attention, he prefers to keep a low profile and concentrate on his work. Since 2014, he has rarely published any articles. However, today he has something to share about himself, technology, and the future, and is writing this article to speak to you and the entire world .

A Personal Note: What I Have to Say

In the last 18 months, I refused all interviews and devoted myself to my work. Before moving ahead, I would like to clarify that all the news (including TikTok videos) about me which surfaced over the Internet in the past 18 months, was well compiled fake news for attracting views. These deliberate disinformation made me extremely upset because they portrayed me as the person that I can never become. I was a mere symbol rather than a real person in all such news coverage. I also sent an email to Today's Headlines in this regard, specifically asking for reviewing and filtering out these rumors before such news is aired publically. However, it didn't help much and these news were just out of sight for only a month. As a result, I think all I can do is to seek help from regulatory authorities and continue to reserve the right to pursue relevant legal responsibilities.

These unauthorized news and videos depicted me as an omnipotent person and got Jack Ma involved. Undoubtedly, I think he can always sleep tight whether I am in Alibaba or not. Certainly, it is the joint efforts of thousands of engineers that guarantee the security of Alibaba Cloud, and my personal contribution is very meager.

The small achievements I have made are not worth discussing. I think 99% of the readers do not accurately understand why I was named in the MIT TR35 in 2017. Everyone just witnessed the excitement and applauded. However, I do not need everyone to applaud for me, but I truly hope that everyone applauds for my work. This is also one of the reasons, why I have almost ceased publishing articles since 2014. I hope you will remember my work and my contribution to society rather than my personal growth experience. Lately, my experience has gained more attention than my work. I am still very dissatisfied with myself and want to work even harder.

In the second half of 2017, I switched my job domain from network security to artificial intelligence (AI). My team has built many key infrastructure systems in Zhejiang, Shanghai, and Chongqing. In particular, I devoted all my efforts to the work in Shanghai in 2018, and the results turned out to be satisfactory. These things were not publicized. What astonishes me is that the newsmakers focused on the hoaxes; have not even figured out what I am really proud of and how my work matters.

So I decided to release an article today. As an engineer, I want to put down my judgment on the future in the hope of helping peers to avoid detours. All the views and opinions in the article are personal and do not represent those of Alibaba.

Science and Technology Progress Aims to Liberate Productive Forces

I divide the development of productive forces into five phases: manual labor, mechanization, electrification, informatization, and intelligence. Each advancement in science and technology leads to the liberation of these productive forces and brings about huge changes in society.

The second scientific and technological revolution happened 140 years ago and made electric power available across all industries and fields. Since the power infrastructure was built by adopting key technologies such as the central power station and alternating current (AC) transformer, the cost of generating electric power was gradually reduced. Subsequently, various electrical applications emerged, and people got new and stable energy.

Now, we know that electric power was first applied to telephones, telegrams, and electric lights. But, It was the demand for electrical lighting that boosted the development of power infrastructure. The primary reason why electric power was dedicated to this single purpose was that there was no dazzling array of electrical appliances at that time. Electrical lighting was also the focus of the competition between the Edison General Electric Company and the Westinghouse Electric Corporation, 100 years ago. It is difficult to say whether a light bulb or a power station was more important than the other in the competition.

I once used a metaphor to describe the cloud computing dilemma. We have a computing infrastructure with a single cluster of tens of thousands of servers, which is the "central power station." However, the "light bulbs" were still missing. We are now using the "central power station" to power the "gas lights." Most of the businesses hosted on cloud computing today are still information systems. However, in ideal conditions, what consumes a lot of computing power should be intelligent systems. They are just the "light bulbs" for cloud computing, which we have always been searching for without gaining any results.

Here, it's important to clarify the differences between an information system and an intelligent system. In my opinion, the essence of an information system is the ability to edit databases. If a service system involves a large amount of manual interaction and relies on manually submitted forms to complete tasks, it is then referred to as an information system. On the contrary, an intelligent system, in my ideal opinion, should aim to automate tasks. The intelligent system inputs a task and then outputs results. The intermediate process is highly automated by machines. The intelligence level of an intelligent system is evaluated according to the complexity of tasks that it can complete.

From this point of view, smartphones are not intelligent. Instead, they are still information systems. In addition, various intelligent systems available in the market are just named as "intelligent". I do not imply that information systems have no value. Instead, information systems are of great value, but they are certainly not apt for the next era. With the rapid development of computing technologies, various information systems have emerged and greatly reshaped the world. It is also an important step to complete the transformation and up-gradation from electrification to informatization to witness how various computer systems are applied in every field for helping people to manage work and provide services more efficiently.

In this transitioning process, the Internet provides a magnifying effect. In my opinion, the Internet itself is not a productive force but the one that connects thousands of information systems in a manner that creates a scale effect. The Internet supports economies of scale. It magnifies the value of a system by thousands or tens of thousands of times. However, productivity is provided by the information system itself. Terminals that connect to the Internet are browsers, iOS systems, and Android systems, and the evolution of these terminals is critical. For example, through the Internet, Baidu connects people with information, Tencent connects people with other people, and Alibaba connects people with information services. Unfortunately, all these are not for the next era.

The first thing to occur in the next era will be the upgrading of information systems to intelligent systems. Companies that seamlessly connect all intelligent systems through the Internet will emerge. The upgrading from informatization to intelligence will bring huge progress across productive forces, and commercial corporations that adjust to such social changes will only survive to justify the survival of the fittest. From a historical point of view, it was an overwhelming process to upgrade the PC operating system to the mobile operating system in the information age. After the release of the iPhone, developers turned to work on iOS software. They no longer develop software for Microsoft Windows, and thereby gave Microsoft a strong blow. Microsoft was likely not be able to maintain the spot if it did not seize the opportunity of cloud computing. From the perspective of business development, similar events are bound to happen, and it won't be surprising if the giants in the information age may become insignificant after just one reform of productive forces.

Now the question arises, what kind of intelligent system does the world need in the future?

Computer scientists across the globe strived to make machines intelligent. Initially, simple expert systems used to rely on experiences and rules to handle simple tasks. However, machines did not know how to handle exceptions that were beyond expert experiences. As a result, data-driven intelligence emerged.

When machines become intelligent to some extent, they tend to process relatively simple tasks and partially liberate manpower. Increasing machines is equivalent to increasing manpower. Both, machine intelligence and human intelligence have their own strengths. Machines have a large amount of computing power and are tireless, which makes refined management of many jobs possible and often results in cost savings.

For example, bus scheduling was planned based on experiences and the number of buses was set for a single route in the past. However, if citizens' travel condition varies, the supply-demand relationship of buses changes. Some routes will be busy, while some others will be idle, which results in a waste of resources. To solve this problem, we need to first calculate the precise number of passengers on each bus trip, and then rely on machine intelligence to schedule buses on different routes. In this way, optimal efficiency can be achieved with the same resources. Therefore, the benefits of using machine intelligence are obvious.

Large-scale Machine Intelligence System Cannot Be Built Five Years Ago

It is evident that the upgrading from informatization to intelligence is arising with the development of productive forces. It won't be incorrect to say that now is the eve of the outbreak. it benefits from the maturity of four key technologies: cloud computing, big data, Internet of things (IoT), and network connection technologies.

We know that the current development of machine intelligence attributes to the study of brain science and the advancement of computing power. Neural networks have evolved with deep learning and have made major breakthroughs in the fields of vision and speech. There is no doubt about the importance of computing power. However, it is difficult for us to succeed in practice with computing power alone. The maturity of the other four technologies is also mandatory.

In the current technical environment, cloud computing, as the computing power infrastructure, provides enough computing power for intelligence. Big data technologies provide data processing methodologies and tools, which act as data infrastructure (there is no monopolized data infrastructure at present, and fragmentation is popular). IoT technology reduces the cost of intelligent devices to a level low enough to provide a foundation for the deployment of rich neuron sensing devices. Network connection technologies, ranging from 4G to 5G, lay an important foundation for high-speed data transmission.

If we can name a technology tree, the large-scale application of machine intelligence requires the four technologies lighted up at the beginning, as they form the foundation ground. Five years ago, the cost of the four technologies was the main bottleneck for any large-scale application of intelligent technologies. But, these technologies have gradually matured today.

Whenever a new technology emerges, the following two problems often arise:

  • The Scarcity of Talent: As we know, a Ph.D. graduate who understands deep learning or other machine intelligence technologies may earn an annual salary similar to a programmer with 10+ years of work experience. It is important to acknowledge that machine intelligence is required everywhere in the industry and is in short supply.
  • The Cost of Technologies: The cost of new technologies supposed to be high. When cloud computing first appeared, it also focused on the capability first and then the efficiency. According to some estimates, cloud computing servers has seen a price reduction for 57 times since its launch. According to the well-known Moore's law, the computing performance doubles every 18 months, which implies that the cost of the hardware with the same computing power will be halved every 18 months. Machine intelligence also follows a similar pattern and is expected to cost higher initially. It takes time to popularize new technologies. But, at the same time, it is not an ideal thing to simply wait for the prices to go down.

Both problems determine that machine intelligence should be first applied to the point where the greatest social efficiency is generated. It is important to find such a business scenario to bring the technology out from the laboratory to society. With continuous nourishment from business practices, technology can grow rapidly.

What Kind of Machine Intelligence System Does the World Need?

Both the problems mentioned above can be resolved effectively over time. However, challenges arise when we encounter problems persisting in the industry today. The real problems revolve around any of the following two aspects:

  • AI Should Not be a Separate Industry in the Future: We already have the incorrect industry classification today. For example, since the appearance of power infrastructure, electric power has become a key production factor because electric power is required in all industries and fields. In my opinion, intelligence is also a key production factor that every industry requires in the future. Therefore, there is no need to make AI a separate industry. If we have a separate AI industry, it will be a dilemma for the companies of that industry to determine their actual course of action. Additionally, these companies will be confused about their responsibilities. Eventually, the current retail industry pattern should be followed, where, each retailer has an e-commerce department that uses the Internet for marketing and sales. In the future, each enterprise should also have a dedicated department that is responsible for the development and training of its intelligent systems. The department should strive to train systems like training pets to make them intelligent. This is not just recommended for an AI company, but for each company.
  • Development Challenges of Machine Intelligence Technologies. The recent growing popularity of machine intelligence begins with deep learning resulting in great breakthroughs in the fields of vision and speech fields. Consequently, commercial enterprises will always be engaged in jobs such as vision, speech, and natural language processing. However, we should never forget that the complete human brain intelligence is from perception to action. In addition to perception and action, obtaining intelligence also requires continuous feedback to complete high-frequency collaboration.

Focusing only on perception is a big misunderstanding and is technically worthwhile. However, in the business terms, the social value it creates is very limited as it only has a relatively restricted contribution to the liberation of productive forces.

To assess the social value of an intelligent system with reference to the development of productive forces, we should recognize how much it contributes to the liberation of productive forces. Focusing only on perception does not create values. After completing multiple large projects, I discover that values are always created during the processing phase. Therefore, it is difficult to clarify whether the input and output values are worthwhile when the focus is just on perception. However, once the action phase begins, productive forces will be liberated and values are quantified. The "action" here means that machine intelligence schedules manpower or other devices.

Actually, from the perspective of technological development, we had the ability to make machine intelligence capable of making decisions a long while ago. For example, search engines and personalized recommendations are typical decision-making instances implemented by machine intelligence. By processing massive data every day, machine intelligence can eventually implement refined matching.

Therefore, I think a complete intelligent system consists of perception and action. The action phase is supported by the decision-making and scheduling technologies. The value of an intelligent system is measured in terms of how much it contributes to the liberation of productive forces.

Unfortunately, I do not think there is an ideal intelligent system in the industry so far. I refer to the current status of the industry as "with intelligence but without systems." Many AI startups possess some intelligence capabilities, such as a few technologies in vision, speech, Natural Language Processing (NLP), knowledge graph, search, and recommendation. However, only a few companies have complete technology stacks. Although companies like Baidu, Alibaba, and Tencent (BAT) have complete technology stacks, they are yet to integrate all the technologies into a complete system of perception and action. In other words, these technologies exist as fragments. There are very few examples of applying complete technology stack to a particular scenario for solving a specific problem. This is the key factor that leads to the emergence of an intelligent system and is also an engineering problem. The challenge for engineering is to integrate all intelligent technologies to achieve complete perception and action capabilities. Meanwhile, we can effectively control costs and provide developer-friendly interfaces.

From the perspective of intelligence technology, self-driving and smart speaker are two scenarios with complete closed loops from perception to action. I think we can use both scenarios to improve machine intelligence technologies, but it is difficult to make them thriving in business at present. Self-driving frees all drivers and obviously boosts the liberation of productive forces. However, it is of little significance to old cities due to the restriction of the current urban road infrastructure, as the urban roads are not designed for self-driving, and it becomes difficult for them to accommodate self-driving vehicles. Therefore, self-driving is more applicable to aviation, navigation, logistics, and other fields. This narrows down the scope for business applications. A smart speaker integrates a number of machine intelligence technologies, chatbots, the holy grail in the AI field is its core technology. As a result, it is extremely difficult to make a smart speaker remarkably intelligent. Additionally, at present smart speakers, play an extremely limited role in the liberation of productive forces in various family tasks. Thus, it is difficult to clarify its value. Nonetheless, both technologies will gradually reveal their vitality as the infrastructure upgrades over time.

Using the aviation industry metaphor, we can compare the current intelligence technologies to parts required for building an airplane. Many parts and engines are already available in the market. However, all the manufacturers are selling parts as final products, namely airplanes. Customers think that they have bought an airplane, while actually they have bought some parts, which contribute little to the liberation of productive forces. The real difficulty today is that we even do not have "airplane" design drawings. So, I plan to draw one and try to make an "airplane."

Building the Intelligence Era

For an airplane to take off, several elements are needed, the first one being, the pilots. Pilots do not necessarily know how to build an airplane. However, they must know how to drive an airplane to help people travel by air. I believe the employees working in the intelligence departments of the future enterprises are just like pilots. They are responsible for training purchased intelligent systems to make them remarkably intelligent. The final intelligent levels of the systems may vary due to enterprise data differences, as well as the technical levels and sense of responsibility of the "pilots."

Channels will definitely be the second element. In my opinion, channels are still offered by infrastructure providers, including operators and cloud computing vendors.

The last element is the airport. The airport is responsible for scheduling and coordinating all flights and provides services for all airplanes. This is the most interesting part. I think the "airport" is a real business model like the App Store of Apple Inc. In my opinion, the most important task of the "airport" in the intelligent era is to provide services to machine intelligence systems rather than to people.

Imagine that 70% to 80% of the future population on the Internet will be machine intelligence systems. They will handle the majority of jobs, and each machine intelligence system will have a separate owner. The owner might be an individual or an organization and will have sovereignty over the machine intelligence system. As each machine intelligence system will exist to complete one or more specific tasks, providing services for all machine intelligence systems will be a huge business model.

Key Path to the Future: The Automatic Coordination between Machine Intelligence Systems

Meanwhile, I also believe that in the current machine intelligence industry, too much emphasis is placed on man-machine interaction. While machine-machine interaction has been ignored. However, the latter is of greater importance. Man-machine interaction is still regarding information systems. The automatic collaboration between machines further enlarges the values of intelligent systems on a large scale.

Therefore, for smarter future, it is necessary to define a language for all machine intelligence systems. Machines should communicate by using their own languages just like human beings and implement simple logic operations. The interaction and collaboration between all machine intelligence systems should not require any manual intervention. It is just like the scenario that your children and the neighbor's children play on their own. You do not need to intervene in their communication, and they will get what they need to complete their own tasks.

Let's take the example of the One Netcom Office platform. The current mainstream One Netcom Office platform mainly commences its workflow starting from combining data from all relevant commissions, offices, and bureaus of the government. Then, based on the converged data, the platform implements data governance, sorts out process flows, and restructures service patterns. The idea of such big data application remains on the road of informatization construction. To promote the implementation of new technologies, we must reform the process flows first. However, the high degree of customization in different regions makes it very difficult to implement large-scale productization nationwide. These are also the demerits of big data applications. Actually, we can adopt an intelligent construction idea. Each commission, office, and, bureau of the government can build a machine intelligence system on its own. The task of the machine intelligence system will be to handle window services instead of civil servants. After a citizen submits an application, the citizen is authenticated. Then, the machine intelligence system of the commission, office, or bureau will send a coordination request to the machine intelligence systems of other commissions, offices, or bureaus based on the required materials. After several rounds of machine intelligence communication and collaboration, the citizen can quickly get the expected results. This automatic collaboration among multiple machine intelligence systems significantly reduces the impact on process flows.

Majorly, it is the collaboration between services rather than the exchange of data. Therefore, even though the interaction and collaboration between machine intelligence systems must be done through the network, the security is controllable. Since each machine intelligence system has its owner, and all the training processes also occur within the respective main body, the data exchange and sharing are not necessary. The owner can set what a machine intelligence system can and cannot produce as an output. All security control measures are done within the intelligent system. In case, the intelligent system is connected to the Internet to collaborate with other machine intelligence systems or use the services provided by the "airport", the default untrusted mode applies.

Whether deploying the machine intelligence systems on the public cloud or on a private cloud is not a major issue. Owners can deploy the systems wherever they prefer. As even today, cloud computing is still in danger of being pipelined just as the operators are pipelined by Internet content providers. Cloud computing vendors may also be pipelined by intelligence vendors in the future because cloud computing and big data are not intelligent.

Group A

To put the above-stated ideas into practice, I was appointed to set up Group A for Alibaba Cloud. The mission of Group A is to build a machine intelligent system on the same lines and speed up the advent of the intelligence era.

In my opinion, this is a task that requires the whole society to work together for the next 30 to 50 years, similarly as we have made all our efforts in developing informatization over the last 30 to 50 years.

Get to know our core technologies and latest product updates from Alibaba's top senior experts on our Tech Show series

0 0 0
Share on

Alibaba Clouder

2,606 posts | 737 followers

You may also like


Alibaba Clouder

2,606 posts | 737 followers

Related Products