news

Talking with Lou Tiancheng about Robotaxi: "The more advanced the assisted driving is, the further away from highly automated driving is the car company"

2024-08-14

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Editor’s Note:Tencent Auto's editorial department calls the past decade of electrification the "stormy era" of China's auto industry. Now, standing at the historical node of 2024, which is known as the "first year of intelligent driving", we can't help but ask: What technical route will the major players in the industry adhere to? How will they build their own competitive barriers? Tencent Auto has specially launched a series of intelligent driving planning. Through interviews, actual tests, cross-comparisons, reviews and other methods, it strives to stand at the origin of history and further understand the huge changes that may occur in the auto industry in the next ten years, so as to provide readers and the industry with a more comprehensive content guide and leave some historical footnotes for the industry.

Author | Zhang Xiaojun

Editor | Shi Ding

Produced by Tencent News "High Beam"

Just when Robotaxi was about to be erased from people’s memory, it suddenly returned to the public eye.

In July this year, Baidu launched the Robotaxi, which attracted widespread attention in Wuhan and also caused some controversy. Many people began to discuss whether Robotaxi has reached a turning point. In this regard, we talked to Pony.ai co-founder and CTO Lou Tiancheng about Robotaxi.

Lou Tiancheng is the first graduate of Tsinghua University's Yao Class. He is also known as "Master Lou" for his outstanding performance in programming competitions. After graduation, he joined Google, where he first came into contact with driverless cars, and then joined Baidu to engage in driverless car research and development. In 2016, Lou Tiancheng and Baidu's former chief architect Peng Jun founded the driverless company Pony.AI, where he served as co-founder and CTO. The company is valued at US$8.5 billion and is seeking to be listed on Nasdaq or the New York Stock Exchange.

Lou Tiancheng has been working in the field of L4 autonomous driving for about 10 years. During the interview, we discussed news about Robotaxi, a brief history of driverless driving, and common routes for autonomous driving.

Lou Tiancheng made some rather sharp comments: for example, "the better L2 is, the further it is from L4, and vice versa", and "when autonomous driving surpasses humans, data becomes a interference factor, and the more is not necessarily better."

Here is a summary of some of Lou Tiancheng’s core views:

① The reason why the carrots are running out of the circle is because they are unmanned;

② Autonomous driving can be divided into 5 levels;

③ Pure vision is a good student, while lidar is a cheating student;

④ The better the assisted driving is, the further it is from L4;

⑤ When autonomous driving surpasses humans, more data is not better, but a distraction;

⑥ Now is not the time for autonomous driving. It will take another three to five years for unmanned driving to be implemented.

⑦ My AI worldview: "Humans are essentially nothing more than AI."

The following is an excerpt from the interview. (For easier reading, the author has optimized the text)

There are five levels of autonomous driving

"High Beam": Let's start with the recent news. Robotaxi became popular in Wuhan some time ago, and Robotaxi has attracted a lot of attention. Why is it at this time?

Loutiancheng:Because it is unmanned. We will achieve unmanned operation by the end of 2022, and Baidu will also achieve this by the end of 2022. The technology has been mature for more than a year. It takes time to prepare from the maturity of technology to the public perception.

It (Robot Run) has a certain scale and is concentrated in one city, so the public perception is higher than before. 99% of it is because it is unmanned. This is the premise - even if there are 100,000 vehicles, each with a safety officer, it will definitely not reach the popularity it has today.

"High Beam": Others would also attribute the popularity to other factors, such as more fleets and larger trial areas.

Loutiancheng:These are the 1%. This 1% has only emerged recently, but it is fundamentally unmanned.

"High Beam": Why is Wuhan the most popular city? Is it accidental or inevitable? Who will be the next Wuhan?

Loutiancheng:There may be a relationship, but it will definitely not be the only one within a certain period of time. All first-tier cities are possible.

"High Beam": Can you sort out the history of the development of autonomous driving, how it was born and developed step by step to where it is today?

Loutiancheng:Autonomous driving technology has been around for 30 years. 2009 was a threshold when a group of people realized that it was possible to achieve full autonomy - before that everyone was talking about auxiliary functions. I was not directly involved, but later joined the team (Waymo).

The definitions of L2 and L4 began at that time. (L2 is partial autonomous driving, and the driver must always monitor the driving environment and be ready to take over control at any time; L4 is highly autonomous driving, and the vehicle can be fully autonomous in most environments and conditions without driver intervention.)

I use time to make you more sensitive - it can run automatically for 1 hour, 10 hours, 100 hours, 1,000 hours, 10,000 hours without your intervention. Each step takes at least a year or even two or three years to complete. This is L4 pure unmanned driving. Assisted driving has another line.

Google achieved 1 hour in 2012, and we achieved 1 hour at the end of 2017. We have been moving forward from 1 to 10,000 hours, and each step has evolved our mindset. Cognition is very important.

"High Beam": What is the key to every technological breakthrough?

Loutiancheng:The vehicle should have some basic functions and some basic time capabilities, and basically take over once an hour. Not many vehicles have achieved this today. You can't easily say that assisted driving can be done in one hour. This involves vehicle modification, including the entire integration of sensors.

1 to 10 are some of the models we mentioned, which are machine learning. They are not necessarily large language models, but various models - they are the main driving force of 1 to 10.

From 10 to 100, it really involves collecting large-scale data, very complex models, and end-to-end. And you must have a certain fleet and collect enough data. From 10 to 100, you need native data and cannot rely on data from other industries. Complete data collection and simulation training system are the key to this stage, and it is not too much to take two years.

The key to going from 100 to 1000 is the indicator system. For example, we achieved 200 last month, and there is a new version this month. I want to know whether we should achieve 180 or 220? I can't rely on guessing. I can't make a decision by sitting for an hour, because 1 hour is too short. Can I test for 10,000 hours? First of all, it takes a long time to test for 10,000 hours; secondly, there is no guarantee for testing for 10,000 hours. For example, it rained more last month, but it rained less this month. How do you know whether the data is strong or weak? For example, there have been many road repairs recently, or it is just bad luck. There are many traffic accidents on Friday nights. Last month, I accidentally had an extra Friday night - there are all kinds of noises that affect judgment.

How to filter out these noises and truly judge whether you are doing well? This is the most critical point of 100 to 1000. Many companies are stuck here.

From 1 to 10 or 10 to 100, we often talk about a word called "resources" - I have a lot of money, data, machines, and I can train models, which is no problem. But if you don't have a "ruler", you don't know whether it will get better or worse. This is the ability to "surpass resources". If you can't do it well, no matter how many resources you have, you can't do it. It won't go in the opposite direction, but it will go in circles.

Let's talk about the story of AlphaGO. How do I know that humans can't beat AlphaGO? They can play against each other, and I want to see who has a higher chance of winning. The quality of driving varies from person to person, and two good drivers drive differently. Is one good or bad driver? It is extremely difficult to evaluate.

I even trained a large-scale model to do this - evaluating the quality of this model is much more difficult than the model itself.

"High Beam": What is your professional term for this model?

Loutiancheng:We call it Context-based metric system internally (aims to evaluate the capabilities of the system according to different scenarios), and I focus most of my energy on this.

This is a very important thing for the company to break through the limit. Many people talk about how much data we have and how big our data is, but forget the difference between good and bad - when you can't judge good and bad, you can't move forward.

"High Beam": Recently, Carrot Speed ​​​​has been running fast. My friend in Wuhan said that it drives too slowly, which makes a human driver feel unsafe - as if there is a truck in front of you, and you need to drive slowly as well, otherwise what if you get hit?

Loutiancheng:This evaluation is difficult to quantify, and it needs to be simulated with a model. Safety has a sense of security issue - if you drive to scare people to death, it is also unsafe. Driving too slowly is unsafe, this is a high-level understanding, and this understanding is very powerful. Sometimes there is an illusion that driving slowly will be safer, but it is not the case. Driving slowly will make people dizzy, and make other drivers wonder what you are doing, making others confused (confusing) and very dangerous.

From 1000 to 10,000, this is actually it. Now I look back and it seems quite obvious. I didn’t understand this at the time of 1000 hours. From 1000 to 10,000 hours, the safety risks you may have are not entirely caused by you, but by other drivers. If you don’t hit others, others will hit you. Overall safety is a very important part of real safety, and this is after 1000 hours.

Most people can do 1000, which is beyond the human level.

Why is it so difficult? I was riding in our car, and one day I suddenly realized, why does this car drive like this? At this time, I dare not doubt that it is wrong, because I doubt that I am not as good as it, it may have thought of something I have not thought of - it drives better than me, I can't just rely on myself to evaluate, because I am not as good as it, I can only tell it what is good.

From 1000 to 10000, there is such an evolutionary system that can help it evolve - you can't teach it. I think it's okay to take care of the child now, and I can still teach him; many years later, it may not be the case. He is better than me. I can only tell him how to let him learn by himself, create an environment for him, find him a better teacher, and tell him what is good, but how to learn is his ability.

You must not feed it all your driving data, because your data is only at this level. If you feed it poor data, it will not be able to learn. First, you have to find very good data.

And, how to reduce the risk of other cars is the most critical factor from 1,000 to 10,000. It is a bit like what you just said, other people's cars have difficulties because I drive to cause them difficulties and cause them to have accidents. So can I not cause difficulties to others and make it easier for them to drive?

"High Beam": Are we now in the process of going from thousands to tens of thousands?

Loutiancheng:Now it has reached 10,000 hours, which is 10 times the level of human drivers.

"High Beam": What are the professional terms for the 1 hour, 100 hours, 1,000 hours to 10,000 hours you mentioned?

Loutiancheng:We usually use the number of miles you drive before something you don't want to happen happens, such as a takeover or an accident. It originated in the United States, where miles are used, such as Miles per Intervention, Miles per Perfectly evasion, etc. We call it MP-X.

"High Beam": From starting your business in 2016 to now, which step was a qualitative leap for you?

Loutiancheng:Every step was difficult. We stopped at 10 hours for a long time, worrying that we couldn’t go on. 10 hours was definitely not enough. If there was no progress for a long time, the team might be confused and the pressure was very high.

This breakthrough is a very important point in the accumulation of confidence, and it was accompanied byToyotainvest.

The more powerful L2 is, the further it is from L4

High Beam: When I was studying autonomous driving, I found that there were many disputes over the route in this industry. The first one was whether to use pure vision or lidar?

Loutiancheng:The biggest difference is the different goals. Different goals lead to different most suitable routes.

If I want to do fully unmanned, I won’t take certain routes. For example, I won’t take the fully visual route.

I really like what Hesai said: a good student can be very hardworking and excellent, but it is difficult for you to be better than a cheating student. LiDAR is a cheating thing - for example, the distance from us to the screen can be known by LiDAR once it is measured. No data, no parameters, no need to learn anything, just look at the answer directly. To achieve full unmanned operation, you need "a cheating student" to do it.

High Beam: WhyTeslaWant to go the purely visual route?

Loutiancheng:Because now it doesn't need to be completely unmanned.

The pure visual cost is lower; more importantly, the appearance of the vehicle will be greatly affected. Cost was a factor four or five years ago, but today the cost difference is not so obvious, but it does have an impact on the appearance of the vehicle.

"High Beam": The second route of dispute is, vehicle-road collaboration or single-vehicle intelligence?

Loutiancheng:The two complement each other, and their conflicts have been very rare since day one. Each of them has its own important things, and both can be done, so there is no need for A to replace B.

But if there is no vehicle-road collaboration, the scheduling is not good enough, but it is not completely impossible to go on the road - this is why we first develop single-vehicle intelligence. It has priority.

The data provided by vehicle-road collaboration is currently difficult to help single-vehicle intelligence. Conversely, single-vehicle intelligence is helping vehicle-road collaboration. You can think of autonomous driving as a device, and now this smart device has better performance than vehicle-road collaboration in all aspects.

High Beam: The third route is the debate between incremental autonomous driving and leapfrog autonomous driving? You are leapfrogging.

Loutiancheng:You mentioned L2 and L4. From the numbers, I seem to have taken advantage because I took a larger number. I shouldn't have taken advantage of you. We have two different goals - that's it.

One is football and the other is basketball. It is more difficult for a top football player to become a top basketball player because many things are done too extreme. Football players are very good and have strong bending ability. People have a bit of hunchback. Basketball players must not be like this. When he is extreme, it is difficult to do another thing.

High Beam:XiaopengI recently went to the United States to test drive Tesla FSD and Waymo, and concluded that Tesla FSD will definitely catch up with Waymo next year. Although Waymo's experience is smoother at this stage, Tesla has more data, a wider range, a larger area, and lower costs. Do you agree with this view?

Loutiancheng:One studies liberal arts and the other studies science. Who will score higher depends on the questions.

Their goals are very different. I compare assisted driving and L4 because FSD and L4 are not comparable. For assisted driving, cost and coverage area are very important, and interaction with humans is very important. As an assistant, what it has to do is to do everything right when it helps you, and to tell you that you have to do it yourself when it can't help. What I just said is all that L4 doesn't have.

First of all, the cost of L4 is important, but not that important. On this basis, you have to handle everything yourself and cannot leave it to others. Others may be sleeping and you cannot wake them up. You have to handle everything yourself. Even if it is a little slower (we are faster than other cars now), safety must be guaranteed 100%. If you understand model training, these are two very extreme plans.

You can think of it as the difference between a family doctor and a specialist. A doctor needs to know about various diseases, but you need to find a specialist to treat them. A specialist is someone who knows a lot about diseases, but the ones he knows about must be the right ones.

"High Beam": Therefore, no matter how good FSD is, it cannot directly transition to L4.

Loutiancheng:The more L2 does, the further it is from L4.

The reverse is also true: the better an L4 company is, the further it is from L2.

"High Beam": What are the essential differences between L2 and L4 technologies?

Loutiancheng:L4 requires at least 10,000 hours without an accident. To put it bluntly, if an accident occurs every 10 hours, who would dare to use your car?

L2 technology is also constantly improving, but the number is getting lower and lower. Three months ago, I might have achieved 10 hours, but this month it is only 8 hours. This is not a problem, because users have never asked me to do so, as long as it is greater than 1. But from 10 to 8, the cost is reduced, and users only need to pay less, which is what users prefer. You originally had to go from 10 hours to 10,000 hours, but it took you three months to reduce it to 8 hours.

You are a better L2 product, but further away from L4.

Musk once shot a video in which he took over once every 40 minutes. He didn’t think there was anything wrong with it, and neither did I.

When autonomous driving surpasses humans, data is a distraction

"High Beam": What is the North Star indicator you are most concerned about right now?

Loutiancheng:In order of importance, they are safety, comfort and efficiency.

"High Beam": Regarding what we just discussed about how to reduce the fear of human drivers or passers-by, what do Robotaxi companies need to do?

Loutiancheng:Do what everyone should do, and let him think you are a normal car. This technology is very complicated, but in layman's terms, it is to make everyone think that it is a person driving the car, and there is nothing special about you.

Don't surprise others - change lanes resolutely when necessary, and don't hesitate when not to change lanes. People generally don't hesitate. If they feel that the lane is a little crowded, they will squeeze in - all actions are like the normal behavior of an experienced driver.

Many people have different ways of doing things, which we call "one thousand faces for one thousand people". Everyone thinks that more data is definitely better, but this is not the case. Because once you have one thousand faces for one thousand people, the more data you have, the worse it is and the more complicated it is.

"High Beam": This is contrary to FSD's point of view. They think the more data, the better.

Loutiancheng:If they could achieve 1,000 hours, they would understand it this way, but now it is less than 1.

When you start to surpass humans, the more data, the better. Humans have 1,000 hours. When you start to surpass humans, data becomes a distraction.

After 1,000 hours, you have to consider the shortcomings of the data. People drive differently in different situations, which will confuse machine learning. It doesn't know who to learn from. It also can't take the average, because there is no average for behavior. Suppose you have all the data, and now you want to surpass this average, the more data you have, the harder it is to surpass it.

"High Beam": If more data is not always better, how much data is reasonable?

Loutiancheng:This involves some internal points. After 1,000 hours, 100 cars are enough.

"High Beam": This sounds easy. It is difficult to increase the amount of data, but it is easy to reduce the amount of data.

Loutiancheng:Getting smaller doesn’t mean I just throw away some data, but I need to do more processing on the data and leave what I want. It’s difficult.

"High Beam": How do you know how much data is reasonable?

Loutiancheng:This is the amount of data needed to infer the size of the model you are training to achieve your goal. The more data you have, the harder it is to converge.

Musk mentioned before that data on a flat road is useless. Because of the iteration of the large language model, its gradient will quickly drop to 0 and cannot converge. This is a technical reason. But the essence is that when you go beyond, you will find that these data will have a negative effect. When you go from 1,000 to 10,000, the data reaches a certain number, it can help you do something, but more is negative and will pull you back.

I have reached 6,000 hours today, but not yet 10,000. After I add that 1,000 data, it will be pulled back from 6,000 to 4,000 hours.

“High Beam”: Since there is so much data for L2, why has it only reached level 1 now? Why is it so slow?

Loutiancheng:I often give this example. Our company gives each employee one computer, but everyone's performance varies greatly. Some people in the company do well and some do not. It's not that you give 10 computers to good people and only one computer to bad people. Everyone has one computer, but there are big differences. There are many things beyond hardware.

"High Beam": He Xiaopeng has a point of view. He believes that in large-scale AI applications, software and hardware must be combined. Do you agree?

Loutiancheng:If we are ultimately pursuing the ultimate in performance, then hardware and software integration is helpful, and I agree with that. But not everything in the world needs to pursue the ultimate in performance.

Now is not the time for ChatGPT on autopilot

"High Beam": You joined Google right after graduation. What is your biggest gain at Google?

Loutiancheng:The biggest gain was exposure to autonomous driving, because at that time only Google had autonomous driving.

"High Beam": Does Google's organizational culture give you any inspiration?

Loutiancheng:Google emphasizes teamwork, and teaches you many ways to collaborate better, such as how to manage code and data, and how to write code together so that everyone can write well and without problems.

It takes everyone working together to get things done, not just one or two superstars.

"High Beam": Why did you start a business with Peng Jun in 2016?

Loutiancheng:I was at Baidu at the time, and everyone had great recognition for our concept, cognition, and determination to do this.

Now, Turnip Run is more focused on unmanned vehicles. From 2016 to 2019 and 2020, everyone believed that it was possible to achieve full unmanned operation. But in the following three years, during the pandemic, everyone was a little skeptical about whether it could be done. It was basically everyone's perception of the matter. Everyone was willing to come out because they wanted to spend a few years doing something that would impact humanity.

"High Beam": You have been in business for 8 years. Has the implementation of driverless cars been earlier or later than you expected?

Loutiancheng:Some are early, some are late. First of all, it is the policy. In the past few years, especially in China, the policy has been to move forward with the technology. Thanks to the policy makers. Another fast thing is the cost. In the early years, the laser radar was 80,000 US dollars. The exchange rate was about the same as it is now, 500,000+ RMB, while today the good ones are only a few thousand RMB. The price has dropped much faster than expected, which is very helpful for commercialization.

The unmanned operation is a little slower than I expected. Of course, it’s not all due to the epidemic, but it is related.

High Beam: Tesla's FSD is very impressive, and the intelligent driving capabilities of domestic automakers such as Huawei and Xiaopeng are also improving. What is the relationship between them and you? Finally, can you describe the market landscape to me?

Loutiancheng:L4's partners include technology companies, car manufacturers, and operating platforms. It is possible that one company can do more than one of these, but from the perspective of the structure, this is the case. Doing more does not mean that the structure is not the same. This is the status of L4. They can also become car manufacturers. If a company suddenly emerges one day, it can make both cars and good technology. This is possible.

"High Beam": If L4 is realized, will there be a need for so many Teslas on the streets?

Loutiancheng:That's not the case. Many people still have to drive. Today, we are more focused on providing travel services and providing basic transportation capacity for cities, not depriving people of the right to drive.

"High Beam": Some people say this is the ChatGPT moment of autonomous driving. Do you agree?

Loutiancheng:L4 has definitely not arrived yet, but we want to achieve it.

"High Beam": When do you think we can see Robotaxi operating on a very large scale in multiple cities?

Loutiancheng:Three to five years. The difference between L2 and L4 is the cost. If you talk about large-scale, the cost must be around 100,000 yuan. Now it is about 100,000 yuan.

But the decline does not mean buying cheaper things, it requires standardized production and production lines. Last year, we established a joint venture with Toyota to specialize in mass production. It will take longer to standardize our solutions, mass produce them, and put them into production lines.

"High Beam": How many autonomous driving companies aiming at L4 can survive in the world?

Loutiancheng:My point is a hand.

The technical threshold to reach 10,000 hours is very high, and it is not something that can be achieved casually; but it is definitely not 1. Its expansion is different from the Internet, which has zero cost expansion, but it requires the deployment of vehicles. Even the car manufacturers are not the only ones. In the end, there are different experiences and different requirements.

"High Beam": After raising so much money and spending so many years, are you satisfied with the results today?

Loutiancheng:Satisfied. I have been satisfied since the end of the year before last - unmanned operation is the ultimate dream of generations of technical people.

"High Beam": What do you expect the success of Pony.AI to be like, and what is your worst-case scenario?

Loutiancheng:Three years and five years (to achieve full unmanned operation). I estimate that the best is three years and the worst is five years.

"High Beam": Is there a possibility that Pony.AI will fall on the road to autonomous driving?

Loutiancheng:Won't fall for any reason I can think of.

"High Beam": What has kept you working on driverless cars for so many years?

Loutiancheng:Not only me, but all the founding team of the company. This is also what my mentor Mr. Yao Qizhi brought me. I was a student in the first class of Yao class.

When a person gives up his original intention, most people will think that he has encountered many difficulties and cannot persist. However, the fundamental reason why people give up their original intention is unrealistic temptation, not hardship. Many people can persist in moving forward under very difficult conditions, but cannot maintain a good attitude in the face of temptation.

We came out not for a decent life, we actually had everything, so we didn't need to do that. We just wanted to get this done, so I wasn't easily tempted. Those who came out at the same time as us and didn't stick with it didn't do it because of hardship or because they had no money, but because they couldn't resist the temptation.

"High Beam": What kind of person is Mr. Yao?

Loutiancheng:I am not qualified to judge him.

He is a person who takes teaching very seriously. He is very serious about our feedback during class. Even when we ask him questions that he can't answer, he will say that he will find a way to answer it for us - he will really come to you to answer your questions.

This kind of "everyone" may forget, but he will not, he attaches great importance to teaching and students' learning.

Human beings are nothing more than AI.

"High Beam": What do you think of the spatial intelligence proposed by Fei-Fei Li?

Loutiancheng:This involves the issue of world view, and I will not comment on it.

"High Beam": Please tell us about your worldview on AI.

Loutiancheng:Sometimes people overestimate their intelligence. I think we live in an unreal environment. I don't believe that our world is unreal, so I spent some effort to find some evidence that our world is real, but the more I searched, the more I found that our world is not real. We are in a simulator, and the simulator arranged our meeting today.

This basically means that you can explore some parameters of the simulator. For example, the speed of light is a very important parameter of the simulator. A lot of AI work currently being done has begun to explore some parameters to a certain extent, but we are not necessarily fundamentally different from those things.

High Beam: What insights does this bring to your understanding of AI?

Loutiancheng:AI surpassing humans is not the pinnacle of AI.

To put it more exaggeratedly, we often ask whether AI will surpass humans. But when you ask this question, are you overestimating yourself? How can you judge whether AI can surpass humans?

"High Beam": Everyone thinks that AI is created by humans.

Loutiancheng:The cars we created can run much faster than we can run. The AI ​​we created can surpass humans. But the more likely view is that human intelligence may not be enough to judge whether AI can surpass humans.

I am not exaggerating the capabilities of AI. AI itself is a fitting and regression process. AI does not create new logic. Today's so-called AI is still far from the level of general artificial intelligence. Today's ChatGPT and today's large language models are all imitating people, nothing more - it's just a process of fitting a large number of parameters.

But I don’t deny that this approach is very good and can achieve very good results, and can even surpass humans - because humans are essentially just AI.

"High Beam": Will humans and robots share the same brain?

Loutiancheng:This is to answer the question of what is the brain. Maybe the brain is just a program of higher-level intelligence. What is memory? Memory is an encoder of some previous impressions. This can be done, but it is not easy for humans to do it.

Today, AI technology has made a lot of progress in some aspects, such as ChatGPT, but it has not reached a particularly high level in many aspects, such as energy consumption. I didn't feel hungry until I had lunch at noon, and I did so many things, but if a robot did the same thing, it would have been exhausted long ago. The energy consumption ratio of humans is very terrible.

The stability of this system is also extremely terrible. If you ask me to build a heart today, I can’t make it. It can beat so many times in its life without breaking down. I can’t make such a thing. So there is still a big gap between what people can build and what they can do, including the brain. But only when dealing with specific problems, AI can reach the level of humans.

"High Beam": So they share a brain?

Loutiancheng:Absolutely, brain-computer technology can do this, and there is no problem in theory.

Because you asked whether it is possible, but I am not saying it is easy to do, it is actually very difficult to do, but it can be done.

There is a trick in the competition: Hold cold card

"High Beam": What do you think of Yang Zhilin? You are both regarded as the academic genius of Tsinghua University and the genius of technology. I am curious about how one genius thinks of another genius?

Loutiancheng:Although we are both in the field of AI, there is no immediate point of cooperation. If we can cooperate, I will still cherish this opportunity. Only those who have experienced the same things will understand the difficulty of cooperation, the support of the other party and the real shining points.

Only those who have won the championship know the difference between the champion and the runner-up.

High Beam: Describe the difference.

Loutiancheng:If you look at the Olympic medals, the difference between the first and second place is more than just one place. Many people don’t know what the second highest peak in the world is, but everyone in the world knows the highest peak in the world.

However, the effort he put in is also huge. When you are not close to it, you may not know how much effort you have to put in. But if you have really walked through it, I believe he will understand.

"High Beam": Is it a burden to be called a technical genius?

Loutiancheng:If this can be considered a burden, it may not reach this height. Because it can be called a burden before. We should set our sights higher. Obviously, there is still a wider space.

"High Beam": Why is your ID for the competition called ACRush? How did you choose it?

Loutiancheng:It was an era where speed was king, and Rush was the fastest running move.

"High Beam": You have won the TopCoder championship for 11 consecutive years. Why did you participate in 11 sessions? What is your driving force?

Loutiancheng:It was because I didn’t win the championship in the 12th competition. Du Yuhao later surpassed me. Although I surpassed him three years later, the latter part was not important. In essence, I didn’t make it to the 12th competition.

High Beam: When was the last year you participated?

Loutiancheng:Just a few weeks ago. I won the runner-up last year, but I didn't win the championship. I usually call the runner-up the "biggest loser". First of all, he is a loser, but he is also the runner-up, so he is the biggest loser.

Racing is something I can stick with for 20 years, and the same goes for autonomous driving.

"High Beam": Is there ever a year when you felt that you might as well not go this year or not go in the future?

Loutiancheng:To use a jargon from the competition circle, being able to do something and wanting to do something are two different things. If you can't do it, saying you don't want to do it is just making excuses.

It is a selection, not an application. You have to be in the top few to go. If you have not been in the top few in the past few years, then you cannot go. But not being able to do it and not wanting to do it are two different things. Being an L4 also involves whether you can do it and whether you want to do it.

“High Beam”: Are there any skills to the competition?

Loutiancheng:Skills are everywhere. In competitions, others will know your current status because everyone is in a racing process. Instead of knowing the result at the end, you will know it roughly during the process.

There is a "hold cold card" in this process. I have some things I know how to do, but I just don't let you know. I use some methods to make you feel that I don't know either, so that you give up the idea of ​​really fighting hard and drown you in it.

Now he can do 80 points, I can do 90 points, if I let him know that I have done 90 points, it will stimulate his potential, he will work hard to do a higher level, and will think about things beyond this point. You control your score at around 80, let him feel that he is good, just pay attention to some details. Finally, you will show your sword and show 90 points, then there will be no time to think about this matter.

Often at the last minute of a competition there will be completely different results. Everyone holds on to that point, but there is always someone who doesn't hold on. Many people adjust the score to the highest point of the person who holds on the least, making others feel that I am almost there.

"High Beam": When racing against the L2 technology line today, will you hold on?

Loutiancheng:This is different. The development of a company requires a certain degree of publicity. Competitions do not focus on the intermediate situation. Even if you are ahead all the way, you will still be overtaken in the end. This is the characteristic of competitions. Companies are not like this. I don’t need my classmates to need it, and they don’t need their families to need it. I need to make friends and have partners. The intermediate results are very important.

"High Beam": If you use this trick more often, won't others know it?

Loutiancheng:I do this all the time, and everyone knows it. But maybe he was lucky.

"High Beam": I remember you had a time when you were counting down in a competition.

Loutiancheng:I may be the world's top player in the last place in competitions, with four last places. It's hard to find someone who can surpass me because many people can't make it to the finals four times. There are usually only 8-10 people in the finals, so it's not easy to make it to the finals four times. I was lucky enough to be at the bottom four times.

"High Beam": If you count down, will you feel depressed when you go home?

Loutiancheng:Of course I will be sad because I take this matter very seriously, but it’s over and I need to move on.

"High Beam": How long have you been lost?

Loutiancheng:A flight.

"High Beam": You have participated in so many competitions. What kind of fun does competition bring to you?

Loutiancheng:Does understanding the difference between the champion and the runner-up count? Knowing many of the same people and having an attitude of continuous learning.

The most fundamental thing is that when you reach the top of the tower, that kind of extreme is a feeling I have always missed.

Previous content