He Xiaopeng discusses large-scale model transformation in intelligent driving: Building a car is like swimming in a sea of blood, but I am more confident now

As a founder, you have to imagine the direction and the brightness for yourself. Sometimes you may not see the direction and the brightness, but you have to be able to imagine it and lead a group of people towards it. So, even when swimming in a sea of blood, "I am not desperate."

He Xiaopeng had a strong premonition that AI models would definitely transform the automotive industry; but he had no clue at all about how to implement autonomous driving at the beginning.

He had his team study papers, visited the United States three times, test-drove Tesla's FSD and Waymo, reluctantly sent away the "beloved general" Wu Xinzhou to join NVIDIA, and even visited Jensen Huang (NVIDIA's founder and CEO). However, unlike his investor and friend Lei Jun, who "cast a wide net" investing in Chinese AI model companies, he had good relationships with many unicorn founders but did not invest in any of them.

"China needs companies like this, but China also needs companies that can understand AI models and apply them in engineering and commercialization—Xiaopeng is one of them," said He Xiaopeng, Chairman and CEO of XPeng Motors.

In the automotive industry, He Xiaopeng is an outsider. Born in 1977, he graduated from the Computer Science Department of South China University of Technology, worked at AsiaInfo for 5 years. Starting in 2004, he spent 10 years founding UC Browser and sold it to Alibaba with a feeling of "marrying off his daughter." He summarized his business genes as "mobile internet + tools." Although this is his 7th year in the automotive industry, many of this CEO's thoughts still come from the mobile internet and smartphones.

In He Xiaopeng's view, so far, no company globally has truly made money from AI models; and what can make money from AI models in a different dimension is autonomous driving—"If you are asked to pay tens of thousands for software or services, you won't buy it. But if you spend tens of thousands, hundreds of thousands, or even millions to buy a car you like, with this capability included, the likelihood is very high."

In May of this year, we had an exclusive interview with He Xiaopeng. In addition to AI models and autonomous driving, standing at the 10-year mark of XPeng Motors' founding, we also discussed with him about friendship and business warfare, peaks and troughs. He had many touching expressions.

When talking about the difficulties of car manufacturing, he said, "Brothers, whoever you want to harm, let them come and make cars."

When talking about the G9 accident, he said, it's not "cutting off one's arm," it's "cutting off one's head."

However, he also said, "Making cars is really tough—it's all world-class opponents, the most awesome people, with long chains, long cycles, and high barriers; but making cars is also really cool."

He feels that making cars is like "swimming in a sea of blood"—"As a founder of a company, you have to imagine the direction and the brightness for yourself. Sometimes you may not see the direction and the brightness, but you have to be able to imagine it and lead a group of people towards it."

Below is the full interview with He Xiaopeng. (For ease of reading, the author has made some text optimizations.)

Seating, Friendship, Stock Price

"They said you probably got so angry that you cried in the bathroom - it cracked me up."

From right to left are Li Xiang, He Xiaopeng, Li Bin, Wei Jianjun

"Qian Wang": In March this year, you, Li Xiang, and Li Bin all appeared at the Xiaomi SU7 launch event. How did Lei Jun invite you?

He Xiaopeng: It was a coincidence. That day happened to be attending the Automotive Hundred People Conference. The three of us may have different thoughts, but because we sat together and talked about this matter. You know, Lei always has a long history with me. I like Lei a lot and also respect him. So I said, "I should come."

I came from Guangzhou - I came in the morning and took the latest flight back; for those who came from closer places, it's not easy to say "I'm too far away" - the two of them thought about it and eventually both came. (laughs)

"Qian Wang": You were the first one to agree, and they hesitated?

He Xiaopeng: They both agreed within an hour. In such matters, if A doesn't go, then B won't go; because A and B don't go, so C won't go either, this kind of behavior is very common. Everyone needs to check their calendars, afraid of other things getting in the way. But when someone says "I'll go" on the spot, it's different.

"Qian Wang": Why did you sit between Li Xiang and Li Bin that day?

He Xiaopeng: They are taller than me, so I look a bit better.

No matter how business is, personal relationships should be maintained. In most cases in the internet industry, this can be achieved, but it's difficult in the automotive industry. This is a feature of the automotive industry: the "battle" in "business war" is more intense. I don't like it very much, but I'm forced to do it. Follow everyone's heart. Preserve your own genes.

"Qian Wang": Is it because you poached Great Wall Motors' General Manager Wang Fengying, so you let Li Bin sit between you and Wei Jianjun?

He Xiaopeng: Oh no! Impossible. We all respect the big brother. Maybe Xiaomi arranged the seating.

"Qian Wang": They considered it (laughs). During the 3-hour event, what were you thinking the whole time? From the photos, you seemed to be the happiest.

He Xiaopeng: Firstly, Lei Jun made a good car, exceeding our initial expectations of him for the first car. Secondly, some of the things he talked about were very interesting to us, giving a new feeling; there were also some pitfalls we had encountered before, not sure if he would encounter them again.

We discussed below, like colors. Initially, we thought it should be colorful, but the colors kept decreasing, down to the best combination of your expectations and actual customer choices, otherwise, having too many SKUs. Generally, in the showroom, you would display the most mainstream colors with just one car. If you use 7, 9, or even 11 colors, some colors have a particularly slow delivery cycle. So, you would reduce the number of colors to reduce management difficulty In the first phase, he did very well, but there may be changes in the future, focusing on it - those that sell well continue to do better, and those that don't sell well adjust the colors. Of course, different cars are different, Xiaomi SU7 is suitable for rich colors; some tough cars don't need many colors.

"Qianwang": After the release of Xiaomi cars, do you have a new view of Mr. Lei?

He Xiaopeng: Mr. Lei has been under a lot of pressure in the past few years. I am very happy that he has achieved great success in the field of automobiles again. I am really happy for him. But the pressure also lies in the fact that being too popular may not necessarily be a very good thing - at least not for me - different people need different lives.

"Qianwang": When did you find out that Xiaopeng Motors' stock price had fallen? - You know, this later became a joke.

He Xiaopeng: I know there is a joke, but I don't pay much attention to stock prices.

That night, they said you were on the hot search again. I said why? They said you were probably so angry that you cried in the bathroom - it made me laugh - the last flight from Beijing to Guangzhou was at 9:30, so I left after 8 o'clock.

My Genes, Anxiety, and Limitations

"If I had the logic of a car brand, I would cut it a bit higher, this is all part of the reflection!"

"Qianwang": Today's topic is large models and autonomous driving. In the past year, when you chatted with CEOs of car companies, did you talk about large models?

He Xiaopeng: No.

"Qianwang": Do you talk about it with Li Xiang and Li Bin?

He Xiaopeng: No.

"Qianwang": You should talk about it with Mr. Lei, right?

He Xiaopeng: Still no.

"Qianwang": Xiaopeng has been firmly committed to intelligent driving from the beginning. Where is the source of your intelligent driving? What are the key points in these years?

He Xiaopeng: Our thinking about new energy vehicles has been clear and simple from the beginning - first is intelligence, similar to the change from feature phones to smartphones; second is from gasoline to pure electric, electricity can power laptops, servers, and the conversion from oil to servers requires electricity, only with larger batteries and more electricity can ensure a better sensory experience for smart cars.

We look at this market from a mobile phone perspective. Most car people think that there are huge differences between cars and phones, and indeed there are, but in the end, many companies' capabilities tend to converge.

In the past few years, Xiaopeng has taken a different path in intelligence. We have always led - the first to put automatic parking in advertisements as important, the first to launch high-speed assisted driving, the first to speak with voice and assistant, the first to put lidar on the car, and the first to have high-definition maps in cities.

In addition, it is an end-to-end model. Originally, you wrote programs using programming methods, saying how to turn left, how to turn right, how to turn around, the rules of the program are endless. The logic of turning right in one place is different at every time point on each road or even the same road. It is very difficult to write rules, even with generalization. We use an end-to-end model for training - here, we see a brand new opportunity for change.

"Qianwang": You look at cars from the perspective of a mobile phone, why is this the starting point?

He Xiaopeng: In the previous entrepreneurial venture, I came out of mobile internet (founded UC Browser). At that time, the investors included NGP Capital (Nokia Growth Partners), invested in 2007. In early 2007, Apple had not yet released the iPhone 1, and everyone thought it was difficult to challenge a global brand with such good technology reserves and customer awareness But it only took 4 years, until 2011. Everyone found out that phones could do more than just make calls and send messages.

At that time, there was a strong belief that cars should be driven by oneself. Everyone thought they should drive themselves. The most important thing about the iPhone was changing the way of interaction, turning buttons into touchscreens, changing from low speed to high speed, and adding unified hardware, including cameras. After the interaction revolution, the scenario changed from making phone calls to sending WeChat messages and watching videos. Therefore, the revolution of interaction and the revolution of scenarios are sequential, or interdependent. Of course, without a doubt, cars face issues such as safety, policies, and user acceptance, which are much slower than phones.

I used to think more about the mobile internet - how will phones revolutionize, how will operating systems revolutionize. We used to work on browsers, and a major issue with browsers is that they are strongly related to the operating system. If you want to create a WeChat, no core of any phone will create content similar to chat by itself, it will definitely be in the App Store, unlike browsers.

"Qianwang": The revolution of interaction and scenarios are interdependent. In the field of automobiles, after the revolution of interaction (reaching a certain level of autonomous driving), how will the scenarios change?

He Xiaopeng: For example, when driving home, the car can automatically drive to the parking spot you remember; originally, cars and seats pursued a five-star safety collision. After the introduction of autonomous driving for 10 years, when autonomous driving cars reach a sufficient scale, there is no need for a five-star collision, and the form of cars and seats will undergo huge changes; the large screen will change, you will not sit in the driver's seat or front row to click, you may use voice, use projection, project to where, then decide.

"Qianwang": You invested in XPeng in 2014, joined as chairman in 2017, when you discussed with Li Xiang and Li Bin in those years, were your ideas aligned? They are also entrepreneurs who entered the car manufacturing industry from the internet. In your opinion, what are the differences between the three of you?

He Xiaopeng: We knew each other, but we didn't talk much at that time. I first asked them about this matter. In 2014, I met with Li Bin, and after the discussion, Li Bin said: he was ready to do it. At that time, the circle was very small, almost no one in China was making cars. Unlike the internet, if you want to invest, you will always find 5 or 10 companies with an idea. At that time, there were none, so we could only set up a team because cars are too heavy.

The three of us have significant, significant differences - they also work in the internet industry, but more in internet cars and media; I work in mobile internet and tools - so the thinking logic is different.

Look at me, I focus more on technology because my main competitor later on was Google Chrome. I focus more on globalization, we put a lot of effort into globalization, later on, we couldn't beat them in developed countries, but we could in developing countries, not relying on technology, but on operations, which had many challenges.

Why do I focus on globalization? Why do I do technology? Technology is easier to globalize. At that time, UC relied on operations, what is operations? Knowing who your users are, based on the positioning of users, making them happier. It's very difficult to go international, relatively feasible in one country The ultimate standard of technology, good user experience, loved by everyone, whether Chinese or foreigners.

"Qian Wang": It sounds like a lot of your entrepreneurial thinking comes from your summary of UC.

He Xiaopeng: Yes, or mistakes.

"Qian Wang": Can you summarize the correct and wrong aspects of your previous entrepreneurship? Many entrepreneurs write letters to their previous investors reflecting on themselves when starting their second venture. Have you done something similar?

He Xiaopeng: Oh no, but there are definitely a lot of reflections. Why do we believe in the change brought by intelligence? Why do we believe that only technological positioning can be done well globally? Why did we initially think globally is a big picture, but we need to move slowly? Why did we enter from a level of 200,000 to 300,000 (cars)?

From the perspective of automobiles, brands should move up and then down. But the internet is not like that. The internet generally either cuts the waist or the bottom. At that time, I didn't have the logic of automobile brands. If I had the logic of automobile brands, I would have cut it a bit higher, this is one of the reflections!

For example, if I understood the automotive system better, I would start with the P7 first, then maybe the G7 system. The cutting angles are different.

Of course, there is a big difference between Xiaopeng, Li Xiang, and Nio. I was originally an investor, not a founder. By the time I joined, the company had been operating for 3 years. They had positioned themselves well, although they had met with me. But to be honest, as an investor, the logic and perspective are different.

"Qian Wang": What were the high and low points psychologically during your time at UC?

He Xiaopeng: There were too many low points, the company was on the verge of collapse and couldn't pay salaries, borrowing tens of thousands of RMB every month because the company was small. Faced with serious challenges, feeling that the company couldn't continue, it was very painful. The high points are not as memorable, the first successful financing.

The day we sold the company, I was very worried that my classmates would be unhappy, going to the office to see their expressions, chatting with them - that kind of uneasiness. Suppose you have a daughter who is getting married, you are most afraid that she will not be happy. Regardless of whether we are there or not, we are very concerned that UC should become a good, profitable company, making the "son-in-law" who marries her happy. These are all memorable moments.

"Qian Wang": Why did you join in 2017, but the company was named "Xiaopeng Motors" in 2016?

He Xiaopeng: It was originally called "Orange Motors". If you are making internet apps, you can use any name. But for cars, if you choose a name and it doesn't pass the trademark, you can't advertise it, and if you can't advertise it, you can't sell cars. At that time, a trademark was necessary. You should know that all 4S car dealerships register car trademarks, and good trademarks are basically already registered.

They said my name probably wouldn't be registered, they positioned it. They took a lot, a lot of names, and in the end, it was Xiaopeng. They said, let's see if the other names are good. I said, this one probably won't be registered in the end. As a result, none of the previous names could be registered.

"Qian Wang": How many names were on your list? Xiaopeng He: Dozens.

"Qianwang": What do you feel about having a car named after you?

Xiaopeng He: At first, I was not very willing, because as an investor, I shouldn't. But later I thought, investors have a saying "help without causing trouble." They said this is a normal choice, not related to you.

At that time, before I joined, it was not highly relevant to me, do you understand? But if I joined, the relevance would indeed be very high.

"Qianwang": I have seen you talk many times about why you decided to join XPeng Motors, which was related to the birth of a child. Is there an updated reason today?

Xiaopeng He: The former boss talked to me yesterday, mentioning a similar story - I won't mention the name, he is too famous - life can be long, but the vast majority of people do not live an exciting life. Some people have a lot of money, their careers are going well, it's a smooth life. I had this kind of thinking in that era too. Nobody wants to suffer, but everyone wants excitement. In this world, without effort, without risk, very exciting, and even making a lot of money, there is only one place: in dreams.

At that time, the most important thing was the birth of a child, plus a phone call from Fu Jixun (GGV managing partner), which had a momentary impact on myself. At least one or two hundred people had asked me before, but at that moment I felt I should do something. Do something entrepreneurial, not related to the internet. At that time, it was not certain to make cars, I thought about it for more than a month. I should do something exciting, not make myself unhappy, and make the child feel cool.

Before, we had a video about screwing, and my child happened to see it. At the end of last year, he asked me, "Dad, aren't you screwing?" "Why is your name the same as this company?"

Before, he only knew me as "Dad," not my full name.

How exactly does the large model transform autonomous driving

"At this time, human rules cannot defeat"

"Qianwang": Speaking of large models, Lei Jun has invested in many large model startups, would you consider investing in one?

Xiaopeng He: If there is a suitable one, maybe.

"Qianwang": There isn't one now, why?

Xiaopeng He: I have a good relationship with many large model companies, like Zhipu's Tang Jie, Xiaochuan, Zhifei - they are all doing large models, doing very well, I applaud them. China needs such companies, but China also needs companies that can understand large models and apply them to engineering and practical use - Xiaopeng is one of them.

Other companies can do basic large models, we actually don't need GPT-4, we don't need GPT-5, a GPT-2 or GPT-3 with self-controllable capabilities is good enough in a vertical field. We don't need such advanced capabilities, just a standard one. The difficulty lies in implementing the standard capabilities into engineering.

For example, if you start a company and say, I will help you decorate, you hire many people to draw design drawings for me. Sorry, you may only need GPT-3, read 10 million global modeling images, it can automatically generate various things based on the information input you provide, you just need to fine-tune it. This capability does not require GPT-4, GPT-5, GPT-2 maybe is OK But this ability to connect is a huge challenge.

"Qianwang": The most difficult part is the connection.

He Xiaopeng: There are countless problems in engineering implementation. How do your data flow quickly? How do you test your simulations quickly? Why do you have so much data? Just now I mentioned the need for 10 million design drawings. You need to know this design drawing - who has good drawings? Who has bad drawings? What designs do Chinese people like? What designs do people from a certain country like?

Like Xiao Chuan, he is working on both large model bases and vertical industries (medical), as well as software-oriented.

"Qianwang": How did the large model transform automatic driving and accelerate the arrival of the turning point?

He Xiaopeng: Very simple. Most people talk about large models as language models. Basically, no one can make money from language models, except for Nvidia or Microsoft. Microsoft hasn't strictly made money either, like OpenAI raised money, not earned money, including Xiao Chuan.

Large models can make money in another dimension, which is autonomous driving. If it improves the autonomous driving capability by 50 to 100 times, turning assisted driving into fully autonomous driving, or close to unmanned driving, this change in capability is something users are willing to pay for and buy cars for.

If you are asked to pay tens of thousands of dollars for software or services, you won't buy it. But if you spend tens of thousands, twenties of thousands, or thirties of thousands to buy a car you like, with this capability, the likelihood is very high. When 10 out of 100 friends around you slowly become 30, 50 friends who use it every day, you will all start using it. This change will be very fast.

It's different from language models. Language models today are not about reliability or robustness, but more like a model for language communication or assistants. It's a kind of "miracle worker". Cars are also "miracle workers", but cars pursue higher safety and stability. In this case, large models are different from language models. However, their underlying logic is similar, using Transformers and a lot of "pre-training", but we haven't started "fine-tuning" yet.

"Qianwang": When you talk about "pre-training", is it the same as the "pre-training" often mentioned in large language models?

He Xiaopeng: Yes. But for autonomous driving with large models, it also needs a "brake" and a "gatekeeper".

For language models, if you ask a question and it answers incorrectly, you don't care. It's different for autonomous driving - if it makes a mistake, you need to ensure my safety. You need a controller, you need a brake. These are two different directions for large models, but the underlying logic is the same as large models today. Therefore, the automotive world will soon become a large amount of data, a large amount of training, generating a large number of rules, and then replacing humans.

"Qianwang": After ChatGPT was launched in November 2022, what actions did you take?

He Xiaopeng: We went everywhere to see what their papers were like, where they came from, and in these papers, why they extended step by step. In the first half of 2017, we used Transformers and extended based on them First, you need to understand the logic. Large models are not human thinking logic, but you can think of them as machine thinking logic, or machine reasoning logic. The content it produces, it may not understand, but it is more accurate than most people, that's the language-based large model.

Visual large models, in fact, are not directly related to robots or AD (autonomous driving). AD involves perception, positioning, planning, execution, control, and the entire user experience system. This whole set needs to be connected. This is a different logic from language-based large models, but they share a common underlying architecture.

Sometimes I have a hunch. Many things come out, and I feel like it's not there yet—like when VR, AR just came out, W3C's Web3, and Bitcoin at the beginning. But when the large models came out in 2022, I took a look, used it briefly, and felt it was a huge opportunity. The premonition comes from the results of rapid thinking after experiencing it, with both explicit and implicit aspects.

Then I went to the United States, looked at the entire logic of large models; visited Jensen Huang; and had discussions with many people in our team. At that time, I felt that large models were very important. But to be honest, at that time, no one knew how to implement large models.

"Qian Wang": In which month did you go to the United States?

He Xiaopeng: April last year, January this year, and June this year.

"Qian Wang": You should be one of the few CEOs of car companies who have specifically gone to the United States for large models, right?

He Xiaopeng: I don't know if others go or not. This is the direction, but I don't know when this direction will come. Many people say there is a good thing. We say, yes. But hardware is different from software. For example, many people say a new technology can make the battery more powerful, this technology is 90% correct. But this technology cannot make the battery more powerful because the battery may be a combination of 15 technologies, if it improves technology A, technologies B, C, D, E, F, G may all decline. The core issue is how you solve it.

Even if you solve it, you may still not be powerful, why? Because to go from mass production to large-scale production, you need manufacturing equipment, right? You need to build a factory, right? You need people, orders, right? After the orders are used, the product needs to be reliable, stable, and maintainable, right? Then you can lower the price, and more people will come in, right?—Software goes from research to large-scale use quickly, hardware is very difficult, especially security hardware.

My biggest concern at the time was, this thing is the future, but how does it land? Especially how does it land in our industry? We are feeling our way across the river. I can't talk too much about this part, we are almost like "Huangpu Military Academy" now. But I strongly believe this thing needs to be done.

When we went to the CES auto show in January this year, some companies said large models are very valuable for AI, but they are not using large models, they are still using programming methods, that is, rules. If there are three people here, so you need to turn a bit slower—this is an example of rule-based methods. Large models are not like that, they throw all kinds of scenarios in, it generates a rule, and you use a reasoning engine to interpret it. As for the rules it generates, you need a controller The reason why XPeng used to do well is because we made the previous generation of autonomous driving capabilities very strong, so we modified the previous generation and turned it into a controller. (laughs) Otherwise, if you don't have the previous generation of autonomous driving, you can't use a large model for this generation. If you have a controller, with safe braking, and then you use the new generation of technology, it can quickly excel in certain capabilities. There is another capability, where if you have shortcomings, you need to use other methods, including controllers, including brakes, to repair and compensate, which requires more data.

Many business owners come from business, sales, and marketing backgrounds. I myself am a technical person, so I may have a stronger sensitivity to this.

"Qian Wang": How does the large model land on the car?

He Xiaopeng: Today it's OTA (over-the-air technology). We don't just put any model you know into the car. Many people say the large model is great, I just put it in and it can be used. But if you ask them how to put it in, they don't even understand autonomous driving, so putting it in, they can't figure it out at all.

For us... it's a very simple logic. You read all the papers, our team understands our architecture, some of the underlying aspects are similar, such as all based on Transformer. Previously, its reasoning logic was different from our data execution flow, so you should rewrite our data flow logic according to its logic, increase data collection, do pre-training. The engine trained is then tested in a simulation environment. We have a very large simulation environment.

Initially, the testing results in the simulation environment were very, very poor, so you keep strengthening the training.

"Qian Wang": Do you read papers yourself?

He Xiaopeng: Our team reads them.

"Qian Wang": How bad were the initial test results?

He Xiaopeng: What was initially written was completely unusable. Firstly, your model is not strong enough, but most importantly, your data volume is not large enough. For example, if you write a program for turning right, assuming there are 100 million scenarios for turning right, you should be able to cover the ability to cover at least 1 million scenarios, because it's relatively simple. But sorry, the large model cannot cover the first 1 million scenarios, it's still 0 or very close to 0 for a long time. So you keep adjusting the model, the key is the data volume, massive training.

Our training costs in the second half of this year alone exceeded $100 million. After May, the entire training scale accelerated.

China is very interesting, many people bought a lot of cards last year, but couldn't use them because there was no scenario training. This is where we have made rapid progress. When you take so much data for training, one day you will suddenly find that in the simulation environment, it far exceeds the current logic. You will realize: oh, the data volume needs to be increased by another 100 or 1000 times - non-mass production companies can't do it, there are not enough cars, you need huge data to improve feasibility.

"Qian Wang": What level can you achieve today?

He Xiaopeng: Today, we have taken the first step and completed some functions in autonomous driving in the large model. Every now and then, we will increase capabilities.

For this, we also rewrote the simulation because the original simulation did not adapt to the testing logic of such a large data volume. Its rules can be 10,000 or 1 million times that of human rules, and you can't test that magnitude with humans Once we improve this system - you greatly increase the amount of data, the entire supporting data system, cleaning system, training system, and simulation system all need to be expanded, you will get something very different. Large models have opened up a new way for everyone. Relying on human programming in unlimited scenarios is not a solution, but "great power will work wonders".

At this point, human rules cannot defeat it.

Silicon Valley Giants and Autonomous Driving

"I don't have a business idol"

"Qianwang": An investor wants to ask you: How much money, how much time, and what basic elements are needed to replicate a Tesla end-to-end FSD (Full Self-Driving)?

He Xiaopeng: Wow, now, it's probably hundreds of billions... First of all, it requires a lot of money, super strong talent, definitely not something that ordinary programmers can achieve with algorithms alone. It requires running a lot of cars on the road, and all these capabilities need to be put into action. NVIDIA has many capabilities, but it takes a long time to put them into action, and the probability is also very low.

Success has a probability. When Xiaopeng Motors just started, others wanted to invest. I told them, "Welcome to invest. If you have an annual income of 100 yuan, you can invest 10 yuan, 20 yuan, and you may earn 10 times, 100 times, or 1000 times in the future, but you need to know the probability of success." They asked: What is the probability of success? I said: Less than 1%. They said: Less than one percent, what kind of business are you starting?

Today, for autonomous driving, if you have to invest hundreds of billions, several years, and thousands of people, with a success rate of only 5%, would you invest? This is one of the thresholds.

Therefore, investing money only represents the possibility of success, not guaranteed success.

"Qianwang": You just returned from the United States recently, this is your third trip to the United States after the release of ChatGPT, what are your schedules and new thoughts?

He Xiaopeng: The two weeks were very busy, I went to 9 cities - meetings, marketing, learning, vacation. Went to Dallas to watch NBA. Went to two cities in Mexico to see Chinese manufacturing. Went to Santa Barbara and some cities, meetings, listening to different people's views on AI. Went to Silicon Valley, San Francisco.

To be honest, the AI gap between China and the United States is widening. Last year, I only saw many people in Silicon Valley transitioning to AI (content), but I didn't see much change. These three trips (to the United States) have been over a year apart, and the changes are bigger than I imagined, especially this time experiencing Tesla and Waymo.

Waymo has not transitioned to large models yet, still using the original algorithm system. In the past, from my experience, Waymo was far ahead of Tesla, but this time, I felt the acceleration of Tesla.

For example, in the past, China relied on area, volume - that is, scale, efficiency (scale in terms of money, scale of people, efficiency in doing things). But in AI intelligent entrepreneurship, trying to catch up with AI quickly through area or volume changes will lead to wrong judgments - that is to say, focusing on height will bring about huge changes. Compared to the Internet era, the intensity of AI transformation will be much faster "Qian Wang": You have experienced Tesla FSD and Waymo this time, you said that Waymo's experience is smoother, but you believe that FSD will catch up with Waymo.

Xiao Peng: I think it will be next year. There are several logics for autonomous driving: first experience, second cost, everyone is also competing in terms of computing power, data, and the overall experience of capabilities. Today, Tesla's cost is lower than Waymo, its computing power is greater, its data is larger, its scope is larger, the experience in certain areas is worse than Waymo, but the breadth far exceeds Waymo.

Today, Waymo and Tesla are not like before with a gap of 10 times, today the difference is probably around 3 times in terms of experience. Solely in terms of experience, it is 3 times, but Tesla is much stronger in other capabilities. As for this 3 times experience, a company today that is AI-driven (wants to catch up) will not take too long.

"Qian Wang": What are the differences between Xiao Peng XNGP and Tesla FSD in terms of technical routes?

Xiao Peng: There are definitely many differences, but I don't know what the differences are, we don't understand Tesla's technical route now.

We don't focus on it because it's not valuable. Tesla doesn't use radar, while the vast majority in China do. Tesla has its own computing power, while the vast majority in China are based on Nvidia. Different computing power, different sensors, different requirements. In the logic of uncertain hardware and uncertain usage scenarios, the capabilities are very different.

"Qian Wang": What do you think of Tesla FSD entering China?

Xiao Peng: If it can enter China, it would be very good, but it needs to comply with China's policies and regulations. It needs to be trained and tested in China, and the scenarios in China need to be well done. Different countries have different scenarios, so training is different. In the U.S., there are stop lines everywhere, but in China, they are rare.

"Qian Wang": If L4 is achieved, can it bring about business model innovation?

Xiao Peng: If L4 is achieved, it will definitely greatly increase the software income or ecosystem income of the car. For example, automatic car wash is one of the ecosystem incomes; automatic parking is the second; automatic charging is the third. You can charge for software separately, as Tesla is doing. You can also combine software and hardware for charging, as some manufacturers do. They are all charging, just in a centralized way or separately.

Fully entering L4 will move towards the ecosystem, but it will take a few more years.

"Qian Wang": An investor calculated that if FSD's Robotaxi (autonomous taxi) is launched, a Tesla is $30,000, with 20 rides per day at $20 per ride, it can break even in 75 days, with a considerable return on investment.

Xiao Peng: This is thinking with the simplest logic, in the long run it may work, but not in the recent few years. There are various additional problems involved. Also, I completely do not believe that FSD can do anything on Robotaxi in the next year, it may be tested in certain local scenarios, but the possibility of doing it on a large scale better is close to 0 "Qian Wang": Where is the bottleneck?

He Xiaopeng: First, you need a cloud-based management system. For example, I took over Waymo's cloud system twice this time. This management system is completely different in different countries and regions.

Secondly, if you want to achieve Robotaxi, there needs to be a huge change in hardware—most people who work on AI large models have overlooked one thing— in the past, the rapid development of China's internet and mobile internet was due to relatively stable, low-cost, and unified hardware interfaces. Our servers, computers, phones, these manufacturers are only responsible for hardware, not so much for other areas, but there has been a change in the phone industry with Apple.

In the future, truly huge AI companies are inevitably going to be involved in hardware. Similar to how we used to have clear divisions of labor, you were Windows, I was Intel, there was a probability, but the probability is getting lower and lower.

"Qian Wang": How do you view Musk working on Tesla while also running the general artificial intelligence company xAI?

He Xiaopeng: He is doing multiple things at the same time. Many of them are long-term dependencies, not short-term ones.

We have been working on robots for several years, and robots have different requirements for large models, different capabilities. Most robot companies that existed before the first half of last year were using past technologies. The famous American robot company Boston Dynamics used to use hydraulics, now it's electric. In the past, if you wanted a robot to perform complex operations, it was basically unsolvable. The environment is infinite, the actions are infinite. A robot has over 30 joints, you want the hands, eyes, feet, thoughts, and mouth to coordinate. If you let programmers code, they can code up to 10,000, code to death, and can only solve 5% of scenarios. Even 10,000 people can't solve it, because the rules are infinite. You have to use AI for everything.

The logic of AI is quite funny. How do you make a large model see this world, it keeps learning what's right and eliminates what's wrong. Hitting someone is a wrong action, it needs to be eliminated. How do you clean it, train it? Very difficult, even more difficult than autonomous driving. However, this is what other large models will have to do better in the future.

"Qian Wang": Have you met Musk this time?

He Xiaopeng: No.

"Qian Wang": Is he your business idol? Who is your business idol?

He Xiaopeng: I don't have a business idol.

"Qian Wang": You mentioned visiting Jensen Huang last year, what did you talk about? Did you learn anything from it?

He Xiaopeng: Firstly, his company's market value was not high 6 to 7 years ago, he seized a huge opportunity, thanks to his persistence. Secondly, his management may not be perfect, but it is very good management in the United States—holding a meeting with all employees every quarter, very few bosses in China do this; reporting directly to him involves dozens of people, which most bosses cannot achieve; he doesn't fire people easily, his employees stay for 15 years or more, with 15 years being the minimum.

In China, it's crazy, and even in Silicon Valley it's crazy, because many employees in Silicon Valley can't stay for that many years. You have to know that 6 to 7 years ago, it wasn't a superstar, just a star In his insights on technology - not just insights, but insights, and enterprise management, we have many differences.

"Qianwang": What does it reflect that his employees stay for a long time, he meets with everyone, and has many direct reports? Does his management experience inspire you?

He Xiaopeng: In a long-term technical field, there needs to be a small group of people who have long-term belief in this matter. Their company was not that big originally, but they worked hard for a long time. Many large companies in China are "fast food", planting trees for three years, blooming quickly, and bearing fruit quickly, so they nurture different companies.

(From him, I learned) that there is a group of super nice people who need a long-term relaxed environment for long-term scientific research. Many large and good corporate research institutes in China do not do well, which is related to the top leadership. The top leader needs to have enough patience and vision. Nvidia was not doing well a few years ago, dropping to a market value of just over 8 billion USD. You have to keep telling them continuously, what the vision is, why we can win - talk for more than ten years, many people won't believe it, I'm telling you.

This year, I also started to communicate a lot with colleagues. I didn't like it before. If I had to communicate with the team, I would just have a meeting. The effect of a meeting is not as good as a small meeting, but it's better than not having one. We now have more all-staff meetings, more punctual. More small meetings with less than 20 people. Originally, many meetings were related to my business, now many are not related to my business. (Let everyone) understand how I think, why I do things this way, know what questions everyone has - some of these questions indirectly affect me, and a few have a direct impact.

"Qianwang": Did you meet with Jensen Huang in April last year? What is the relationship between Wu Xinzhou leaving XPeng Motors for NVIDIA? (Wu Xinzhou was formerly the Vice President of Autonomous Driving at XPeng Motors and is now the head of NVIDIA's Autonomous Driving China team)

He Xiaopeng: I met Jensen in September (last year). I was the one who sent Xinzhou to NVIDIA, and I went to see him (in China).

"Qianwang": Did Wu Xinzhou's resignation make you very sad?

He Xiaopeng: Both sad and happy. He will get better, and we will get better too. Everyone has their own strengths and weaknesses. In this world, at least in many companies in China, very few are decided by one person whether the company lives or dies, and in the end, all companies can do better - this statement is a bit empty.

"Qianwang": But this is also the pain of executives, indicating that this company can do well with or without me.

He Xiaopeng: A good company should be able to do this, but this is a longer-term matter.

"Qianwang": There is a perception in the AI industry: large model companies will eventually become application companies, models and products are like the south slope and the north slope, no matter from which slope, in the end, there must be both models and applications. Is this view also applicable to autonomous driving technology and the automotive field?

He Xiaopeng: I basically agree. In the future, to do this large AI software, because the hardware is not standard, there are many differentiations, so software and hardware need to be integrated.

"Qianwang": If GPT-5 is achieved, how much will it change autonomous driving?

He Xiaopeng: I don't know the specific changes of GPT-5. If it must be divided, autonomous driving should be the small brain of the robot in the long term, and GPT or similar large models will become more and more like a quasi-brain, more long-term it is a brain - different divisions of labor for a person "Qian Wang": Will the future multimodal large models, autonomous driving models, and robot brains be unified?

He Xiaopeng: I dare not say, it's still a bit early now.

"Qian Wang": Do car companies have an advantage in making robots?

He Xiaopeng: In the future, the vast majority of robot companies should be car companies. Some robots are more complex than cars, and some cars are more complex than robots, but 80% of the logic is the same.

"Qian Wang": How many more years until autonomous driving arrives?

He Xiaopeng: I already have a clear target timeline, but let's not discuss this for now.

Those Enemies

"I don't agree with all these statements"

"Qian Wang": In China, with Huawei entering the field today, its intelligent driving capabilities are considered top-tier, and its style is so aggressive, have you felt the impact?

He Xiaopeng: (laughs) Yes! Of course. I have been closely related to Huawei for many years in my previous entrepreneurial ventures, and I have always believed in Huawei's capabilities. In the end, I believe our focus in some areas exceeds that of Huawei. Focus is a key factor for success.

"Qian Wang": When you look at Huawei today, is it different from how you viewed Huawei in the era of mobile internet in the past?

He Xiaopeng: Huawei is a company that Chinese people can be proud of. As a competitor, it is a very formidable opponent, and we should learn a lot from it. However, when the top two fight, the people below often suffer. There will always be fights, whether in China or overseas, right? It's actually a good thing for us.

"Qian Wang": Some industry insiders also suggest that Xiaopeng, why don't you just stop selling cars like Huawei and become an intelligent provider like Bosch, collaborate with everyone, it's said that Huawei has calculated that the gross profit of making cars is not as good as making phones.

He Xiaopeng: Hmm, it depends on the stage. We will provide some capabilities to a few partners, but not to many. It's already difficult enough to work with one partner, like Volkswagen.

I started my entrepreneurial journey earlier in the B2B sector. I dare say that Huawei is the best company in B2B. B2B has its own difficulties and comforts. Once it stabilizes, it's very comfortable, but stability is hard to achieve.

"Qian Wang": After so many years of focusing on intelligence, have you ever hesitated? For example, did the design of the G9 model car waver?

He Xiaopeng: No, it's just a choice between two configurations. Many people think that the higher-end car should have fewer configurations, but our previous cars had these configurations, and they are actually the same, it's just a matter of perspective. I agree, in the future, we should reduce the configurations and make them more refined.

"Qian Wang": Some believe that not only are there many suppliers for intelligence, all mainstream car companies will do it, and companies like Li Auto and BYD may not have such strong intelligence capabilities today, but they don't think they can't catch up. For you, what is the moat constructed by intelligence?

He Xiaopeng: I don't agree with all these statements.

There are many barriers, including technological barriers. Today, the vast majority of people cannot do AI. What used to be called small AI is now called big AI. The requirements for technology on big AI will further increase, and the probability of success will further decrease.

We have been doing intelligence for 10 years, and only in the past two years have we switched to big AI, previously it was small AI. We invest 3.5 billion in AI per year, including training, manpower, which most companies cannot do The basic network effect of intelligence is that the stronger the AI, the more data it needs, the more data it needs, the more costs it incurs, and the more cars it needs to sell to support this system. This cycle will make it impossible for all supply chains in the future. The supply chain sells you a software, but if it is not continuously trained, who will cover the cost? But sorry, in the future, many collaborations must be long-term.

Large models are not favorable for small companies. Because its time cost, business logic, R&D barriers, and training costs are all barriers. When it reaches a certain scale, there will be data barriers, and then when the data reaches a certain scale, there will be new barriers.

The barriers are very high, but one key point is: can it address user pain points, high-frequency user experiences must be used. If it can achieve not touching the steering wheel for a week in the city, regardless of whether the driver is a boy, girl, old, or young, it will be a strong demand in the future. It's just not done well today, but this demand will be huge. Just like a search engine, do you think its barrier is high? Far less than autonomous driving. But there are not many things that a search engine can do well. I have worked on search engines for several years - wow, it's really difficult! Really, really difficult.

So, you absolutely don't have to believe. What did they do before? They drove well on the highway, and could barely drive in the city - the rules end here.

"Qianwang": There is also a view that large models actually reduce the barriers to entry for autonomous driving. As long as you have money, adapt the foreign open-source architecture well, and the rest is about computing power and graphics cards.

He Xiaopeng: I don't quite agree. At the bottom is computing power, one level up from computing power is the Bill of Materials (BOM), and one level up from computing power is the model or AI Operating System (AIOS). Going up again, on top of the model's AIOS is data. One level above data is the globalization policy of autonomous driving. Above that is the experience. Today, Waymo is only strong in one aspect: the experience.

Can you do it as long as you have enough money? Having enough money is only from the perspective of computing power, below computing power is BOM, and above computing power is OS. No one can directly use a regular large model to run autonomous driving. Everyone can see that today all models are viewed in terms of non-real-time, low reliability execution, but autonomous driving is different, it requires millisecond-level, at least hundreds of milliseconds, or even tens of milliseconds.

Data is very difficult. All tier1 suppliers will encounter a huge problem: where does your data come from? If your data can be obtained from OEMs, why can you share it with other OEMs? The initialization of each OEM is a new start. How do you meet the requirements of globalization and policies and regulations? This is a problem involving hardware, software, and ultimately the experience.

Basically, large models make it impossible for tier1 suppliers who used to do intelligent AI to truly deliver something like an SDK in a 5-year perspective - I don't think they can do it "Qian Wang": Some car companies have money and can invest more capital.

He Xiaopeng: Yes, that's the probability of success. I'm not specifically referring to any particular company—look at them now, they haven't even achieved high-definition map cancellation, you have to pay to buy it, and high-definition maps simply don't provide a good experience. In China, assuming there are 10,000 roads, high-definition maps can cover 50 roads, leaving 9950 roads uncovered, how do you use it? You simply can't.

(Most of the time for car companies), others tell them that this thing can be done, but what are the problems, difficulties, rhythm, and probability of success in doing this thing? Having money just means it can build a foundation, but can this foundation work? Can it work well? It requires capability.

"Qian Wang": Let's talk about a very practical issue. Industry insiders say that in car sales, the intelligent driving label ranks very low in the considerations for buying a car, probably eighth or ninth. Can autonomous driving help you sell cars?

He Xiaopeng: Two years ago what you said was correct, now the ranking will rise, in XPeng's user ranking it's in the top three. Now, for cars priced above 200,000, it ranks around fifth; for cars priced below 200,000, it ranks around tenth. Autonomous driving for cars priced below 200,000 is very poor, in the past, hardware BOM and software capabilities couldn't support cars at this price, so we need to do intelligence well.

But I believe: first, we need to improve autonomous driving by 30 times in the next 18 months—remember, not 30%. If in two years, if you drive every week, except for getting in and out of the car, you only need to steer two or three times, and it's very safe, would you use it? You would form a habit. If you can't do it today, my first goal is to make it happen. Second, I want to achieve another thing, affordability. To be able to use all these capabilities for cars priced below 200,000, this is very difficult.

If autonomous driving doesn't make it to the top three in this field today, there are two reasons— it's not good enough, and not all cars can use it. Once these two reasons are resolved, it will happen quickly. Starting in 2024, the next 10 years will be the decade of intelligence.

"Qian Wang": In the process of improving intelligence, how do you promote car sales?

He Xiaopeng: First, improve the experience. XPeng didn't pay enough attention to the experience in the past, but we have been intensively focusing on it this year, including significant improvements in OTA experience; second, turn itchy points into pain points, because a good experience doesn't necessarily mean it's a user's essential need; third, turn high costs into medium or low costs.

Just like AI and large models, accumulate slowly and erupt. Maybe in 12 months, maybe a little off, 24 months, but it won't be longer. It's very close, approaching the turning point.

"Qian Wang": You mentioned that 2025 will be the ChatGPT moment for autonomous driving, what will consumers see?

He Xiaopeng: Lower costs, wider range, better overall experience. Everyone will see: wow, autonomous driving is actually a good experience. This is the first year of initialization. After that, leading companies will distance themselves from companies that used algorithms or rule logic before, and the distance will grow further and further "Qian Wang": Is it like Waymo?

He Xiaopeng: Yes, because of different accelerations.

"Qian Wang": How much time did you personally spend on the big model in the past year?

He Xiaopeng: It's hard to judge. I often look at things here, often discuss, but if you say 30%? Definitely not. I spend much more time than the CEOs of traditional car companies.

"Qian Wang": Overall, the automotive industry is still a relatively closed industry.

He Xiaopeng: Because I am in the automotive industry, it's not good to say this to other bosses, I dare not say this to you (laughs) — I dare not say many things now.

Can't afford to make competitors unhappy.

They are all big shots.

The G9 accident is a Waterloo, but also my luck

"It's not a hero cutting off his arm, it's a hero cutting off his head."

"Qian Wang": From Xiaopeng's establishment to today, exactly 10 years, what insights do you have at this point?

He Xiaopeng: Brothers, whoever wants to harm someone, let them come and make cars.

Making cars is really tough — they are world-class opponents, the most powerful people, with long chains, long cycles, and high barriers; but making cars is really cool — it can change people's way of life. I used to think that autonomous driving would not happen for many years, even if I was doing it, I didn't think it would. But last year I started thinking slowly, this year I started to see the value — I think it can be done.

"Qian Wang": How is the battle of EVs different from the battle of UC before?

He Xiaopeng: At that time, the competition was not very intense, only Tencent, Baidu, 360, OPPO, not many. The competition in the automotive industry is too fierce! Doing mobile internet is a short chain, data can be iterated, users can be operated. Now it's a long chain, a mistake can cause problems in two years, changing one thing can easily lose 20 million, 200 million, or if a model is gone, 2-3 billion will be lost in an instant — it's too difficult.

Now you have to be very accurate, do very well, and need a bit of luck to have a chance.

"Qian Wang": What kind of battle is this?

He Xiaopeng: I just feel like swimming in a sea of blood now.

"Qian Wang": Is the pressure greater this year?

He Xiaopeng: I've been feeling particularly good recently. I think "steady progress". I am very confident that there will be significant differences in the next few months.

"Qian Wang": Because the 150,000 to 200,000 level car is coming out?

He Xiaopeng: Many reasons, definitely not related to a specific product. In fact, a car is like a barrel. I positioned the market competition a bit earlier than most companies, made comprehensive planning, organization, and adjustments to multiple capabilities earlier, and patched up the shortcomings of the barrel. Plus, the overall market momentum will make some of my strengths more valuable for a period of time, one of which is intelligence.

The overall operational capabilities of the company have been greatly improved — remember, not slightly improved, but greatly improved. Give me another 18 months, and my operational capabilities can compete well in the sea of cars, both in China and overseas.

"Qian Wang": Do you have any memorable ups and downs swimming in this sea of blood in the 7 years since joining Xiaopeng Motors?

Interview with Xiaopeng He: Finding Balance in Company Growth

Xiaopeng He: When your cars sell well, and many people tell you, "I really like your car," you feel great. But when strangers on a plane shake your hand and say, "Are you Xiaopeng He? Keep it up, I bought your stocks," that's where the pressure lies. You can create a technology like NGP (Navigation Guided Pilot, intelligent navigation-assisted driving), and I was excited about it. No one can turn highways into fully automated driving, everyone learns from you, you drive a wave of technological advancement, and you feel great. But when you don't do well—one mistake we made before is: never neglect the human aspect, "one foot on the gas, one foot on the brake."

Car manufacturers need a lot of people, it's easy when sales are good, to bring in many people with "one foot on the gas"; when facing challenges, to kick people out with "one foot on the brake"—this is wrong. We have made this mistake more than once. I hope to make fewer mistakes in the future, because if you don't "proceed steadily to far," your employees will be affected. Cars require too many people, costs are too high, unlike the internet industry. Oh, why are manufacturing workers suffering, it's true.

"Qian Wang": "One foot on the gas, one foot on the brake," do you feel indebted to your employees?

Xiaopeng He: Of course, that's not right. But for a company to survive, you must first ensure survival, you have to do right by the majority of employees. From a different perspective, you shouldn't hit the gas pedal hard first, you should proceed slowly and steadily. "Stability" is too important.

"Qian Wang": So what should you do? If not "one foot on the gas, one foot on the brake," can you give a specific example?

Xiaopeng He: When things are good, don't think everything is good, when things are bad, don't think everything is bad. We made a mistake before, increasing the number of employees from a few thousand to 10,000 in just over a year. It was too difficult! A company cannot absorb that. If you have 100,000 people, and you hire 10,000, maybe it's OK, it's 10%. But if you have 3,000 people, and you hire 7,000 to 10,000, it's very challenging.

"Qian Wang": Why did Xiaopeng plan to recruit 4,000 more people this year?

Xiaopeng He: We need that many people. We are currently recruiting steadily, and we will develop faster in the latter part of this year and next year. We won't have a large-scale recruitment, nor will we have large-scale layoffs—I dare not say 100% won't happen, but I will try to ensure it doesn't.

"Qian Wang": The G9 incident is often mentioned in Xiaopeng's growth journey...

Xiaopeng He: One of the Waterloo moments.

"Qian Wang": You mentioned last year that 12 executives were reduced to only 2, how was that decision made?

Xiaopeng He: Actually, there are more executives in our company.

Xiaopeng's executives all want to do great things, are very responsible, and pursue efficiency. However, to make a car that can be like this, to make a car that is very good, flawless, strong in every aspect, requires professional business and management capabilities. Some are good at 0 to 1, some excel at 1 to 10, and some are great at 10 to 100. Xiaopeng Motors is in the 1 to 10 range, in the 1 to 2 phase Special thanks to them, who have done a lot at XPeng. Some retired, some started new businesses. However, as XPeng transitions from 0 to 1 to 1 to 2, there will be methodological adjustments, including my own methodology.

"Qian Wang": What has been the biggest change for you in the past two years?

He Xiaopeng: Stability.

"Qian Wang": What is your division of labor with Fengtou (Wang Fengying, President of XPeng)?

He Xiaopeng: She manages products, she manages sales. I handle most of the rest. I'm not giving a comprehensive answer, but roughly that's how it is.

"Qian Wang": What is your reporting relationship?

He Xiaopeng: We often have meetings together, there is no reporting relationship, we are all part of the core management team. And there's Brian (Gu Hongdi, Vice Chairman and President of XPeng), we discuss many things together.

"Qian Wang": How did you learn from the failure of G9 and find new methodologies for X9?

He Xiaopeng: There are many details, such as simplifying configurations.

Why am I particularly confident now? We have found a series of self-renewal, versatility, and maximizing our strengths to super strengths in thinking and some behaviors. Many people care about sales volume. Even if you increase sales by 5000 units/month, by 10000 units/month, you won't survive in the long run. Increasing sales a bit just makes you do better today, but you need to consider why you can increase sales in the future. It's not like the internet, where cars have high volatility and strong homogeneity in competition. The internet is cost-free, but cars have costs.

More should be considered on how to transition from elimination rounds to all-star games, what is your own logic? What capabilities do you need? What weaknesses do you need to address? How much revenue do you need? How much sales volume? How much gross profit? How much profit can support your R&D at that time? - You need to think from that perspective.

Many short-term actions done now are of no use. It just means, it keeps you alive for a few more months.

"Qian Wang": Please answer these questions yourself.

He Xiaopeng: Sorry, I won't answer that, it's our company's business.

We must think about it, reverse this matter. So, G9 was very lucky for us. At that time, we probably had 5-6 mistakes combined, if we made 1-2 fewer mistakes, maybe sales would have increased by 5000 units/month, and we would have felt that we just didn't do exceptionally well, but not particularly bad - remember! With this mindset, the transformation is different.

Many people now have sales, but still find it difficult to survive the transformation period. The larger the company, the harder it is to self-revolutionize, self-transform. We are a company of over 10,000 people, and it's already very difficult!

"Qian Wang": 16,000.

He Xiaopeng: Just think about how difficult it is for all of us.

Cars are changing in a non-stable state - this is the difficulty of transformation.

"Qian Wang": Some say that after the G9 incident, you were a "hero who cut off his arm."

He Xiaopeng: (Silent for 5 seconds...) Haha, you can understand it that way

Turbulent Times are Coming

"Whether you are a bear, a hero, or a hero, you may become a xióng"

"Qianwang": Why did so many car companies choose to start alliances last year? You and Volkswagen, Didi formed an alliance, Nio and many car companies formed alliances (providing battery swapping services), and Huawei took a radical step by making its car BU independent. Why did a series of alliance actions all happen in 2023?

He Xiaopeng: Because turbulent times are coming. Before the turbulent times, everyone is forming alliances, just like the Spring and Autumn Period and the Warring States Period. In the Spring and Autumn Period, there were hundreds of states, while in the Warring States Period, there were only a few, and finally unified under the Qin and Han dynasties.

But in the automotive industry, it is impossible for 1-2 companies to dominate the market.

"Qianwang": The automotive industry is not a winner-takes-all industry.

He Xiaopeng: Even if 7-8 companies remain, it is still a winner. It may not dominate 100%, but it can dominate a large proportion of the market.

A major change means that whether you are a bear, a hero, or a hero, you may become a xióng—just like when we say China is the cutest panda. (laughs)

"Qianwang": Why did you choose Volkswagen and Didi as alliance partners? Was it an active choice or a passive one?

He Xiaopeng: It was a mutual choice. Volkswagen and Didi are both excellent. Volkswagen is one of the largest global car companies, with strong hardware and globalization. Most people in China underestimate the difficulty of globalization. Many car manufacturers have failed, with countless lessons learned... Many overseas competitors have many thoughts, which may be right in their environment but wrong in yours; and your thoughts may be wrong in their environment, with many interesting perspectives.

I visited a Volkswagen factory some time ago and came back with a few ideas to implement. But now I don't have the energy. I plan to do it in the second half of next year.

"Qianwang": What are those ideas?

He Xiaopeng: We want to make the testing logic in the factory more fully automated and rich, to ensure a significant improvement in the quality of the cars leaving the factory. When you initiate something, you can't do it this way; when you stabilize something, you can do it this way. Because there are too many changes during the initiation phase, the cost-effectiveness is not high, but the more stable you are, the more you can improve the quality several times.

"Qianwang": There is a saying today that Xiaopeng has a safety net, being acquired by Volkswagen. Do you agree with this statement? Are you open to this possibility?

He Xiaopeng: I definitely won't answer that question.

"Qianwang": Xiaopeng recently launched the MONA series new car M03. What is the significance of this car for Xiaopeng and for autonomous driving?

He Xiaopeng: This car is "unique and beautiful under 200,000 yuan". MONA will enable Xiaopeng to enter the high-level autonomous driving market below 200,000 yuan.

In the past, Chinese car manufacturers have been competing in the range of 200,000 to 500,000 yuan, and no powerful car manufacturer has entered strong competition below 200,000 yuan—this represents a turning point for Xiaopeng as well.

"Qianwang": What are your expectations for MONA after its launch?

Translation:

He Xiaopeng: Will have a great impact on the entire industry - oh, it turns out that cars under 200,000 can also achieve high quality; oh, it can also be made highly intelligent; oh, it is worth spending so much money and time to do it well. In the past, cars under 200,000, in order to control costs, more or less switched from oil to electricity.

"Qian Wang": Everyone now compares the launch of new cars by car companies to "playing cards", why did you play this card this year? How do you view Nio's card play, while Apple doesn't play cards - evaluate the playing styles and cards played by each company.

He Xiaopeng: It's all related to their own layout. Our logic for playing cards is, seeing that autonomous driving will bring about a "qualitative change" and can appear more affordably, so it is possible to launch cars in the 150,000 to 200,000 range. This is about technology and cost rhythm.

In the past, when we wanted to launch cars in this price range, we didn't have the ability to make it profitable, but now we can, this is the improvement in system capabilities. If you don't limit the BOM, it's not difficult to make a car with many functions; but if you limit the BOM and still want to make a good car with many functions, it's very difficult.

"Qian Wang": Will Xiaopeng continue to make a further push into the mid-to-high-end market?

He Xiaopeng: Of course. We have never stopped in the mid-to-high-end market. The X9 is the first pure electric MPV. We will definitely do better in this market.

"Qian Wang": In your opinion, in the midst of major changes, how many companies can survive in the elimination round? How many can survive in the All-Star round?

He Xiaopeng: I estimate there are still more than 10 companies in the elimination round. In the All-Star round, there are 7 companies, everyone ranks, whether you can be first, third.

"Qian Wang": What rank can you achieve in the All-Star round?

He Xiaopeng: My goal is first. At lunch today with the core team, I said, if you don't reflect on yourself according to the logic of being first, don't even talk about being first, you can't even achieve top three - you must aim for first.

A lot of our strategic layout is - why can we be first? If you are the leading company, what are the few things that need to be done first? What can be in the top three? What can be in the top five, top ten? What things must be done? - You have to prioritize. Reflect in this way. Who do you think is currently the number one in this industry? What kind of capabilities are in the top five of this industry? Reflect on whether your capabilities are in place.

"Qian Wang": How to achieve first place?

He Xiaopeng: Now, actions speak louder than words.

"Qian Wang": What kind of stage is Xiaopeng in today?

He Xiaopeng: It is at a stage of experiencing headwinds, undergoing self-transformation, accumulating capabilities again, and entering the eve of tailwinds. Comparatively, G9 allowed us to see, think, and act earlier. Misfortune and fortune always depend on each other.

Entrepreneurs have a big heart. When a problem arises, the first thought is how to solve it? Feeling fortunate to have seen it, all these problems can be solved. But when something good happens, it can also be painful to turn around, oh, can this good thing be stable tomorrow? - This is about constantly accumulating capabilities, supplementing weaknesses, and turning strengths into more stable or longer strengths.

Entrepreneurship in the automotive industry - bitter, tough, easy to be sad. But now I am in a very comfortable, stable, and upward stage. We did a lot last year, and only now at the end of this year did we play our first card. In the second half of next year, the first half of the year after, or the second half of the year after, the cards played will be better and better.

"Qian Wang": What is the most concerning issue for you at the moment?

He Xiaopeng: I mentioned 2+3 issues internally. 2 major issues - the first is strategy, planning, OKR, and the second is management, organization, system.

3 minor issues - the first is the mindset of operation, which should be company-wide; the second is customer experience, looking at it end-to-end rather than from one's own position; the third is holistic thinking. Many times we say we want to help customers, we do something right, but looking at it over a longer period, many decisions are wrong. In the past, we made too many isolated decisions, holistic thinking is difficult because the automotive chain is long, time is long, and the cost of mistakes is high.

Why are many automotive professionals afraid to make decisions? Because if you make a mistake, sorry, these 30 million, these 12 months, have been wasted. The internet is different. Oh, made a mistake, come on! Let's fix it together and get it done tonight.

These are all positive strategies.

"Qian Wang": I heard you transitioned from democracy to centralization?

He Xiaopeng: Relatively centralized. During times of transformation, many things require quick decision-making, quick execution, and quick optimization.

"Qian Wang": Why did you lose your temper recently?

He Xiaopeng: I've lost my temper quite a bit recently. Many are minor temper tantrums.

On a larger scale, we have some professional teams that can achieve higher scores in certain areas, but they feel they are already doing well - I am very dissatisfied with this attitude - we need to challenge ourselves, have higher expectations of ourselves, rather than making excuses.

"Qian Wang": Have you ever felt despair swimming in a sea of blood?

He Xiaopeng: (10 seconds of silence...)

As a founder of a company, you have to envision a direction for yourself, envision brightness. Sometimes you may not see the direction and brightness, but you have to be able to imagine it and lead a group of people towards it.

So, I am not in despair.

Author: Zhang Xiaojun, Source: Tencent Technology, Original Title: "He Xiaopeng Talks About Systematic Transformation of Large Models in Intelligent Driving: Making cars is like swimming in a sea of blood, but I am more confident now"