news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Now is not the time to compete in the driverless industry.

Text｜Tian Siqi
Source: Jazzyear (ID: jazzyear)

When Tesla released its second quarter financial report, Musk officially announced that the launch of its Robotaxi product would be postponed from early August to early October. Hou Xiaodi, who also rooted his business in Texas, USA, like Musk, once again confirmed his identity as a "prophet".

Previously, Hou Xiaodi was best known for the journey of TuSimple, the "first stock in autonomous driving" that he co-founded, from trust to suspicion, and from cooperation to separation.

After going public in April 2021, Hou Xiaodi's net worth once exceeded US$1 billion. However, with the initiation of investigations and supervision by the Committee on Foreign Investment in the United States (CIFUS), Tucson has experienced a turbulent change of management rights. Hou Xiaodi was first promoted from the position of CTO to CEO, and then was removed by the board of directors in October 2022; then he and another co-founder removed the former board of directors through a super vote 10 days later, but never returned to the company's management; in March 2023, Hou Xiaodi finally announced that he had completely left Tucson.

Before delisting in February 2024, Tucson's stock price also plummeted from a high of nearly $80 to below $1.

Some say that Hou Xiaodi's relationship with Tucson is like Oppenheimer's role in the Manhattan Project. Just as Oppenheimer had to face doubts and scrutiny from all sides in the post-war years, Hou Xiaodi's decisions and actions at Tucson are always under the microscope, and every choice he makes may become the focus of controversy.

Even so, Hou Xiaodi still proudly tells "Jia Zi Guang Nian" today that Tucson is the world's first company to test fully driverless heavy trucks on open roads, and has achieved the first half of the driverless game - creating a system with safety redundancy.

In May this year, CIFUS reached a settlement with Tucson. According to Tucson’s announcement, the investigation did not find any violations by the company, including the previously rumored “technology transfer”. After leaving Tucson, Hou Xiaodi once again started a L4-level driverless truck startup and established a new company, Bot.Auto, in the United States.

When L2 assisted driving gradually replaced L4 unmanned driving in public opinion, Hou Xiaodi insisted on L4 research and development, driving on this cautious and brave road. But he saw the direction different from the mainstream as an exciting opportunity: "Either be a good student who scores 100 points every day, or use the once-in-a-lifetime opportunity to impact a revolutionary vision. Even if there is more uncertainty and less industry consensus, I am willing to do a miracle with a small probability. Of course, everyone in our team has unique skills and we are united. I believe we can make this miracle happen."

His revolutionary vision is to make driverless trucks operate at lower costs per kilometer than humans based on the concept of Transportation as a Service (TaaS), thereby creating more social value. From a financial perspective, the node for the realization of his dream has been preliminarily determined to be the first half of 2026.

In this exclusive interview, Hou Xiaodi commented on the delay of Robotaxi and industry events such as the domestic "Carrot Run" trend, and made many judgments that may not please everyone, such as "the public has a lot of criticism about driverless cars because everyone thinks they can drive."

In the tide of the times, Hou Xiaodi and TuSimple's story is destined to become a tragic turning point. This reminds him that creation and protection are equally important and protection is more difficult, so this time, he chose the role of guardian. "In the past, I put too much of myself in a technical position, and strong technology is equivalent to one beauty covering a hundred ugliness," Hou Xiaodi said, "but now I still have to turn myself into a hexagonal warrior to protect the company in all aspects."

The following is the full interview, which has been edited:

Safety Challenges of Robotaxi and L4

Jiazi Guangnian: First of all, I would like to congratulate you on once again becoming a true prophet regarding the Robotaxi delay.

Hou Xiaodi: Then everyone should get used to it. I have made many predictions. I have said before that scaling law is not the answer and I firmly oppose scaling law-only doctrine. For example, at the GTC conference, everyone thought NVIDIA was omnipotent, but I felt that NVIDIA's performance improvement in edge computing has almost reached its limit.

I often pour cold water on people when they are enthusiastic, which essentially reflects my different attitude towards future expectations. If you want a bold prediction, you just need to be optimistic. But if you want an accurate prediction, you have to base it on knowledge and logic. Take high school physics as an example. Amateur scientists can talk about turning water into oil or perpetual motion machines, but as long as you have learned the conservation of mass and energy, you will know that he is fooling around when you hear about perpetual motion machines.

For me, when I see a prediction, what I really need to think about is the theoretical support behind the problem. What physical and mathematical laws will constrain the future development trend of technology? We should respect these limitations.

Jiazi Guangnian: You said before that Musk's Robotaxi could not be made, and now that it has been postponed again, it has partially verified your statement. What do you think of the decision to postpone it?

Hou Xiaodi: The current performance of Robotaxi is far from the real L4. A few months ago, a friend of mine praised Tesla so much, so I suggested that he take a test ride on Waymo. As a result, he said to me the same day: "Xiaodi, I have taken a Waymo, Tesla is really rubbish."

(Note: Waymo is an autonomous driving company under Google's parent company Alphabet, which has launched driverless taxi services in California, Texas and other places.)

Robotaxi is not something that can be released by fixing one or two bugs or delaying it for two months. So I continue to predict that October 10 will still not be the time for the real release of Robotaxi, and even next year will not be the time when it can be used. At most, clues about when Robotaxi will be released in October this year will be announced.

Jiazi Guangnian: Why can you feel that Tesla Robotaxi cannot be made after riding in Waymo?

Hou Xiaodi: Waymo's behavior can give you stronger certainty. When I test drove my wife's Tesla, which was the latest version of FSD, I saw that many behaviors jumped a lot, and it was not designed from the perspective of minimizing traffic accident liability. Many algorithm details were poorly done, which made me feel that it was not at the true L4 level.

In addition to the technical essence, there is also the commercial essence. There is a joke that what if someone is killed while using FSD on a Tesla? My answer is a pun: It's not their business. In terms of the division of responsibilities, Tesla certainly said that the driver is fully responsible (the system clearly prompts that when using FSD, the driver must keep his hands on the steering wheel from beginning to end and continue to pay attention to the road conditions), and it has nothing to do with us; secondly, this is not its business model. Tesla's business is to sell cars, and FSD is the added value of selling cars. If you want to consider how to sell more cars, you can't go deep into a limited area like L4 and solve all corner cases in this area.

Jiazi Guangnian: So you think the logic is not to minimize traffic accidents, but to maximize commercial benefits while minimizing Tesla’s liability.

Hou Xiaodi: First, minimize Tesla's liability, and second, maximize commercial benefits. There are three scenarios: best case, average case, and worst case. Tesla optimizes the best case scenario. Assisted driving itself has its value, and as long as consumers recognize it, its mission is accomplished.

Jiazi Guangnian: Musk did not introduce too many details on how to seek regulatory approval at the earnings conference, but he always emphasized that Tesla is a so-called universal solution, which may be better than Waymo's more localized solution. Waymo's solution is too fragile.

Hou Xiaodi: This is a typical sophistry. How do you define vulnerability? I can run all over the world and do it in every city, but you can only do it in a small area, so you are vulnerable. But the question of the L4 competition is whether you can do it well in this small limited area. As a result, he said that I will not compete with you in this, but in that. This is like asking who is more beautiful, Xu Gong from the north of the city, and he insisted that I will not compete in beauty, but in who is smarter; what if Xu Gong from the north of the city is smarter and more beautiful than you? He said that I run faster. Anyway, it is not a fair competition in a limited environment.

There is no technology that can be used for autonomous driving anywhere in the world, just like there is no athlete in the Olympics who can do both long-distance and short-distance running. If you are working on a limited scenario, do it well. If you want to sell cars all over the world, don't set up a memorial and say that you are at the level of L2+. I don't like to confuse the public through marketing. The more noise there is, the more pressure there is on those who do serious work.

Jiazi Light Year: Currently, the technical challenges of L4 are still quite large. What specific aspects do they refer to?

Hou Xiaodi: Generally speaking, it is not difficult for those who know how to do it, but it is difficult for those who don’t. I think our team can overcome the difficulties that others have seen one by one, and make some additional innovations and developments based on the paper.

The core of autonomous driving L4 is how to complete a stable system, especially using unstable modules to complete a stable system. For example, if the camera fails to detect an object in one frame, this will definitely happen. How to ensure that our system does not crash is the key.

The dumb way is to reduce the failure rate from once every 100,000 kilometers to once every 1 million kilometers, so that the system is stable. The smart way is to innovate at the architecture level, so that when one module fails, there are other modules to back it up. We have several more layers of back-up solutions, so that we can handle failures at any time in the system. Civil airliners are a typical redundant system. Aircraft usually have two engines, and they are independent of each other in terms of functional safety design. The safety of the system is achieved through a more advanced architecture, not by creating a single engine that never fails.

Jiazi Light Year: That is to prepare for safety redundancy.

Hou Xiaodi: Specifically, it is to make safety redundancy feasible in engineering. Under the current level of technology, using mature components provided by the industry ecosystem to build a system with safety redundancy is the most important task for autonomous driving.

Some people also say that L2 can gradually develop into L4, but I disagree. We spent 70% of our time on L4 research on redundant systems. No matter how good the L2 function is, it can only complete 30% of the L4 function at most, and the remaining 70% of the functions are not touched. Tesla is not an L4 company, but an L2+ company. Many of its designs are not for safety or redundancy purposes.

Transportation as a Service‍

Jiazi Guangnian: You have always insisted on the L4 route, but now the industry and the public seem to be discussing L2 assisted driving more than autonomous driving.

Hou Xiaodi: The success of Bot.Auto and I has nothing to do with whether it is mainstream or not. We don’t need everyone’s approval. If we can create real economic and even social value, that’s enough.

The technical route is determined by objective laws, and the correctness of the route has little to do with whether it is a consensus. This is especially true in the field of autonomous driving. The public has a lot of criticism about autonomous driving, not because everyone is an autonomous driving practitioner, but because everyone thinks they can drive. If you think about the "consensus" of autonomous driving technology discussed by a group of human drivers, it must contain a lot of unreliable things. If we really develop according to the judgment of the public, it will be a dead end.

Compared to people who can drive cars, there are far fewer people who can drive rockets. Therefore, the field of rocket manufacturing is relatively lucky, and there are not many tragedies where lay consensus guides the technical route of experts. From this perspective, the LLM field is the hardest hit area of ​​this kind of "route noise".

This kind of route noise, in turn, has a great impact on the practitioners of autonomous driving, and even harms them. Should they bear the pressure and oppose the noise, and stick to the long-term goal, or should they first say what investors like to say, and get financing to survive and meet the short-term goal? Most people actually chose the latter.

Either be a good student who scores 100 points every day, or use this once-in-a-lifetime opportunity to impact a revolutionary vision. There is more uncertainty and less industry consensus, but I am willing to do a miracle with a small probability. Of course, everyone in our team has unique skills and we are united, so I believe we can make this miracle happen.

Those who invest in what investors like and talk about consensus views every day are just enjoying the process of entrepreneurship. But I don’t enjoy the process of entrepreneurship, I only look at the end. In the end, this business can be accomplished, and the light of hope emitted by this ultimate vision is attractive to me. To get things done, we must respect the inevitable laws of science and business, and we don’t need to cater to public perception. In many cases, we even have to withstand the tremendous pressure of violating public perception and stick to the truth.

Jiazi Guangnian: If something is different from the public perception, does it increase its appeal to you?

Hou Xiaodi: Being different from the mainstream is just the beginning. The next essential work is the truth-seeking process based on facts. If in the end, this "differentiated cognition" is logical, then it can be said to have great value and potential. But discovering different cognitions means new opportunities! Of course I will be very excited. A typical example is that all companies in the United States that make driverless trucks are SaaS, which is the Software as a Service model. Of course, they can also package it as Hardware as a Service, but in essence, they are all selling semi-finished products. Only our company, Bot.Auto, is Transportation as a Service, and we sell transportation services.

Jiazi Light Year: How to define Transportation as a Service?

Hou Xiaodi: If you have goods, I will deliver them for you. Our customers do not need to know whether the goods are delivered by unmanned or manned vehicles. Anyway, we have sufficient transportation capacity and affordable freight. Others are selling unmanned driving software, but the problem is that there are only 20 customers in the United States who can buy unmanned driving software.

Jiazi Guangnian: What is your revolutionary vision under the concept of Transportation as a Service?

Hou Xiaodi: Make the operating cost per kilometer of autonomous driving lower than that of humans.

Jiazi light year: When can it be achieved?

Hou Xiaodi: The first half of 2026. We still have two years to try our best.

Jiazi Light Year: Is this time point an industry consensus?

Hou Xiaodi: Of course it is not a consensus. No one believes it, and peers are not able to give a time point. For example, the story told by Waabi (an autonomous driving company founded in 2021) is that simulation is the best in the world, everything is just like the real world, and there is no problem that simulation cannot solve. But I don’t believe this story, because the last 1% of the problems in simulation are the most difficult. It is not economically efficient to be absolutist on this point.

Of course, this is just my technical judgment, and then the path judgment is deduced. To be fair, everyone lives in his or her own beliefs, and the process of entrepreneurship is to realize one's own beliefs.

Jiazi Guangnian: You said that many people are criticizing driverless cars because they think they can drive. Is this similar to how after the college entrance examination everyone is criticizing the composition topic instead of discussing math problems?

Hou Xiaodi: That's right. Most people can see the similarities between L2 and L4 in terms of phenomena, but they cannot see the gap between L2 and L4 in terms of system architecture and design. Airplanes and rockets are both flying in the sky. It is easy for laymen to see that airplanes are flying faster and faster and will replace rockets in a few days. But if you have studied physics, you will know that airplanes cannot fly out of the atmosphere, but rockets can. Outsiders' criticism relies on overly optimistic extrapolations, while real experts will return to the essence of physics.

Jiazi Guangnian: The Chinese app Luobot has also been very popular recently, and people have been criticizing it on both the technical and social levels. Have you been following its operation?

Hou Xiaodi: First of all, I have to say that this is a good thing. I applaud Baidu. The early promoters who stood in the spotlight faced huge misunderstandings from the public. The process of educating the public is beneficial to the entire industry.

But I still have to say that discussions at the social level are extremely boring, extremely disorderly, and extremely ineffective, so let's not get involved. Alan Kay (winner of the 2003 Turing Award) said that people always overestimate short-term breakthroughs in short-term technology, but underestimate the long-term impact of technology. How much impact can the deployment of 400 driverless cars have on taxi drivers? The deployment of 400 cars today has a good effect, but if anyone is worried that there will be 40,000 cars deployed tomorrow, it is purely groundless. At present, no one has the ability to promote on such a large scale, which is precisely the technical productization problem that driverless driving needs to solve.

Thinking that driverless cars will be available tomorrow is a manifestation of overestimating short-term breakthroughs. At the same time, the public tends to underestimate the long-term impact. For example, will our tax structure and travel habits change in the future? What changes will be made in urban planning? More suburban areas may become more livable, and people can work on their commutes. This is the long-term impact that really needs to be considered.

Jiazi Guangnian: As you said before, we will imagine whatever is in science fiction. But now there are so many things that can be done with mobile phones, which would not be written in novels 100 years ago, but it has indeed brought significant changes to our lifestyle.

Hou Xiaodi: Technological invention is a single point breakthrough in the complex network of human society, and a single point breakthrough will induce a large number of complex related changes. The limitations of human imagination may make it impossible to think clearly about so many related factors, so we cannot blame science fiction for not being specific enough. We must admit that these connections are very complex, and it is difficult for us to predict what the social form will look like due to new technologies, but we should not go to extremes in thinking and shout about unemployment when a new technology comes out. As a social individual, there is a risk of unemployment whether there is new technology or not. As a society as a whole, new technologies bring about broader social changes. This change includes the rise of new industries, the decline of old industries, and the migration of employment. It is by no means a simple and crude unemployment problem.

End-to-end, rule-based and principle-based

Jiazi Guangnian: How do you understand the end-to-end solutions that are currently being promoted in the autonomous driving industry?

Hou Xiaodi: Many things are confused by people who don't know the subject. It's like you tell Guns N' Roses that you are a heavy metal fan and that you are their favorite. Then Guns N' Roses will be angry with you because they are not heavy metal but hard rock. AC/DC is also hard rock, while Judas Priest, Black Sabbath and Metallica are heavy metal.

People always ask me, are you end-to-end? To maintain academic rigor, we are really more than end-to-end. But what they actually want to ask is, have you adopted more advanced technology? Of course we have. The current situation is that "end-to-end" is not an academic concept, but an emotional carrier. People who don't understand the field can easily pin all the unfulfilled dreams and unfinished business in the field of machine learning on the specific technical concept of "end-to-end".

Jiazi Guangnian: So you think the concept of end-to-end has been replaced by others, and people will equate end-to-end with high-tech and precise.

Hou Xiaodi: The advantage of end-to-end is that you can think of a part of the neural network as modular, and the communication and connection bandwidth between modules is very high. The second advantage is that error propagation can be achieved during module training, and the error can be transmitted to earlier modules. For example, in motion planning, it can directly affect the training of perception. But the end is not the means. You don't have to stick to the specific technology of "end-to-end", many other technologies also have such characteristics.

And I must criticize the view that "connectionism solves everything". There is always a misconception that everything from end to end must be neural networks. If I add something that is not a neural network, I will not be pure and revolutionary enough. This makes people feel uncomfortable. The three major schools of artificial intelligence - connectionism, behaviorism and symbolism, each have their own strengths. We cannot use one doctrine to solve all problems. We must be inclusive.

In this sense, I really like DeepMind's style. For example, AlphaGeometry recently solved the problem at IMO very well. From AlphaGo to AlphaGeometry, DeepMind has always been inclusive. It has played with all schools of AI. It does not get involved in the debate of ideologies, but builds systems to ultimately solve problems.

(Note: Google's DeepMind used AlphaProof and AlphaGeometry2 to reach the level of a silver medalist in the 2024 International Mathematical Olympiad)

Jiazi Lightyear: Is the lack of end-to-end explainability a major obstacle?

Hou Xiaodi: Yes! What's even worse is that some people package the lack of explainability as something they are not ashamed of, but proud of. I have to oppose this even more. Musk says every day that we don't write code anymore. But his words are just advertising and marketing. Engineering practice is much more complicated than what he said. How many engineers have been tired of NaN (Not a Number) caused by non-convergence in the training process?

At present, the actual training of an end-to-end network is still a "workshop-style" development model full of mystery and requires manual polishing by "experts", rather than the development model of an automated factory. To see which details need to be improved, you must first freeze other irrelevant parts of the network, and then prepare "appropriate" training data like a Chinese chef cooking, and then continue training and testing iterations.

The most extreme argument I have heard is that the distance of autonomous driving can be conquered, which is a corner case of tens of millions of kilometers and a large amount of computing power. This is a typical case of living in one's own fantasy world. If the person responsible for listening to the story (investor) has such a strong subjective attitude first - there is a saying that if the leader is good, the followers will be even better - then think about how outrageous the person responsible for telling the story (entrepreneur) can make up around "data-driven" and "end-to-end" vision. The two complement each other, and this is how bubbles are created.

Jiazi Guangnian: There are many car companies in the industry that have stated that their end-to-end solutions will be available soon.

Hou Xiaodi: Musk is talking about a set of rhetoric that cannot be falsified in the short term. Don’t ask why we can’t do Robotaxi end-to-end. The answer is, don’t worry, we are training the next version. Now it’s just that the computing power or data is not enough. I posted a picture on WeChat Moments before, with a carrot hanging in front of a donkey. This is the status of Robotaxi and Tesla shareholders.

Image source: X

Moreover, Musk is a successful person. Believing in Musk and following Musk is also the political correctness of the current venture capital circle. On the other hand, this unfalsifiable rhetoric is very effective and suitable for all ages. Second-rate and third-rate technical teams can perfectly explain their technical gap as a gap in data and computing power. If the data and computing power are not enough, then you just need to pay! This is a perfect closed loop.

Jiazi Guangnian: Are there any other schools of thought besides theirs? Do you think you are more of a traditional rule-based school?

Hou Xiaodi: Rule-based itself is an unclear concept, and people can be very subjective about what rule-based is. This makes it easy to abuse it, and eventually it becomes a "weapon" to express subjective emotions, rather than a "tool" to explore objective logic. My former colleague, Wang Naiyan, former chief scientist of Tucson, mentioned principle-based, rather than rule-based, in his blog, and I agree with this.

People have stigmatized rule-based methods, as if everything rule-based is composed of tens of thousands of if elses. But if you really want to argue, the neural network itself is actually a huge rule-based system. Open the network, there are linear operators and nonlinear operators. Each nonlinear operator is a specific implementation of if else, and the neural network is a huge rule-based system composed of billions of if elses.

If anyone in the industry has written tens of thousands of if else statements using a programming language, it is simply because his engineering level is too low. Under no circumstances should a rule-based system allow tens of thousands of if else statements to exist. This is the bottom line. The second bottom line is that, for example, when a vehicle can cross a double yellow line, one, two, or three conditions must be clearly stated. This is what I call rule-based. I have a strong opinion that as long as I am still leading the team to develop L4 autonomous driving, our system must include rule-based, or principle-based. I will not allow an L4 system that is completely non-principle-based to be on the road. This is the bottom line for safety.

The foundation of our system is a foundation model. But on this basis, we emphasize the design and implementation of explainability. We call this framework Foundation to all (F2A) internally.

Jiazi Guangnian: You have been pursuing complete explainability.

Hou Xiaodi: L4 must be explainable and cannot rely solely on unclear data black boxes.

There are three elements here: requirements, design methods, and engineering implementation. Explainability is a requirement. Whether it is principle-based or learning-based, they are all design methods, and writing if-else is engineering implementation. Requirements exist objectively. Don't deny requirements just because the methods have limitations. That's why I said that L4 must be explainable.

There is a high probability that different design methods are intertwined. Its value lies in highlighting the main contradictions, not in arguing. If we argue, all neural networks are rule-based systems. Finally, the quality gap between engineering implementations is greater than that between humans and dogs. We cannot use poor engineering implementations to negate the direction-guiding role of design methods.

What Trucks and Robots Have in Common

Jiazi Guangnian: What are the differences in the technical challenges between trucks and passenger cars? Why do we feel that the progress of autonomous driving for trucks seems to be slower than that for passenger cars?

Hou Xiaodi: Trucks are closer to commercialization, which is why I want to continue working on unmanned trucks after leaving Tucson. Trucks are less popular than cars mainly because VCs like to invest in things they can touch every day, such as humanoid robots, but the traffic rules for trucks are much clearer than those for cars. They can correctly define what actions I should take under what circumstances, which is the most important thing for me.

Jiazi Guangnian: Let me ask you a question off topic first. What kind of robot do you think has the most commercial prospects?

Hou Xiaodi: The current pursuit of humanoid robots is like: you see, technology has advanced, so we should look at this field; there is a labor shortage, so we should look at this field. At present, the companion robots that everyone can generally agree on are still far away, and the current application market of robots is in factories. Many practitioners think the same as I do. The early application scenarios of robots are all certain scenarios, so why do they need legs? But legs are indeed a communication explosion point that can make the public have infinite science fiction associations.

Jiazi Guangnian: That is to say, VCs are more interested in to-C products like sedans and cool products like humanoid robots, while trucks don’t seem cool enough. But the investors who contacted you at least have some interest in trucks. What kind of people are they and what are their characteristics?

Hou Xiaodi: Investors who are interested in trucks are more pragmatic. Today, more people are interested in unmanned driving, and more are interested in trucks than cars. In the past few years, more people were interested in cars. Now the focus has narrowed. Investors who can concentrate on autonomous driving of trucks focus on whether it can make a profit. That is very straightforward. We just look at the operating costs and how many years it will take for the technology to make a profit. This is a very direct calculation.

Jiazi Guangnian: Can we say that VCs who only like to look at cars are now looking at robots?

Hou Xiaodi: Now people think that driverless trucks are robots to some extent. Robots don’t necessarily have legs and arms to run in factories. They can be a hanging box that runs on public roads. Some evaluation criteria for robots can be applied to driverless driving, so driverless trucks can meet the criteria. And driverless companies are much cheaper than robot companies.

Jiazi Light Year: Which of the current costs are higher?

Hou Xiaodi: I can give you an example of Android phones. I have been an Android user for many years. Early Android users used to make fun of themselves: "I am either charging or flashing my phone." But as the mobile phone system matures, the frequency of maintenance events such as charging and flashing is getting lower and lower, and L4 autonomous driving will also go through this process.

Our main task is to make daily maintenance more cost-effective, time-saving and process-automatic. Although Bot.Auto has world-leading technology, we do not claim to be a technology company, but a pragmatic operation company. We use the world's most advanced technology to operate well. We are not a group of high-ranking algorithm scientists. We are just down-to-earth practitioners in the freight logistics industry. The value created by technology must be reflected in operations, not in how beautiful the formulas derived from the technology itself are.

Jiazi Guangnian: Specifically, how do we ensure that maintenance becomes easier?

Hou Xiaodi: For example, can we detect when the camera is dirty? When should we clean the camera? Who should clean it? How should we clean it? How long will it stay clean after cleaning? There are probably thousands of such questions. We want to think about doing this again with brand new automated technology when there is no driver intervention, based on past experience.

The industry is basically facing two types of problems. The first type of problem is how to make a safety redundant system, which is the first half of the game. Tucson was the first and only one in the world to complete it. A few days ago, (American self-driving truck company) Kodiak completed a driver out test in the desert, but its speed was very low, it was not on a public road, and there was not much interaction with other vehicles. The main selling point was hardware redundancy, which I think should also be applauded. But Tucson is still the only one in the world to complete a fully unmanned heavy truck test on an open road.

Image source: TuSimple

After the first half is completed, you will get a system with good performance but also very expensive. The second half is to consider making the expensive system cheaper, which is what we are currently doing.

Jiazi Guangnian: Will you also pay attention to the developments of other unmanned truck companies?

Hou Xiaodi: I know a little bit about the industry dynamics, but I don’t pay attention to it from a competitive perspective. Now is not the time for competition at all. There are three stages in the industry, namely mountain climbing, running, and boxing. We are all in the first stage. When a company makes the operating cost of autonomous driving equal to that of human drivers, the first stage will end. The second stage will end when the additional capacity provided by autonomous driving exceeds the current shortage of truck drivers. In the third stage, it will be direct competition.

For example, if there is a snowy mountain on Earth that no one has climbed before, and I want to be the first person in the world to reach the top, should I rush to climb it now? No, I will exercise, survey the terrain, and then climb it when everything is ready. This has nothing to do with whether I know that others also want to challenge this snowy mountain.

Most Wanted to Recruit True Believers

Jiazi Guangnian: How many people are there at Bot.Auto now?

Hou Xiaodi: More than 40 people.

Jiazi Guangnian: You said before that driverless companies don’t need a lot of people, just like model-level companies, it’s not a very manpower-intensive business.

Hou Xiaodi: In the past few years, we did need a lot of people because there was nothing in the ecosystem. But now it is no longer a wilderness, and most of the "wheels" we need are already popular. Computing units, sensors, etc. are much more mature than in previous years. There are countless open source platforms for parallel training of large models, and there are also various computing resources on the cloud. The sufficiency of the "wheels" will make us rethink the organizational structure.

Five years ago, I thought that a company like Midjourney, where 10 people could create a product with millions of users, would be impossible to emerge. But now there will be more and more such companies, and fewer people will focus on more core products.

Jiazi Guangnian: Then it will be as easy as assembling a computer host. In this case, what kind of people would you most like to recruit?

Hou Xiaodi: What I want most is to recruit true believers.

Jiazi Light Year: How to define a true believer?

Hou Xiaodi: Technology is important, but the championship and runner-up in the Olympic Games are usually not determined by technology. Entrepreneurship is a long-distance race, not a sprint. Many driverless companies died of acute diseases, not chronic diseases. We should ensure that the company is stable and survives for a long time. As for an engineer, whether he solves a specific engineering problem in one week or one and a half weeks is no longer so important.

Second, when capital is abundant, I will have a laissez-faire mentality, and there will be more unexpected drives brought about by individualism. A hobby project can achieve a great career, just like Google always says that Gmail was incubated by their 20% system (Note: Google once stated that every engineer can freely dispose of 20% of their working time).

But now the cost of capital acquisition is very high, and the Federal Reserve will not cut interest rates, so financing is very difficult. We need to transform at the company level, from bottom-up to top-down, from individual momentary intellectual flashes to organized corporate unified strategies. In this case, the freedom and time given to individuals will become less and less, which is actually related to the scarcity of resources.

Jiazi Guangnian: Is this only true for the driverless car industry, or is this a trend that exists in the entire technology field?

Hou Xiaodi: I think it's the entire technology industry. Some other companies I know don't give employees much room to play freely. To be honest, we also need to consider to what extent Google's 20% system was just for publicity.

Jiazi Guangnian: However, the popular model-level companies or star teams now don’t seem to have many people, at most a dozen people, with one or two geniuses as team leaders, and they have produced very good products, such as Sora and Pika.

Hou Xiaodi: In fact, most people in most companies only produce one or two extensions of the paper, especially at the level of pure software. Some people also say that it is all Llama model shells. I don’t agree with such extreme statements, but in this case, the flash of personal ability has indeed not brought about earth-shaking changes.

Jiazi Guangnian: So you hope that the people you recruit can implement the company's top-down ideas.

Hou Xiaodi: Five years ago, we would open some special projects to cater to our star employees, but now we won’t do that. Our core value is whether the operating cost can be lower. You can’t say that I have published a great paper, so I have to do basic research work. I need a free atmosphere. Maybe the company in the past had the same freedom as academic institutions, but it has not been established in recent years.

Jiazi Guangnian: This will make the candidates feel that you are too authoritarian.

Hou Xiaodi: If you want freedom in research, you can choose the academic circle. Or you should look for a large company with funds and a monopoly. We small companies must have the mentality of living on the edge of bankruptcy. Even if we just raise money, we must know that the money will be burned out soon. With this mentality, we really don’t have the freedom to do open research.

I would also like to share a "top-down" corporate governance idea. This has little to do with technology, but is about how a small company can use professionals.

Let's take lawyers as an example. There are two types of lawyers: corporate counsel and litigation lawyers. Corporate counsel will often tell you that this is not compliant and that is risky, but if you do not incorporate these suggestions into the top-down corporate strategy and accept them without hesitation, then in the end, there will be very little that can be done. Litigation lawyers fight lawsuits every day and live on the edge of life, so they will not tell you what cannot be done. They will say that it is risky for you to do this, but the risk is greater if you don't do it, so you still have to do it. A good litigation lawyer does not give you risk warnings, but directly gives you an overall top-down strategy.

If you rely too much on "professionals doing professional things" without your own top-down strategic backbone, you will easily enter a state of not seeking credit but avoiding mistakes. But startups are always sailing against the current, and entering such a state is very dangerous and may lead to death. The essence of a small company's survival is whether it can create miracles in the company's general direction.

Jiazi Guangnian: From serving as Tucson’s CTO in the early days to becoming its CEO, and now founding Bot.Auto, what has been the most difficult transition?

Hou Xiaodi: Now we need to strengthen our study in finance, legal affairs, financing, and public relations. I used to think that it was enough for professionals to do professional things, but now I think I need to formulate a core strategy to bring all levels together. Just like different performers playing different instruments, they all return to the main theme that the conductor is responsible for. The company needs such a conductor who deeply understands the company's vision and mission and makes all parties move in the same direction. This is the transformation that the CEO needs to make.

Now I pay less attention to details. If an algorithm problem takes a month to be solved, I might slam the table and do it myself before, but now this will not happen again. If you can't finish the work, I will not do it myself. But I will spend more time on the big picture and get in sync with everyone on principles and strategy.

Jiazi Guangnian: From the beginning of your business to now, have the goals and principles you have believed in for a long time undergone any other changes?

Hou Xiaodi: Let me start with the things that don’t change. I’ve said before why I want to work on driverless cars. There are three reasons. First, my work must be meaningful, can promote productivity and be challenging; second, it must be technology-driven, and technology must create value; third, the business model must be simple, and I can’t make a business model that I can’t explain clearly. These three points are still the same when I work on Bot.Auto now.

What has changed is that I used to put myself too much in the technical position, thinking that strong technology can cover up a hundred flaws. But now I still have to turn myself into a hexagonal warrior to protect the company in all aspects. In order to avoid dying from acute diseases, we must think carefully about all aspects. At the same time, the company needs a long-term strategy to move in one direction and unite employees with strategy.

In my previous experience, I played the role of CTO most of the time. I felt that the company was not heading in the right direction, so I became the CEO, only to find that the problems of the previous "30 years of Wanli" were already too deep to be reversed. This time, since I can fully perform the duties of CEO from the beginning, I must ensure that the company is always on the right strategic path from the beginning.

*Ms. A also assisted in the interview for this article

For media cooperation, please contact WeChat ID: ciweimeijiejun

If you want to communicate with us, you can reply "join the group" in the background to join the community