This article is published by netease Cloud.


Author: Liu Chao

At present, the word “Internet” has been generalized, which is probably attributed to Luo Zhenyu’s “Luo Ji” thinking to vigorously promote the concept of Internet. It seems that enterprises cannot open their mouths without adding the word “Internet”. So the term Internet + is everywhere.

An important force supporting the Internetization is cloud computing. But now the cloud has been extended to do not know where to go, what financial cloud, invoice cloud and so on, let us who do computing, network, storage, middleware, big data, and these can not be clearly distinguished from the extension.

Now artificial intelligence is hot, but also joined in the category of cloud computing, so all kinds of intelligent cloud come out, intelligent education, intelligent medical treatment, intelligent transportation and so on.

Those who work hard in traditional industries, helplessly see Internet companies make a year’s money in a single day, so they want to get on board one after another and set foot on the tide of Internet, cloud and wisdom.

Let’s take stock of the present situation and look forward to the future.

First, looking back at those popular Internet applications, the history is so similar

The Internet has all been hot. From the three earliest portal websites, search engines, social networking sites, e-commerce, microblog, group buying, mutual finance, video websites, O2O, live broadcasting, and bicycles, all of which are hotly pursued after Google’s withdrawal, wave after wave, the waves behind the Yangtze River push the waves before, and the waves before die on the beach.

They all go through this cycle.

Must first have a creative individual, to be a no product or invention of a business model, when the creative recognised by the market, welcomed the outbreak of the market, everyone flocked, do homogeneity product, an emerging markets into a red sea, in the red sea life and death battle, some out, some failed, some are mergers, In the end, only two or three remain, either becoming giants or being acquired by giants.

None of these waves have escaped the cycle. The only difference is that the cycle is getting faster and faster. The portal race has been going on for years. Startups are increasingly finding that each round of funding seems to get bigger and bigger, but it doesn’t take long.

The market is impatient, you don’t take aggressive marketing strategy, competitors will, you are cheap, they are cheaper, you are free, someone will dare to subsidize.

Investors have no patience, not in the short term to come up with proud results, you will not have the next round, so do not spend money, was crushed by the opponent’s momentum, no next round, is a dead, hard to spend money, the opponent crushed, before winning, no next round, or a dead.

Core employees don’t have the patience, if you want to have a rapid development, people have to find cows, the cows are often very expensive, will make you melt money soon, of course, some entrepreneurs can through vision and options to keep cattle man, but it is effective, people tend to social area wide, have an independent thinking, not so easy to fool live, people tend to like, Easy to steal, so in addition to painting the pie, you need to keep cashing in on some pie, such as your business to improve in a short period of time, or to another height.

So as the cycle gets faster and faster, there is less and less time for Internet companies, and the Internet companies that win tend to have the rapid speed of iteration to make an impact on traditional industries. Bosses in traditional industries want to gain the same competitive advantage by adding Internet capabilities. So what exactly is going Online?

Second, what is the Internet company, what is the AI company

In his lecture at AI Conference, Ng mentioned his understanding of what an Internet company is and what an AI company is, which I think is very profound.

(1) Shopping mall + website ≠ Internet company

If you are a traditional mall, just make a website, that is not called Internet.

Real identification of an Internet company, there are the following points:

  • A/B testing, let the data speak for itself: When you have A page that needs to be improved, how your website is designed, how your APP is designed, is it up to you to reward each layer and let the boss decide? Big boss often is very dangerous, because he doesn’t understand customers’ preferences, and the subjective preference, in the actual test, are often very different results, so don’t let the boss clappers, but let the data speak, through online test the results of the two kinds of schemes, finally come to the conclusion that, although the rapid ascension can’t do, but can ensure each time change, is positive.
  • Shorter cycles: Your application must iterate fast enough, and iteration should be based on data, taking small, fast steps based on data feedback, which requires the organization and process to be adaptable. Unlike traditional companies that upgrade every few months, Internet companies upgrade almost every day. When you open various apps, you will find that the interface changes at every turn.
  • Engineers and PMS make decisions: How can we get online quickly? If every time online are more than one hundred people to open a conference, let the boss to make a decision, it must not soon, should let the engineer and PM to make decisions, they are the ones who are really hear gunfire, you need to make them independent responsible for a piece of content, independent decisions, independent, independent responsible for all the PM work in parallel, makes updates fast.

(2) Traditional technology company + machine learning/neural network ≠ AI company

If you’re a traditional company with just one division using machine learning and neural networks, you’re not an AI company.

A real AI company should have the following characteristics:

  • AI companies tend to acquire data strategically: Data is the cornerstone of an AI company, so if you want to make your company smart, acquire data strategically. In fact, many companies have already done so, all those so-called electronic, everything all want to APP, don’t accept cash and offline purchase and payment, through a variety of terminal to infiltrate your life, you feel very convenient, open an APP to do this for a while, a while open another APP does other, your outline of the whole picture is clear. Therefore, you will often be subjected to a variety of advertisements, which shows that they have been big data, but not intelligent, accurate portrait, but not intelligent push, or the algorithm is a little dull, but the strategy, you can feel.
  • AI companies usually have a unified data warehouse: data should be connected. If you need to communicate with 50 departments to obtain the full amount of data in your company, it means that there is no big data, let alone intelligent. Your company should have a unified user center, a unified commodity center and a unified big data platform. No matter which product you use, your users should feel that the products are interconnected rather than isolated, which has requirements for the structure of the organization, system and data.

In AI companies, product managers need to learn to use data and demand accurate feedback when communicating with engineers. Product manager in addition to the customer thinking, thinking, there should be a data like Wu En says, be a robot of artificial dialogue, draw a box, the product manager in a small, there is no point, want to know where the data come from, how to get these data, how to make use of these data, how to quantitative evaluation of the effect of the robot of artificial dialogue if there is a promotion, It doesn’t have to be emotional, it has to have data.

Three, the traditional Internet companies should consider the three major structures

To get your company online as quickly as possible, you need to think about three major structures.

  • The first is IT architecture, how to reduce CAPEX and OPEX, how to reduce the burden of operation and maintenance. IT architecture is getting more and more complex, and IT is more and more expensive to be able to play an IT architecture well. If you are not an IT professional, but a financial, manufacturing, medical, you will find that the cost of this piece is getting bigger and bigger. Cloud is an inevitable trend, the so-called professional people do professional things.
  • The second is the application architecture, how to achieve rapid iteration, how to resist the high concurrent traffic after the Internet. Servitization is a necessary trend. With each small team responsible for a separate service, iteration is faster. Each module is an independent service that can be independently extended to withstand high concurrency from the Internet. For example, when traditional financial payment meets Internet payment, the frequency is suddenly N times higher, and it cannot be handled without servitization.
  • The third is data architecture, how to form a unified data platform and give data for digital operation. Another advantage of servitization is that similar data can be centralized. With the data, how to operate, provide accurate recommendations to customers, realize big data and further realize intelligence must be considered.

IT architecture Trend 1: From resource flexibility to architecture flexibility, the gap between Internet companies and traditional companies is getting bigger and bigger

At present, cloud has become a consensus in the industry, but the current state of cloud computing is still in the cloud computing 1.0 period, that is, only the realization of resource level flexibility.

What is elasticity?

Say popular point is flexibility, mainly contain two aspects of flexibility, one is time flexibility, that is, want when to want when to want, one is space flexibility, that is, want how much, how much.

People go cloud because physical machines don’t have that kind of flexibility. Physical machine is not flexible in four aspects, one is not flexible purchasing, procurement short period is a week or two, long a month or two, two is not flexible granularity, unable to purchase a nuclear 1 g1m bandwidth machine, 3 it is to reuse is not flexible, the same physical machine, one used, it is difficult to recover into shape, let another person to use, unless the reshipment system, four is not flexible operations, A lot of things to go to the machine room to fix, and the machine room is often in the most partial place.

Virtualization technology solves this problem, point can solve the procurement flexibility problem, can be large or small to solve the problem of granularity flexibility, that is, to solve the problem of reuse flexibility, interface operation and maintenance to solve the problem of operation and maintenance flexibility. The only unsolved problem is scheduling flexibility, which requires manual operation and maintenance, so the scale is limited.

The scheduling technology of cloud computing makes the cluster scale very large, and can be automatically scheduled, so as to truly achieve resource elasticity from the perspective of customers.

But what if cloud computing doesn’t matter? For example, normally we only need 10 virtual machines, but in the double 11 seconds kill scenario, we need 100 virtual machines. With the flexibility of cloud computing resources, we only need a point on the cloud platform, and soon 90 virtual machines will be created. But what about the applications inside?

It still requires our operation and maintenance personnel to install it one by one, which is too complicated to achieve true flexibility.

So we started thinking about how cloud computing could manage applications.

We divide the applications into two types, one is called universal applications, such as Spark, Mysql and Hadoop. These applications are universal and can be installed by anyone, but their operation and maintenance are relatively complex, so they should be handed over to cloud platforms for operation and expansion. The other is called your own apps, where the cloud platform can’t help you because only you know how to install your own apps, so the cloud platform often provides script-based tools to do this, such as Chef, Puppet, Ansible, etc. However, one of the biggest weaknesses of scripts is that it is very difficult for scripts to run successfully once the environment is different, which makes it very difficult for script-based installation tools to implement cross-cloud migration.

Hence the container.


Container is another word for Container. The idea of Container is to become a Container for software delivery. The characteristic of container, it is pack, 2 it is standard.

In the days before containers, it would take three ports and three changes to get goods from A to B. Each time the cargo had to be unloaded, disassembled, and then loaded onto the ship and put in order again. So when there were no containers, the crew had to stay ashore for several days each time they changed ships.

With the containers, all the cargo was packed together, and the containers were all the same size, so every time the ship changed, one box could be moved across the whole, which could be done by the hour, and the crew could no longer go ashore for a long time.

This is the container “packaging”, “standard” two characteristics in the application of life.

Deploying any application also includes a lot of loose ends, permissions, users, paths, configurations, application environments, etc. This is like a lot of piecemeal goods. If they are not packaged, they need to re-check each environment of development, testing and production to ensure the consistency of the environment, and sometimes they even need to re-build these environments, just like unloading and reloading the goods every time. A little difference in the middle, may lead to the failure of the program.

So how does a container package an application? Or to learn containers, first of all, there should be a closed environment, the goods will be encapsulated, so that the goods do not interfere with each other, isolated from each other, so that loading and unloading is convenient.

The other is how to standardize the container so that it can be shipped on any ship. The criteria here are the image and the running environment of the container.

The so-called mirror is the moment when you weld the container, and the state of the container is saved. Just like sun Wukong said, the container is fixed at that moment, and then the state of this moment is saved into a series of documents. The format of these documents is standard, and anyone who sees them can reconstruct the exact moment in time. The process of restoring the image to runtime (that is, reading the image file and restoring the moment in time) is the process of running the container.

With containers, cloud computing truly achieves the full flexibility of the application and resource layers.



To summarize briefly, although containers have become a trend from a technical point of view, in the real implementation process, cloud computing 1.0 era will apply the process of cloud, many traditional companies have not yet completed. However, at large Internet companies, container-based cloud computing 2.0 technologies that enable fully elastic scaling of architectures are already being used on a large scale. Container’s three platform battle has come to an end, Kubernetes has won, now in the Internet company’s forum, to talk about container technology, has felt very old-fashioned technology, but in many traditional enterprises, cloud has not been completed, this gap is huge, using the rapid iteration of container technology, Would be a dimension reduction blow for companies that haven’t gone cloud yet.

Five, IT architecture trend two: cloud containerization is irreversible, IT personnel are becoming more and more expensive, only large scale can reduce the cost, through automation and intelligence, let a small number of high-paid IT personnel to manage large-scale clusters

Some people will say, you always say that cloud, how good cloud, can reduce costs, but the real practice, although the cloud technology is constantly iterating, but my operation and maintenance costs, but also higher and higher.


Originally with the physical machine, in fact, I only need to understand the basic knowledge of Linux can operate and maintain, then you cheat me on virtualization software, virtualization software is very very expensive, can use virtualization software certificate more expensive. Later problems binding virtualization software, then you cheat me with open source virtualization software it, who knew there was no free, open source without adaptation and customization, basically can’t use in a production environment, but if once the custom, and turned into a private software, by this time my operations staff, not only should use virtualization software, Also need to develop virtualization software, can develop virtualization software is very expensive, and if their operations it is not even half one hundred people team, simply do not, then you should use a container, but after using the container, calculation of network storage technologies cannot little, but also to understand the application of the whole stack engineers are more expensive.

Can cloud computing really reduce costs? Did I use fake cloud computing?

Let’s take a closer look at how cloud computing can be used.

The simplest way to use it, of course, is to use the public cloud, which is equivalent to buying an apartment. The cost is very low. At the same time, the isolation is logical.

If you don’t want to compete for resources, you can use a dedicated host, which is equivalent to buying a whole building. All the virtual machines on this physical machine are yours, while many PaaS services such as databases are still on the common cluster.

So there are a lot of companies that want to build their own private cloud, like living in a villa, but in fact, it is a rural home, because you may need to rent a piece of land in someone else’s data center, and then either build their own cloud platform, or bid for cloud platform. No one would think that when living in a villa, they take the land to build it by themselves, and repair it themselves if there is a problem. In this way, the cost of deployment operation and maintenance update is very high, and you need to raise a professional operation and maintenance team to do this.

This approach often only solves the psychological security problem, is it really safe? Not necessarily. Just like self-built houses in rural areas without perfect property and security, it is difficult to ensure that the cloud platform hosted in other people’s data center will not be threatened by the unprofessional operation and maintenance behavior on the neighbor’s rack. You build the cloud platform, the neighbor rack with physical machine directly, we often see some small company in company a VPN, office network and computer room of the physical machine connected to office network wifi is very unsafe, if a hacker can reach the neighbor’s physical machine, is likely to black to your physical machine above.

Unless you are a big millionaire, such as telecom, banking and electricity, you can own the land and build the data center by yourself. You have a operation and maintenance team with hundreds of people and a very perfect machine room operation and maintenance process, which is equivalent to Trump taking a piece of land to build his own villa, recruiting property and hiring security guards. Of course, it is the safest. Most traditional enterprises may not enjoy such treatment.

Netease cloud exclusive cloud service, commercial villa cloud service. Land netease to take, house netease to cover, property netease to manage, security netease to recruit. Equivalent in netease’s room, and draw a piece of place, this place above the deployment of cloud resources are all your own, but the hardware, network, security, installation, upgrades, repair, you don’t have to tube, there is a very professional operations team to do this, these operations ensure safety compliance operation, you only need to pay attention to your own application.

Why should netease launch exclusive cloud? We believe that the trend of cloud containerization is certain, and there is nothing wrong with the road described above. It is inevitable that technical talents who can understand OpenStack and Docker will become more and more expensive. Only with increasing scale can the cost advantage of cloud be reflected. Such as Google, Google inside operations engineer is quite expensive, there are hundreds of thousands, even millions of dollars in annual salary, they master the most advanced technology is no problem, but they will through all kinds of automation, intelligent technology, even managing global millions of machines, such cost stand down, it is not very high, so Google fully sustainable. If you just operate and maintain a cloud platform with dozens of nodes or at most hundreds of nodes, you also need to hire some people who are so expensive, ordinary enterprises can’t stand IT, so for most enterprises, they should hand over the IT architecture to the most professional talents.

Personally, I even think that there will only be two cloud platforms in the future, one is large-scale public cloud platform and the other is self-made cloud platform of local tyrants. The other forms will disappear with more and more advanced technologies, higher requirements of operation and maintenance personnel, and more and more expensive annual salary of operation and maintenance personnel.

Vi. Business Architecture Trend 1: The impact of the Internet has become inevitable, and rapid change has become the core competitiveness. DevOps restructuring of organizational structure, process and culture is an inevitable choice

In terms of business architecture, I emphasize servitization here. Is how to avoid the black swan of development and operations.

The so-called black swan is a problem that we have not encountered before. We imagined that the application would not change, so that the stability has become a bubble because of the impact of the Internet.

The Internet requires your application to change fast, so you have to change. Once your traditional application, such as traditional card payment, is connected to an Internet platform, such as online payment, the throughput you face is many times greater than before. This is a new problem, and it is more risky than not to change.

So, as you can see in the graph above, because of the rapid change, we have to break down the application into microservices, and each module is iterated independently and released independently, so that we can cope with the change. Because can not carry, so we still have to use micro services, originally a program to carry, now multiple applications to carry, it is possible to carry.

However, with the disassembly of micro-services, there are more services and more versions, and operation and maintenance becomes a big problem. In general, traditional companies, the proportion of development and operation and maintenance is quite high, so the operation and maintenance pressure is very great, and it will be very unstable.

Therefore, at this time, containers should be used to advance the deployment of the environment to the development stage, so that developers can not do hand in hand, but from the time of developing the code, we should care about the configuration of the environment and generate container images.

This time, although each development more than some work, but each module developers, just to maintain their module container mirror, workload is not particularly big, and if all the deployment environment all in the hands of a few operations, is very easy to get wrong, this is the equivalent of more than each development as much as 5% of the workload, Thus reducing the workload of operation and maintenance personnel by 200%.

Some people say that operations will be happy, but development will not want to mirror, so it needs DevOps culture to break down the wall between development and operations to speed up iteration and ensure the stability of the system.

Netease has a quality management platform dedicated to DevOps to get the whole process through.


There is also a fault drilling platform, is the use of deliberate way, the simulation of some system errors, thus the stability of the test system and fault tolerance, only after the fault drilling system is not a problem, is the true stability, not moving it, because you don’t touch it, don’t know why, suddenly he hung up the whole system is not available, if you often fault, The way to avoid black swans is to keep the system usable during the drill.

Therefore, the impact of the Internet has shattered the illusion that constant is equal to stability. Rapid change has become the core competitiveness, and DevOps culture has become an inevitable choice. This is what we often say: high frequency beats low frequency.

7. Business Architecture Trend 2: Service to form a capability reuse center, quickly launch products, open up the data platform, and occupy the high-end of the new industrial biological chain

The second effect of servitization is to form a hub for reuse of capabilities.

If the chimney system is built like the original, information will be independent of each other, and every time a new application is developed, the cost will be very high, each system has its own user management, each system has its own product management. Actually there should be a user center, has a commodity center, and the user service center and goods all make it, so that when you want to develop a new system, you as long as the center of the calling user interface is ok, don’t need to develop a user center system, the ability to do this when the center of the reuse more and more, The new apps you create get thinner and lighter and faster and faster. In this way, information is also interchangeable. A user listens to music, buys goods, takes a bus, buys seafood, takes online courses, and so a complete portrait of the user can be drawn.

Only data get through, data can become one of your core competitiveness, only the ability to reuse, to launch new products as soon as possible.

Now the so-called New retail, and fourth Party logistics, is to build such data centers and capacity centers. The platform will not do the most difficult and tiring work, such as the delivery of express goods. However, how to circulate logistics, where to send goods, and how to distribute warehousing have been completed in the capacity center and data center. New retail is the same, your supply chain, in the end should be what goods, how much, when and where to send, these unified data center and ability center will help you calculate, you just honestly open your own store can. So, no matter you are a shop, or a terminal logistics, your throat in the hands of others, hard work are you to stem, money is others earn.

Data Architecture Trend 1: Strategic data collection, integration, and feedback are the foundation for companies to become competitors at the poker table in the age of AI

The frequency and real-time of data collection and data application is a big difference between Internet companies and traditional companies. Many traditional companies have their own operating platform, but its operations from the terminal to collect data, and then use the Excel form to arrange, once every two weeks to do consolidation, written report, and then feedback to the top, top according to the report, to make the next phase of the operation strategy, so as to guide the next phase of the operation.

This mode of operation is no problem in traditional industries, but it is impossible for Internet companies. You might imagine, for a home appliance business, and 618 pairs and how important it is, the turnover of most of the year is done in these two days, and the key is in the morning after four hours, in the four pairs of 11 in the morning hours, all operations at the scene, they looked at the big data platform of real time data, It is necessary to launch a marketing strategy quickly to ensure a big sale within these four hours. If it is not completed within these four hours, it is likely that it will fail to meet its annual performance target. Such real-time and sensitivity is unimaginable in traditional industries.

So strategic data integration, is a premise of digital operation, so that your data will be collected all of the terminal, there are transaction data, can be stored in a database, there are also burial site to browse data, can be placed in the log repository, as well as customer service data, also can be collected, in a big data platform, Conduct unified analysis and feed back to operations in real time through BI.

There is a very popular program this year, luo Yonghao and Luo Zhenyu had a nine-hour long talk, in which Luo Yonghao talked about a card table theory, that is, he wants to engage in the largest computing platform — mobile phone, so that when the next wave, let alone win, at least can be on the card table. The same applies to AI, not to mention winning in the AGE of AI, strategically collecting data, real-time feedback data, and becoming a competitor at the poker table for AI companies.

Data architecture Trend ii: Although artificial intelligence is in full swing, there are many concepts but few implementation, artificial intelligence model has poor universality and small trial range, which requires scenario-based implementation. General artificial intelligence has a long way to go.

Artificial intelligence mainly goes through three stages. In the first stage, we call it the expert system stage, hoping that the experts can summarize some knowledge and tell the machine. But knowledge of this matter, ordinary people may not be able to do, may be experts can, such as language experts, or financial experts. Can knowledge of language and finance be expressed as a slightly more rigorous mathematical formula? For example, a language expert might summarize grammatical rules such as subject, predicate, object, definite form, complement, the subject must be followed by the predicate, the predicate must be followed by the object, summarize these rules and strictly express them soon ok? And it turns out that this is not going to work. It’s too hard to summarize. Take the subject, predicate and object example, many times in the spoken language, the predicate is omitted, others ask: who are you? I replied: I liu Chao. But you can’t require the machine to speak standard written language for speech semantic recognition, which is still not smart enough. As Luo Yonghao said in a speech, every time he looked into the phone, he said in written language: Please call x and X for me, this is a very embarrassing thing.

This stage of artificial intelligence is called expert systems. Expert systems are not easy to succeed, on the one hand, knowledge is difficult to sum up, on the other hand, the knowledge summed up is difficult to teach computers. How can you program it to a computer if you can’t tell it, because you seem to think there’s a pattern?

So people thought, well, it looks like machines are a completely different species from humans, so let the machines learn by themselves. How does a machine learn? Since the machine is so powerful in statistics, it must be able to find certain patterns in large numbers based on statistical learning.

In fact, there is a good example in the entertainment industry, visible.

A netizen has calculated the lyrics of 117 songs from nine albums released by well-known singers in the Mainland. The same word appears only once in a song, and the top ten adjectives, nouns and verbs are shown in the table below (the number after the word is the number of occurrences) :

What if we wrote a random string of numbers and took one word from the adjective, noun and verb in order of the digits and joined them together?

For example, take PI 3.1415926, the corresponding words are: strong, road, fly, freedom, rain, buried, lost. Connect and touch up a little:

Tough kids,


Still on the road,


Spread your wings and fly to freedom,


Let the rain bury his confusion.

Do you feel something? Of course, true statistics-based learning algorithms are much more complex than this simple statistic.

Statistical learning is easier to understand a simple correlation, however, such as one word and another word always appear together, the two words should have a relationship, to express complex correlation, and the formula of statistical methods are often very complex, in order to simplify the calculation, often make a variety of independence assumption, to reduce degree of difficulty of calculation formula, in real life, however, Independent events are relatively rare.

So people began to reflect on how the human world works from the world of machines.

Human brain is not store a lot of rules, also is not a record of a lot of statistical data, but by neurons trigger the implementation, each neuron has input from other neurons, when receives the input, will produce an output to stimulate other neurons, and a large number of neurons interactions, eventually form the results of the various output. For example, when people see a beautiful woman with dilated pupils, it is not the brain making a regular judgment based on the proportion of body size, nor counting all the beautiful women they have seen in their life. Instead, neurons fire from the retina to the brain and back to the pupils. In this process, it’s hard to figure out what role each neuron played in the final result, but it did.

So people started to model neurons with a mathematical unit:

This neuron has input and output, and the input and output are represented by a formula. The input affects the output according to its importance (weight).

So n neurons can be connected together like a neural network, the number n can be very, very large, all the neurons can be divided into many columns, and each column can have many columns, and each neuron can have different weights on the input, and therefore each neuron has a different formula. When people input something from this network, they want to output a result that is correct for human beings. For example, in the above example, input a picture with the word “2”, the second number in the output list is the largest. Actually, from the perspective of the machine, it does not know the input picture with the word “2”, nor does it know the meaning of the output series of numbers. It does not matter, human knows the meaning. As for neurons, they don’t know that the retina is seeing a beautiful woman, nor do they know that the pupils dilate to see clearly, but when they see a beautiful woman, the pupils dilate, and that’s enough.

For any neural network, no one can guarantee that the input is 2, and the output is always the second largest number. To guarantee this result requires training and learning. After all, the dilated pupils of beautiful women are the result of years of evolution. The process of learning is to input a lot of images and adjust them if the result is not what you want. How to adjust it is that each weight of each neuron is fine-tuned to the target. Because there are too many neurons and weights, it is difficult for the results generated by the whole network to show an either-or result. Instead, it makes slight progress towards the result and finally achieves the target result. Of course, these adjustment strategies are still very skillful, need algorithm master to carefully adjust. Just as the human sees the beauty, the pupil does not enlarge to be able to see clearly at the beginning, so the beauty runs away with others, the result of the next study is that the pupil dilates a little, rather than dilate the nostrils.

It doesn’t sound that logical, but it can be done, just so willful.

The universal theorem of neural networks says, suppose someone gives you some complicated fancy function, f(x) :


Whatever this function looks like, there is always a neural network that can take any possible input x, and its value f(x) (or some exact approximation) is the output of the neural network.

If the function represents a law, it also means that the law, no matter how wonderful, no matter how incomprehensible, can be represented by a large number of neurons, by a large number of weights.

This reminds me of economics, so it’s easier to understand.

We think of each neuron as an economically active individual in a society. Therefore, the neural network is equivalent to the entire economy and society, and each neuron adjusts its weight to the input of the society and makes corresponding output, for example, when the salary rises, the vegetable price rises, and the stock falls, what should I do? How should I spend my own money? Isn’t there a pattern? Sure, but what’s the pattern? It’s hard to tell.

The economy based on expert system belongs to planned economy, and the expression of the whole economic law is not expected to be expressed through the independent decision of each economic individual, but to be summarized through the strategic view and farsighted knowledge of experts. Experts never know which city street lacks a sweet tofu vendor. Therefore, experts say that how much steel and steamed bread should be produced is often far from the real needs of people’s lives. Even if the whole plan is written in hundreds of pages, it cannot express the small laws hidden in people’s lives.

Macroeconomic regulation based on statistics is much more reliable. Every year, the Bureau of Statistics will make statistics on the employment rate, inflation rate, GDP and other indicators of the whole society. These indicators often represent many internal laws, although they cannot be accurately expressed, but they are relatively reliable. However, the summary expression based on the statistical law is relatively rough. For example, economists can conclude whether the housing price will rise or fall in the long run and whether the stock price will rise or fall in the long run when they see these statistical data. If the economy is on the rise, both the housing price and the stock price should rise. However, based on statistical data, it is impossible to summarize the law of slight fluctuations in stocks and prices.

Microeconomics based on neural network is the most accurate expression of the whole economic law. Everyone adjusts the input from the society, and the adjustment will also feed back to the society as input. Imagine the subtle curves of the stock market, which are the result of individual trades. There is no uniform pattern to follow. And each person makes independent decisions based on the input of the whole society. When certain factors are trained for many times, they will also form macroscopic statistical laws, which is what can be seen in macroeconomics. For example, every time a large amount of money is issued, housing prices will eventually rise, and after a lot of training, people will learn.

Netease has applied artificial intelligence, a powerful technology, to anti-spam work. Since netease launched its mailbox product in 1997, our anti-spam technology has been constantly evolving and upgrading, and has been successfully applied to various product lines of hundreds of millions of users, including video entertainment, games, social networking, e-commerce and other product lines. For example, Netease news, blog albums, cloud music, cloud reading, Youdao, BOBO, Kaola, games and other products. In general, anti-garbage technology has accumulated 19 years of practical experience in netease, and has been quietly escorting netease products behind. SaaS services are now open as cloud platforms.

Reviewing the development of netease’s anti-garbage technology, we can roughly divide it into three key stages, which basically correspond to the three stages of artificial intelligence development:

The first stage mainly relies on keywords, black-and-white lists and various filter technologies to do some content detection and interception, which is also the most basic stage. Due to the bottleneck of computing capacity and the development of algorithm theory at that time, the technology of the first stage can barely meet the use.

In the second phase, based on the computer industry has some update algorithm, such as bayesian filtering (algorithm) based on probability theory, some color, texture recognition and so on, these more excellent mature, we can do better on the basis of these algorithms of feature matching and technological transformation, achieve better effect of anti spam.

Finally, with advances in artificial intelligence algorithms and advances in computing power, anti-waste technology has evolved into a third stage: the stage of big data and artificial intelligence. We will use massive big data to conduct behavior analysis of users, make portraits of users, evaluate whether users are junk users or normal users, improve user experience by means of man-machine recognition, and understand semantic texts. There are also artificial intelligence-based image recognition technologies that can more accurately identify pornographic images, advertising images and some contraband images.


The second application of netease artificial intelligence is qiyu Fully intelligent cloud customer service, which has gradually developed from the keyword matching of the first generation to the probability-based NLP and the third generation based on deep learning neural network.

AlphaGo events, let artificial intelligence began in full swing, we find that many of the areas are like this, when Daniel didn’t come up with corresponding solutions, most of them are no way to see, since TensorFlow, greatly reduces the threshold of machine learning and artificial intelligence, so many companies began to come up, known as the artificial intelligence, There are often more concepts, less landing, and infinite extension.

Actually is the formation of deep learning model, and the applicable scope is very small, generality is relatively poor, often can only do a special thing, for example, if we have a lot of electricity data, we can use to do customer service, but the nature of the service model cannot be used for spam, we have a lot of email spam data, can train the detecting garbage data model, However, this model cannot be used to recommend music, so there is still a long way to go for general AI. Currently, it is necessary to find a very focused scene in the industry, so that AI can be applied as soon as possible.



Understand netease Cloud:

The official website of netease Cloud is www.163yun.com/

New user package: www.163yun.com/gift

Netease Cloud community: sq.163yun.com/