On the eve of the 2021 Shanghai auto show, the actual video of BAIC's polar fox Alpha S Huawei HI model with Huawei self-driving system ADS began to scan the screen online. The Huawei ADS in the video shows a high robustness of the algorithm in a busy city with complex traffic, just like a veteran driver who has been driving for many years.

The emergence of Huawei's autopilot production has attracted everyone's attention, causing as much heat as the Shanghai auto show two years ago, when Huawei made a low-key visit and announced a full-scale foray into the auto industry.

This seems to be the consistent style of Huawei. If you don't make a sound, it will be a blockbuster.

On April 16, the Star Channel came to Huawei Shanghai Research Institute to interview Su Qing, the head of Huawei ADS. Starting with Hayes Chip, Suqing led the development of Huawei Leonardo da Vinci AI chip architecture and is currently the president and chief architect of Huawei's BU smart driving product line.

Suqing is the kind of interviewee who can capture the media quickly. He is quick-thinking, speaks extremely fast, has a very forward-looking vision for self-driving, and is extremely confident, and is not hesitant to share his views on other companies in the industry-of course, he does not forget to add professionally: this is my personal opinion. It doesn't represent the position of the company.

We sorted out the actual records of Suqing interviewed by the quasi-Star Channel and other media, without a cut, Enjoyit.

About Alpha S Huawei HI Edition

Media: which cities does our system support now?

Suqing: maybe I'll briefly talk about the composition of several modes of the system first. It's not a simple Robotaxi, it has NCA, ICA and ICA+ modes. I think the questions you may ask and what you experience today are mainly NCA mode. NCA mode is completely automatic, a bit like the Robotaxi experience.

When we are in mass production at the end of this year, we will open four cities in the north, Shanghai, Guangzhou and Shenzhen, and we will open a batch of new cities about every three months. This is the experience of NCA. We can also see that after all, there are still many second-tier, third-tier and fourth-tier cities in China, and everyone has to buy a car and use it. At this time, we will provide an ICA+ model because our cars will learn the whole traffic environment and self-composition map by themselves.

As long as you have driven the car, or your neighbor has driven it, the car will automatically learn about the road conditions, it will make up the picture in real time, and then the car will reach a similar Robotaxi, but it will definitely be a little worse, because its data is not complete enough to achieve such an experience.

In particular, the more popular highway, including the inner ring, central ring and outer ring of Shanghai, do not need a map at all, and can achieve the same basic experience. Therefore, this thing can be generalized by the end of this year, and there is no problem.

Media: a city every three months?

Suqing: more than one, one batch every three months.

Media: what is the approximate volume in a quarter?

Suqing: I can't tell now, maybe six at the beginning? I'm just giving an example, about this level.

Media: what is the scope of the test now?

Suqing: the whole country is already in generalization, and the first-and second-tier cities are all in generalization.

Media: can you tell me all the hardware configurations?

Suqing: we have two configurations, the standard version is the computing power of 400TOPS, and the deluxe version is equipped with the computing power of 800TOPS.

Media: can you tell me the life of lidar?

Suqing: it's no problem for lidar to see it on a passenger car for 10 years.

Media: mass production?

Suqing: mass production, spot.

Media: when will it be delivered?

Suqing: November and December this year.

Media: in which parts of China will our car owners test?

Suqing: the first is to go north to Guangzhou and Shenzhen, the other is the national highway network, and the ring roads in all major cities are also running, which is the first batch to cover. We began to run in second-tier cities in the second half of the year.

Media: are we talking about all the roads in the city?

Suqing: it's all in the city, but Beijing is a little special. Beijing cannot enter from within the Fifth Ring Road of the law.

Media: I haven't experienced the function of AVP in the past two days. What's the R & D progress and mass production plan of this area?

Suqing: AVP is actually the first part to be completed. I think you may not be very interested in parking. We can arrange for you to experience it next time, because AVP is definitely the best of our mass production cars, and everyone should be amazing.

Media: can you achieve L4 AVP? People can get off?

Suqing: I think so. Now people always like to get rid of their hands, eyes and feet. I know it's good as a financing or a gimmick, but to be honest, I've been on autopilot for so many years. I actually don't like that.

I think what needs to be solved more is not to set up a Demo, in a specific business district or a specific Building, which is not what I want to do, what I want to solve is to solve the problem of your daily commute to and from work.

Obviously, one problem is that it is impossible for every commuter, wage earner, his Office and the community at home to build this map and the map of the garage, which is impossible, and no one can do it.

What I would like to solve is to use the car self-learning technology to solve the automatic parking of everyone commuting to and from his office and home garage every day.

What I want to pursue in the first step is not that this person gets out of the car and leaves the car. What I want to pursue is that when this person arrives at the gate of the community, the car will tell you that you don't have to worry about it now. As long as you activate this function, the car will automatically park you in the parking space. This is the problem I solved in the first step.

Media: can you briefly expand on the NCA, ICA+ and ICA, you mentioned?

Suqing: to put it simply, the NCA model is that you can see that there are prefabricated high-precision maps in the car. ICA+ does not have high-precision maps, but the car will automatically learn the map according to the environment of the car or the environment he has driven. This is ICA+.

In the place where you drive for the first time, there is always such a place, no one has ever driven, no one else has driven a car, this is a complete ICA mode, we can see that Tesla is now an ICA mode, divided into these three.

Media: how do users feel like in ICA+ mode?

Suqing: you will find that ICA+ is based on a zone between NCA and ICA. The more times you drive, or the more times he drives, the closer his experience will be to NCA. When you drive less, the experience is a bit like ICA, which is a self-learning process of gradual improvement.

Media: can be understood as whether there is a map, maybe the confidence of this system is different, in a certain case it is easier to quit?

Suqing: to put it simply, if you go to a strange and complicated city by yourself, you will slow down and be careful because you don't know if there is a gap in front and whether pedestrians will jump out. This kind of problem is actually the same, as far as cars are concerned.

Media: yesterday our engineer said that there is no way to achieve peer-to-peer after being downgraded to ICA+?

Suqing: not exactly, point-to-point, which means that you can search this target point on the map at any time of departure. But in ICA+, in theory, there is no global map, so in the places you have been to, the simple point is to commute to and from work every day, which is actually possible because you have driven it.

But if you want to generalize to all locations, it really can not be achieved, you can understand that its map is an incomplete map, it is easy for you to understand the matter in this way.

Media: because the accuracy is not as high as that of high-precision maps, so the ability may be a little weaker?

Suqing: the accuracy of the map is enough, but the data is incomplete. Let me give you an example, when you only drive once, your bike lane may be constructed, your far away lane may be missing, your opposite lane may also be missing, you have to drive more and more to be more complete.

It's kind of like playing StarCraft, remember? At first, the map is black, and the place you drive is white. This is the process, very similar.

Media: the weather we experienced today is better. I don't know how to deal with weather like storms and typhoons, as well as night patterns and tunnels.

Suqing: there is no difficulty in the tunnel. I don't know why people always talk about the tunnel. There is no difficulty in the tunnel. The tunnel is just a positioning problem and there is no GPS. But then again, if you drive in the city, you can't rely on GPS even if you drive under the viaduct, which is not realistic, unless you are doing Demo for fun, which is obviously not difficult.

The storm can take a look at a video released by our car show last year, which is equivalent to heavy rain, so it's not a problem for us, is there any other scene?

Media: at night?

Suqing: some of our classmates have experienced the night mode, and it is not difficult at night. On rainy days, cars behave more carefully and conservatively because the sensors are blocked, but there is no visible difficulty at night and during the day.

Media: how do we divide our work with BAIC? What is the difference between these car companies that Huawei cooperates with?

Suqing: this is a good question. I still don't have a very clear interface for division of labor, because we all work together to build this car.

If you must share, BAIC may be more in the relatively traditional part of the mechanical system and chassis system of the Cover.

Huawei helps him with the computerization of the whole car, including autopilot, cockpit, and back-end cloud. If you divide it for a long time, it's probably like this, but it's not that simple.

Differentiation frankly is a serious problem. What do you think is the difference between smartphones? What is the difference between mobile phones? The more complex the electronic system, when the development cost of each subject reaches billions of dollars, it should not be differentiated in this place. This is a serious problem.

Media: is the cooperation between Huawei and BAIC a model or a series of platforms?

Suqing: a series of models, because this kind of thing is a huge investment for us and BAIC, and it will not be just one model.

Media: are Changan and GAC, too?

Suqing: same.

Media: when will a series of models come?

Suqing: from the first half of next year to the first half of next year, you will see a large number of cars on the market.

Media: just now you said that the completion rate of our experience car is only 30%?

Suqing: algorithm.

Media: how many percent of the cars can we test by ourselves?

Suqing: you can't say 100%. Once you become such a complex software system, there is no 100%. It is iterated every two or three months, and the iteration is quite large. Take the emergency brake as an example, it is about 70% or 80% less. Let me just give an example.

Media: how long ago is the stable version of the car we made?

Suqing: it only adapted for two months, the version is the same version, it only adapted for two months.

Media: what have car companies taught you?

Suqing: to take a simple example, at first all the people who made Robotaxi had a big pile of sensor towers on the roof of the car, like a tower.

Frankly speaking, we still envy that the algorithm would be much simpler. When we first did it, many years ago, we also wanted to put a tower on it, even a shorter one, which was firmly stopped by one of our big clients, and you would never be allowed to do so.

So you can see that today's ADS cars look the same as ordinary cars, and this is a very important point that we and the car factory have learned.

On the Planning of Huawei's ADS Department

Media: how many people are there in ADS? How long does it take to get on such a car?

Suqing: the polar fox is the first car with deep cooperation. The car itself may have been developed for 3 years, and it should be faster later. There are always many problems in the first generation. You can think of a lot of later imports, I guess it should be about 24 months, any shorter may be very difficult.

Media: how big is the ADS team?

Suqing: there are more than 2,000 people on autopilot.

Media: are the key R & D in China?

Suqing: all in China, yes.

Media: can you tell me about the division of the people who make boxes, lidars and algorithms in a team of more than 2000 people, and what is the proportion?

Suqing: you can think that there are about 1200 people in the pure algorithm, which can be divided into several large chunks. We call it big perception, that is, the vision and laser are all in the big perception team. There is also a second prediction, and the third is called PNC, regulation, which will be subdivided in PNC.

You can think that the size of each team is about 200,300 people, and the remaining 1000 people are just doing the other things you just said.

Media: in addition to the polar fox, are there any new models and brands to be added in the next stage?

Suqing: yes, maybe now the company has released three, Jihu, Changan, Guangzhou Automobile, and there are some other big factories behind, which you will see later.

Media: is it our ADS going out to sea or their domestic models?

Suqing: first of all, their models in China.

Media: what do you think of Robotaxi? Will you consider doing business?

Suqing: first of all, I would like to express a personal position: I will not do Robotaxi even if you kill me. Robotaxi is a result and should not be a business goal.

For Americans, the taxi-hailing experience is very poor. I have been on business trips to the United States for so many years, and the experience is very poor; the taxi-hailing experience in China is very good, and to be honest, it is not expensive. If you really say Robotaxi, then China has already realized it, but that Robo is an individual, and there is no problem with this experience at all.

Today you turn it into a computer, and the experience is not going to get any better, frankly. I firmly do not think that this thing will change the basic disk of experience, the basic disk of travel, in China, absolutely not.

Second, Robotaxi is one of the most difficult problems, technically, because it needs to sweep away all Cornercase. So why do I say that he is a result? you must have a long and imperfect process before everything is perfect, and you need to take over.

With the evolution of time and the maturity of technology, it can be realized one day. But it will take a very long time, and it will require a large number of cars. It is by no means tens of thousands of cars as we are talking about today.

When you dare to say this sentence, when you have hundreds of thousands of millions of cars running N years, the data tell you can, you can. So it is a result, it should not be a goal.

My personal opinion: all companies that aim at Robotaxi are doomed. And the last person to reach Robotaxi is the passenger car, that market must be mine, but not now.

Media: in addition to BAIC, Huawei's Changan and Guangzhou Auto provide several cars to build cars. Why did Huawei choose these car companies? what is the degree of Huawei's participation in the cooperation?

Suqing: there are many reasons for choosing customers. To be honest, the speed of domestic partners is indeed faster than that of international factories. This is also the reason why you can see that domestic car factories come out first, and this may also be the reason for China's speed.

The second reason for choosing several car factories is different. Like BAIC's first customer, BAIC is very sincere and cooperates very well. Today, you see that Huawei's plan is not bad. What you saw three years ago may not be like this. Maybe it was really a prototype thing at that time. At that time, BAIC chose to trust Huawei and cooperate deeply with Huawei.

And they have really done a lot of work on the self-driving chassis, which is a big reason for the deep cooperation with BAIC.

In fact, Chang'an is also similar, there are similar stories, of course, there are different reasons, of course, there are also commercial interests.

Media: what is the priority of autopilot in Huawei BU?

Suqing: from my point of view, autopilot is absolutely the first, not a little bit of the first.

Media: what are Huawei's future investment plans for autopilot?

Suqing: there are more than 2, 000 people now, and we spend about $1 billion a year. I guess the growth rate will be about 30% a year in the future.

Which echelon does Huawei autopilot belong to in China?

Suqing: definitely number one.

Media: what will the payment model be like in the future? will consumers pay a lump sum after buying this car in the future?

Suqing: there are two kinds. First, the one-time payment model. Second, subscription mode. There will be both.

Media: does the subscription model have anything to do with Huawei?

Suqing: it has something to do with the car factory. Why does what I do have nothing to do with me?

Media: when will ADS be profitable?

Suqing: I'm not in a hurry. Huawei makes a profit for 10 years in everything it does. The only thing I have to do now is to make the technology the best in the world and then solve the real problem. Autopilot, in fact, I don't think there is any need to worry about profitability.

Let me give you an example. I remember the beginning of 2006, when Nokia was very popular. At that time, we told the company that we wanted to build smart phones, and a bunch of people said that you were crazy and told me that the user penetration rate was only more than 0.000. It was a toy for enthusiasts, and it was a toy for you guys of science and technology, saying that there was no market for this thing.

So the first thing is to judge whether this thing is right or wrong in general. If it's right, the market doesn't have to worry, depending on whether you can do it well.

On the supervision and responsibility of autopilot

Media: how is the safety responsibility of this car divided?

Suqing: we always adhere to the experiential approach, reaching the level of L4, but legally it is L2. There is no ambiguity in this.

And I think if autopilot wants to develop rapidly and bring you a better experience, it must be done. Is to decouple function, experience and legal liability, otherwise the car factory will be very careful to provide you with the safest, but basically equivalent to waste function, in fact, we see a lot of car factories do this.

Media: when can autopilot cross the L2 stage? Is the law a little loose now?

Suqing: let me be frank, what people say about laws and regulations now is a lie, that is, technical problems.

Today you have seen that our car is already L4, but I will tell you clearly that I dare not let the driver leave that car. If you can take over even once in 100 million miles of 1000 kilometers, you will soon get there, right? Before the MPI is extremely big, don't talk about L4, it's all Demo.

Then, as soon as we talk about the law here, I really see that the law in China is already very loose, and the state is very supportive of autopilot. I don't think it's appropriate for you to bring the law out again.

About the technology of Huawei ADS

Media: is the self-learning map sent to the cloud and then redistributed to all vehicles from the cloud?

Suqing: depending on the choice of different automakers, this can always be kept at the end of the car, or it can be reintegrated in the cloud.

Media: then Huawei provides the best solution, relying only on the car end or mass production car, in fact, you don't need to make a map car to run a high-precision map?

Suqing: if you are dealing with daily commuting, point-to-point commuting, it is OK.

Media: NCA is Huawei's own application of our team to do high-precision map data collection?

Suqing: we have two parts. In fact, we want to give a brief introduction to our system. Our whole map system is called Roadcode,Roadcode, which is composed of two parts, one is called RoadcodeHD, and the other is called RoadcodeRT.

The meaning of HD can be understood as what people think of as traditional high-precision maps, which are done by a special map-making team and are offline. RoadcodeRT is a self-learning map of the car. These two things are two in one.

I haven't done this before and haven't noticed that the infrastructure of the whole city is changing so rapidly. I find that the urban roads of the whole Shanghai are constantly renovated, and the traffic lights are changing much faster than I thought. If you just use the traditional RoadcodeHD technology, you will soon be dead.

So RoadcodeRT itself will constantly update HD, to precipitate the data after learning, which is a process of precipitation iterative cycle.

Media: last night we found that the side of the vehicle encountered takeout guy will be more tangled, is there any good solution?

Suqing: you are right. You will find that there is a blind spot in the laser coverage behind the side of this mass production car. It needs to be repaired by vision. In addition, this car is not in the final state of mass production. As a matter of fact, it will only take two months for BAIC to adjust its chassis to be able to use and be able to switch to autopilot on the road after the Spring Festival.

So you can understand that the completion of the algorithm is only 30% and 40%. The problem will be solved when you buy it.

Media: is this solution to add sensors?

Suqing: no, optimize the algorithm.

Media: can the position of the side object be accurately detected only by vision?

Suqing: this car actually has two circles of vision sensors, a long-distance one and a fisheye, which we also use.

You will find that the feature of vision is that the farther the distance is, the greater the measurement error is. When the distance is shortened, its measurement accuracy will be rapidly improved, even higher than the laser.

Your problem is a side lane or nearcut-in problem, at this time close visual ranging is not a big problem, in principle.

Media: so when we are driving, we will look around not only in parking conditions?

Suqing: of course, or wouldn't it be a waste?

Media: would you like to ask pre-fusion, since it is pre-fusion, will all this information be summarized into a whole set of computing centers or neural networks to digest it?

Suqing: you can think that all the information is the input of the network, but it is not a network. Different networks perform different functions.

Media: you mentioned that there could be no takeover for 1000 kilometers. How did you get this data?

Suqing: to be honest, I haven't found a better indicator for measuring autopilot with MPI data so far, but there are a lot of calculation skills and technical methods in the value of MPI, which is why I don't want to talk about it.

As you can see, to put it simply, MPI has something to do with a few things:

First, it has something to do with statistical methods, and then it has something to do with time and space. Time and space means that what kind of road you choose to run at what time has something to do with all three things, and the value can be as low as an order of magnitude.

Why does what I just said have something to do with statistical methods? We can see the statistical results in California. If you have hundreds of cars, you can pick out better samples in a certain period of time and count it continuously for a better period of time, the MPI value will be very beautiful.

When we do the MPI value in this way, frankly speaking, it doesn't make any sense. It's more likely to accumulate all the cars in all the time. At this time, the statistical significance of MPI is a real MPI.

All I can say is that any MPI is a core secret in the autopilot team. Frankly, I can't tell you a specific number, it's not a simple number, it's a big watch. In the so-called California statistical method, I can really do 1000 in Shanghai.

But in terms of real historical statistics, I can only say that I haven't reached 1000, which I have to tell you, and I bet no one in the world, including Waymo, can do 1000.

Media: will there be any difference between your self-study and Tesla's shadow model?

Suqing: to tell you the truth, Tesla's current model only sees the concept and has not explained the details. from our practice, there are at least a few things, whether you call it the shadow mode or the car end intelligence.

We have two major technologies, one is the RoadcodeRT, mentioned just now, which solves the problem of self-learning and self-composition of the whole traffic static environment, including the AVP mentioned just now is also realized by this.

The other is that we call it DDI,DDI, which may be more like your shadow model, that is, DDI will constantly learn the driving behavior of the car owner, not necessarily taking over. Maybe if the behavior of the car itself is not consistent with that of the car owner, he will grab the behavior of the car owner to do iterations, maybe it is the shadow mode you mentioned.

Media: your pre-visual perception is unique.

Suqing: what is unique?

Media: four cameras, telephoto + wide angle + binocular, rarely seen in mass production cars.

Suqing: yes, because the eyes are more difficult, in fact, we didn't do it, but we just did it.

Media: what problem have you solved?

Suqing: there are a lot of problems with binoculars. To put it simply, it is not easy to make good use of binoculars in terms of algorithms, because the essential problem to be solved by binoculars is depth measurement. However, it is very difficult for depth measurement itself to be relatively stable and can be generalized. Most binoculars can only achieve 20 or 30 meters, and we far exceed this data.

Media: at present, this architecture has three lidars and many cameras. This set of architecture can be done on other models, basically not in terms of the number and type of sensors.

Suqing: pretty much, you'll see it's almost the same in this generation of cars. We usually do a small upgrade every 18 months and keep iterating up.

Media: you just said that autopilot may need millions of cars, and now Tesla has a million.

Suqing: good question. I remember who said a question before that makes a lot of sense. What is big data? Big data's focus is not on the word 'big', but on data quality and completeness. This is the essence of big data. Autopilot is actually very similar. There are two key questions in the data. First, the quality of the data itself. Second, the dimension of the data.

On these two issues, I think there is a big problem with Tesla's data.

What is dimension? Just rely on a few simple visual data collection, this data high precision positioning of nothing, the dimension is very low. It is obvious that the car data dimension of ADS is several orders of magnitude higher than it, and the data dimension is extremely important, and the data dimension represents the degree of information richness and differentiation.

Second, the quality of the data itself. You will find that the data itself is generated by algorithms, and the complexity of your low-order system leads to relatively low quality of the data itself. Tesla is currently in this state. If you want me to guess, Tesla's data has long been saturated and has not improved the ability of the system.

In fact, what we lack in ADS is not data, but the algorithm has a lot of problems to solve, I am absolutely not short of data.

Media: to solve these problems, the first logic is to perceive in place, and the second logic is to predict each other's vehicles in place. Which is more difficult?

Suqing: good question. It was very difficult for us to work on the first day, and then we found it difficult to predict. After the prediction, we found that the regulation and control was difficult, and now the regulation and control is over. When we come back, we find it difficult to perceive and continue to cycle.

What you must ask me to say, in terms of the technical complexity of the industry as a whole, everyone knows that it is difficult to perceive at the beginning. If you ask me, in terms of theoretical and technological maturity, forecasting and regulation are the real problems, which many people may not be aware of.

Media: algorithms mainly rely on neural networks for deep learning, and sometimes there are black boxes in deep learning. Do you think there will be a breakthrough in algorithms in the future?

Suqing: first of all, there is not only a neural network in the autopilot system, but only a part of it. In terms of calculation, it accounts for the absolute majority, but definitely not in terms of code scale. First of all, let's clarify this problem.

Second, I firmly disagree with the idea that AI is a black box. Its calculation mode has changed from scalar calculation or linear calculation of CPU to probability calculation based on linear algebra. It is completely explainable from the point of view of probability, and there is no problem at all.

You can think of it as something like probability, statistics and what was in the past. You can't say that probability and statistics can't be explained. I totally disagree with this view.

On the redundancy of autopilot system

Media: so we rely on lidar for long distances?

Suqing: no, that's not how we divide it. People used to ask the question, is your perception pre-fusion or post-fusion or some Redundancy (redundancy) technology?

First of all, I think the idea that sensors don't have Redundancy is nonsense. Then the post-fusion technology was abandoned by us two years ago, and now we are all pre-fusion technologies.

The feature of pre-fusion is to put all the information together and send it to the NN network for processing. It is not a simple question of which sensor uses which information. You can also simply understand that sensors have an Attention mechanism for each other.

On the other hand, the characteristics of different sensors are different. For example, millimeter wave speed is sensitive, but the measurement is a mess, and the vision is better for semantic measurement; the laser is better for geometric measurement, and it itself merges these together.

Millimeter wave we take the raw data of millimeter wave directly and use its original point cloud.

Media: will it be difficult to get raw data from suppliers?

Suqing: there are two questions. First, most of the Tier1 is not willing to open up the original data to you, but Huawei is bigger, and people are also willing to open it. Second, the raw data of millimeter wave is relatively dirty, which is actually more difficult to deal with. We are now using NN to deal with it.

Media: do binoculars make laser redundant?

Suqing: this thing can not be called redundancy, in fact, different sensors, different performance, advantages and disadvantages have a fluctuation.

Media: a head autopilot company proposed true redundancy, taking radar and LIDAR as a subsystem, pure vision as a subsystem, independent testing, and multiplying the takeover rates of the two subsystems to reduce the required test mileage in a statistical sense.

Suqing: frankly speaking, I guess it was written by their Marketing, definitely not by their research and development, otherwise I would doubt their R & D ability. It is not only your perceptual system that really determines your takeover rate, but also has a lot to do with your regulation and control, even bigger than your perceptual system.

These systems don't have what you call a truly redundant design, right?

Second, you can't handle most of those case, that are difficult to handle with 80 times the number of sensors. I bet you. So the logic of doing statistics with this multiplication algorithm is absurd.

True redundancy is a very Marketing saying, if you want to do a good job in sensing, you should do sensor fusion, not redundancy, which is a serious waste of sensors. Their technical level is definitely not like this.

About ADS's competitors

Media: some multinational companies are still playing the concept of L3, but Huawei insists on continuity optimization. What do you think of this traditional big company going to make a breakthrough in responsibility, while we go to break through two different differences in continuity?

Suqing: in fact, you see, the ideas of everyone in Europe are not exactly the same. I make a personal evaluation, which does not represent the position of the company.

I personally think that among the big three in Europe, the public in BBA is actually relatively forward in thinking, which has something to do with their exploration of autopilot for so many years. The ideas of other families are still in the process of evolution.

In fact, Tesla, I'm sorry, I still have to mention Tesla and Tesla. I think it has taught everyone, including us, and the car factory a lot of things.

When you see the car factory, you will find that what is the essential change of an industry? If you look further ahead, you will see clearly.

To go a little further, we used to be steam engines and electric steam, and then the energy revolution or the power revolution, and then the computer was invented, and then the computer was changing everything. In fact, this is the process in the past 30 or 40 years. The computer is changing everything. The mobile phone was changed last time, and the car was changed this time. This is the view of us and Tesla.

In the traditional car factory, first of all, my base is the car, and now there are some single points of the computer, so I take the car as a foundation, and then I try to embed the computer into it. This is the view of the traditional car factory.

Our views are different, our view is based on computers, cars are computer-controlled peripherals, this is the essence of different views, will lead to different views on everything.

So you will see that traditional car factories use this as an idea to make a lot of small boxes, a function plus a box, a function plus a box, but our view itself is a computer, a big computer, and hang up the car. this is essentially different.

Media: so is this the fundamental reason why Huawei doesn't build cars?

Suqing: if you don't build a car, I think it's a matter of business choice. If you don't build a car, the market will be bigger.