news

tencent news "high beam"

author: aodun

editor｜liu peng

xpeng motorsthe okr of li liyun, head of autonomous driving, is almost an open secret in the industry.

"every word xiaopeng says will be my okr." at last year's 10.24 technology day, li liyun, who had taken over xiaopeng motors' autonomous driving team for only two months, made his first appearance in his new capacity. that night, he issued a military order: in 2023, xngp will be opened in 50 cities; in 2024, it will be usable and easy to use throughout the country.

nine months later, one year after taking over the autonomous driving team, li liyun led the team to complete the okr ahead of schedule and exceeded it. on july 30, he xiaopeng announced at the ai ​​intelligent driving conference that xngp will be officially upgraded from "available nationwide" to "easy to use nationwide", achieving nationwide opening "without restrictions on cities, routes, or road conditions".

2014, the year when xiaopeng motors was founded, was also the year when li liyun officially joined the autonomous driving industry. he was one of the first people to engage in autonomous driving at baidu's us r&d center, responsible for the overall architecture and algorithms of the driverless car behavior prediction system and intelligent decision-making planning system; after working briefly in the x lab of jd silicon valley r&d center for a year in 2018, he officially joined xiaopeng motors in june 2019, initially serving as the person in charge of the intelligent driving decision-making planning algorithm and simulator, mainly responsible for the research and development of xiaopeng motors' high-speed and urban autonomous navigation assisted driving system ngp; in august last year, wu xinzhou, former vice president of xiaopeng motors' autonomous driving, moved to nvidia, and li liyun officially took over as the person in charge of the xngp project.

unlike any previous stage, the year that li liyun led the xiaopeng intelligent driving team was a year in which big models vigorously stirred up the intelligent driving track. everyone and every company involved seemed to be undergoing a transformation. just last month, xiaopeng motors' autonomous driving center completed its latest organizational adjustment, focusing on ai and end-to-end, and newly established three major functional areas: ai model development, ai application development, and ai performance development.

before the ai ​​smart driving conference on july 30 and after the 10th anniversary conference, tencent news "high beam" had an exclusive interview with li liyun. looking back on this year, he believes that the most important thing is people, activating and coordinating organizations and people. "with such a complete organization, everyone can give full play to their strengths. with everyone working together, we have gradually achieved our goals."

when talking about how he xiaopeng intuitively understands the progress of internal autonomous driving, li liyun gave an example to illustrate: "in fact, it is very simple. it is to push relatively stable and the latest software to the cars of the senior brother (he xiaopeng) or other internal colleagues. through test drives, he will see the progress and evolution of xiaopeng motors in intelligent driving capabilities."

during the interview, li liyun said "sure" nearly 20 times. he believes that in the end, there should be many car companies that will have intelligent driving capabilities, but most of them may only have 60 or 70 points. the car companies that can really achieve 80 or 90 points must be very sure, self-developed, have strong mid- and back-end capabilities, and are truly oriented towards ai reform. "we believe that there may be only a few in the world, and xiaopeng will be one of them."

the following is an edited transcript of a conversation between tencent news' "high beam" and li liyun:

“looking at the whole year, the most important thing is still people”

tencent news "high beam": he xiaopeng has always been all in on intelligent driving. do you have close communication on a daily basis? how does he evaluate your work?

li liyun:our intelligent driving department is a happy yet very challenging department. the interesting thing is that we basically do not write ppt reports.

how can the eldest brother intuitively understand the progress of internal autonomous driving? in fact, it is very simple. we just push the relatively stable and latest software to the eldest brother, or to the cars of me, tingting (yuan tingting, senior director of autonomous driving products of xiaopeng motors) and other internal colleagues. we will test drive the car ourselves first. through the test drive, he will see the progress and evolution of our intelligent driving capabilities.

the communication between me and my senior brother is more agile than you think. our software does not have a new version every week or two. sometimes there will be an update or some problem will be solved or some small features will be added every few days. during the process, all stable software versions will be experienced by my senior brother.

as we push the latest software, he will drive it more and more, so we will discuss the driving experience on a daily basis. sometimes we will discover some new needs or areas for improvement. the senior brother will also find problems and feedback them to us. in such natural communication, we will also often discuss the long-term investment and planning of qizhijia and ai technology.

tencent news "high beam": does this high-frequency communication bring you a lot of pressure?

li liyun:xiaopeng and everyone who cares about us can experience the changes or evolution of xiaopeng motors' intelligent driving at the first time. i feel very lucky and interesting. this is related to my personal experience.

i studied for a phd in computer science in the united states. at that time, artificial intelligence and deep learning were just emerging. when i was about to graduate with my phd, neural network deep learning had already completely overturned traditional deep learning, but artificial intelligence was not as popular as it is now.

i started my career in the internet industry (baidu), but i was also fortunate to be involved in the earliest wave of smart driving or autonomous driving in china. i came into contact with robotaxi in 2015 or 2016. at that time, i thought robotaxi was very interesting, but after a period of accumulation and growth, what is even more exciting is that your work or daily efforts, learning and growth can all be presented in front of you, visible and tangible, and you can even buy it yourself, or your relatives, friends, and colleagues can buy it and use smart driving products on a daily basis. this is a particularly exciting thing.

i think robotaxi is more of an operational service, but it is very exciting to own a car. the software and functions of this car are developed and iterated bit by bit by ourselves from scratch. i think this is very interesting. so do you think it is stressful? i think it is a motivation. what is more interesting is that the further explosion of artificial intelligence has gradually made this possible.

tencent news "high beam": you have taken over the intelligent driving team for more than a year. can you briefly summarize your feelings during this period?

li liyun:this year is worth reviewing. looking back on the whole year, the most important thing is still people. because to do anything well, you need the best team and people. of course, we are also very lucky. because of xiaopeng's firm investment, we already have many outstanding classmates settled down. because we have achieved some achievements and industry recognition before, we can also attract tingting and more new and outstanding talents in ai.

what i do is actually very simple, but the most important thing for me is to combine the old and the new, to activate and adjust the organization and people. with such a complete organization, everyone can play their own strengths. with everyone working together, we have gradually achieved our goals. for example, last year we set a goal of 50 cities, but we have actually exceeded the task. on january 1 this year, we launched the app in 243 cities. now we have achieved that the app can be used nationwide.

in the second half of this year, you can pay attention to the subsequent 1024 press conference. we are more of a start-to-end, end-to-end approach. ai is definitely indispensable in the end, or i think it is the only way to embrace ai. end-to-end is actually a presentation of ai. just like xiaopeng said, although our car is an l2 assisted driving car, we will achieve an l3-like experience by the end of next year.

what i want to emphasize is that we will let users feel the speed of iteration. there will be a small version every month and a relatively large version every two months, ensuring that the capabilities are constantly improved in the process of starting from the end.

tencent news "high beam": how to combine the new and the old?

li liyun:on the one hand, it will stimulate the abilities of old classmates and let them perform new roles. for example, the old classmates may have abilities in the middle and backstage, and they are heroes behind the scenes, but now sometimes they are asked to go to the front line to do front-line business. in this process, he has a new role and vitality, and he has grown. more importantly, in this process, he has also led the people around him and set a good example. we call this cross-functional cadre training. now many dedicated people behind the scenes have come to the front to do front-line business. i think the entire center is more dynamic.

more importantly, as we continue to invest and persist in this matter, our business is still making rapid progress, so we can attract a lot of new ai talents. i think the competition for talent is very fierce now, but i am very lucky to see that xiaopeng can continue to attract good people around the world.

tencent news "high beam": last time, xiaopeng joked that all friendly competitors should be more gracious and tap less and dig less.

li liyun:we just hope that our competitors will be more cautious and refrain from poaching our employees. but more importantly, we ourselves have strong appeal, both to new talent and to many excellent old classmates who have returned. this organization is very vital and is growing from generation to generation.

although there will definitely be personnel turnover, i feel very gratified to see that people in the background will enthusiastically step to the front to take on greater responsibilities and meet greater challenges. the entire organization will be more visionary and motivated to truly do a good job in intelligent driving.

tencent news "high beam": i recently saw that xiaopeng's smart driving team has made some structural adjustments. what is the thinking behind this?

li liyun:it is a more agile organizational structure for the ai ​​era. recently, our autonomous driving center has set up three new functional areas: ai model development, ai application development, and ai performance development, to accelerate the evolution of ai capabilities and the ai ​​transformation of the organization.

tencent news "high beam": what are the considerations in attracting talent in the future?

li liyun:i think there are some changes in the quality and requirements of talents. domain is no longer as important as before.

we don’t want young, smart people with ai backgrounds or mathematical rules to join us, and we want people who can use ai to solve business problems. in the future, everyone will compete on data and ai efficiency, so the efficiency and capabilities of the entire infrastructure are also very important. therefore, we think that people’s specialties and preferences may change to a certain extent.

after the organization, way of doing things, business and the entire vehicle-side ai penetration are fully connected, the subsequent evolution will become faster and faster. the core of speed lies in your acquisition of effective data and then the change in efficiency of iterating with effective data.

intelligent driving capabilities and experience must always improve, not "shock therapy"

tencent news "high beam": in the past year, everyone has been rolling end-to-end. looking at the timeline, did you predict this trend internally before this? what preparations did you make?

li liyun:xpeng motors has always been very determined to invest in intelligent driving. since its inception, it has invested heavily in intelligence. at that time, there was an autonomous driving center, which has continued to grow to today. the current investment is also very determined. we invest 3.5 billion yuan in ai every year, and our talents are also spread all over the world.

in terms of end-to-end, we don't do it because others have done it or because we have seen something. for xiaopeng, we do it because we believe in it, so we see it, and so we do it. i think xiaopeng motors is leading this.

as early as the beginning of 2022, we actually started to make relevant arrangements. at that time, we had just completed the verification of the urban assisted driving function based on high-precision maps in several cities. we believe that in order to truly achieve the goal of being able to drive and use the car nationwide, we must embrace the capabilities of ai.

so from then on, we used ai end-to-end cognition to recognize the road network, get rid of our dependence on high-precision maps, and then gradually formed our own unique xiaopeng motors end-to-end and xiaopeng motors' three networks (neural network xnet + regulatory control model xplanner + large language model xbrain). today, it has the ability to be driven and used throughout the country, and has achieved the country's first end-to-end large model mass production.

tencent news "high beam": from the perspective of external perception, end-to-end was originallyteslait was proposed by musk in a live broadcast to demonstrate the first end-to-end solution, and after that, everyone started to follow up in a very short time.

li liyun:i don't deny this perception, but this is just a timeline. for xiaopeng motors, embracing ai is a very certain thing.

xiaopeng motors is an oem with intelligence as its label and core. we must ensure that our products are always progressive in both capabilities and experience. even during the end-to-end promotion process, we will not resort to "shock therapy."

xiaopeng motors has two characteristics: on the one hand, we are very confident in ai intelligent driving; on the other hand, we must ensure that the user experience is continuously improved during the product launch process.

in 2021 and 2022, we launched the predecessors of xngp, the expressway ngp and the city ngp, which actually relied on high-precision maps. at that time, we would also emphasize how many scenarios we processed and how many lines of code we wrote, but at that time we actually already knew that in order to make it accessible and easy to use across the country, we must be confident in ai. so at that time we had already started to conduct preliminary research, make judgments, and do this, but at that time, in order to ensure that the intelligent driving experience of the products we launched was gradually improved, we still did it based on high-precision maps and heavy rules.

so i don’t deny this, but at least among chinese car companies, we are the first to invest in and deploy end-to-end ai.

tencent news "high beam": some time ago you said that "fsd will find it difficult to handle the 300 million electric donkeys in china", and this sentence quickly went viral. from a technical perspective, what is your judgment on the implementation of tesla fsd in china?

li liyun:we particularly respect tesla, which is one of the global leaders in smart cars. it has done very well in the united states because the local road conditions and road rhythm are not so complicated, and the driving speed there is indeed faster.

but i think that the road conditions in the united states are very different from those in china. i have lived abroad for many years, and the intensity and competition are very different from the road conditions in china. in foreign countries, sometimes there are some narrow bicycle lanes, and there may be some scattered cyclists next to them, but it is unlikely that there are all kinds of small electric vehicles like in guangzhou. the infrastructure construction in many chinese cities is very different. some cities have special electric lanes and non-motorized vehicle lanes, while some cities do not. non-motorized vehicles often compete with motor vehicles for lanes. they will not only be on your right, but also on your left. sometimes they will suddenly cross, and some will come from the opposite direction. so i think the traffic conditions are very different.

so i think that if tesla fsd comes to china, there will be different challenges. i believe that no matter where it comes from, it needs to be optimized for china. this is an inevitable process.

tencent news "high beam": in june, when xiaopeng went to the united states to experience tesla fsd, what was the first thing he talked to you about?

li liyun:xiaopeng called me immediately and said that he felt tesla was making rapid progress because of its investment in ai and end-to-end capabilities.

we talked in detail about the differences between the road conditions there and china. for example, the highways in the united states are faster, but the traffic game is not as intense as in china. for example, there are not 300 million electric donkeys, and the traffic lights are simpler. there are all kinds of traffic lights in china, and there are very few waiting areas and roundabouts in the united states, but the road conditions in china are more complicated. sometimes the car has to go to the far left to turn right, or to the far right to turn left, etc.

but i think what is more important is that after talking with xiaopeng, on the one hand, i know where we can do better, and on the other hand, it also strengthens our confidence in choosing this path. insisting on changing the r&d paradigm from the original more people- and rule-driven to data- and scenario-driven will greatly improve our efficiency, so we are more determined about this matter.

xiaopeng experienced our latest version after returning to china, and he also felt that we had done a lot of good things in the latest version. for example, after embracing the big model and end-to-end, we performed very well in scenarios such as small roads, detours, and u-turns. especially for u-turns, if you use rules to solve it, it is difficult to see the u-turn opening, and there is nothing on the other side, because you are driving and looking at the same time, and it is difficult to solve this problem with rules.

overall, we are a little excited and a little thrilled, but more importantly, we have a sense of mission and motivation.

tencent news "high beam": from the 520 ai day to the 730 ai intelligent driving technology conference, xiaopeng motors has achieved the goal of being able to drive the car all over the country and being easy to use all over the country. on 827, it released the "end-to-end four steps". please briefly introduce the thinking and implementation path behind this. in addition, does this mean that end-to-end is already a recognized and sustainable technology route? are there new solutions and paths that may emerge in the future?

li liyun:we have now reached the stage where the system is easy to use across the country. we have opened up scenarios that were previously difficult to cover with rules, such as roundabouts, u-turns, and small roads. why can we do this? not only because we are gradually turning to ai data to drive problem solving, but more because we have very rigorous verification engineering considerations.

there are 2,595 cities in china, including prefecture-level cities and county-level cities. however, in terms of mileage, our test fleet has actually traveled 7.56 million kilometers, which should be the top level in china and even in the world. we care very much about the actual safety of users when using the car. we not only have this intelligent driving capability, but also hope to verify this through the verification fleet.

the goal of end-to-end is not end-to-end, but to effectively utilize data. the essence of end-to-end is an extreme imitation learning, which requires the ultimate utilization of data without loss. in order to achieve this goal, our basic strategy is: streamline the model, open up the data channel, and establish a more powerful ai architecture. we believe that relying on end-to-end capabilities can help us solve most corner cases on the road, thereby realizing the l3+ autonomous driving experience.

tencent news "high beam": how do you ensure the actual effect during the high-frequency iteration process? is there a hard rule that front-line engineers and related teams must experience the car in person?

li liyun:we can start with the high-speed ngp. at that time, i was the person in charge of the high-speed ngp project, not the person in charge of the center. i had just returned from wuhan. at that time, everyone, whether it was writing algorithms, making models, working on the middle and back-end, not to mention the students who made products, even those who were more inclined to supplier management and supplier parts development, needed to ride this car. after experiencing it, we knew what kind of product we should provide to users.

when i was doing high-speed ngp in 2020 and 2021, because i had just returned to china, the first thing i did when i came back was to get my domestic driver's license, and then the first thing i did was to drive like crazy on the highway. it's not that the speed is crazy, and i dare not say that i am familiar with all the highways in guangdong, but at least i am familiar with all the highways in guangzhou and the surrounding highways. i have driven almost every inch of the highway several times, and really experienced all kinds of high-speed scenes and working conditions. i also know the value of high-speed ngp to users. now if you buy a car worth 200,000 or 300,000 yuan, if it does not have advanced assisted driving functions, especially on the high-speed ring road, this car may be a bit outdated or some functions are missing.

later, when we reached the urban ngp stage, we certainly couldn't drive on every inch of road, but i also drove through a lot of cities. at that time, i drove all over the country with the product testing and r&d team, basically in the yangtze river delta, the pearl river delta, and cities around beijing. we drove through every city, such as the connecting roads between urban areas, suburbs, and cities and surrounding satellite cities. i really saw the vastness of china and the complexity of road conditions. it was also at that time that we really felt that it would be very difficult to sustain high-precision maps based on heavy experience, so we became more determined and believed that using ai end-to-end to solve this problem was the right direction.

now that we can drive and use our cars all over the country, we like to explore some small roads or go to places with low coverage. i think it's very interesting. throughout the process, we encourage and even require students who are directly related to cars, whether they are algorithms or models, to ride our cars and experience them.

tencent news "high beam": in the previous process of opening cities, which city do you think has the most complicated road conditions? is it easier to open cities after end-to-end? the generalization ability is stronger.

li liyun:in fact, i have been to many cities, many of which i never thought i would go to, such as lianzhou, some cities in northern guangdong, or many small cities in the yangtze river delta. i think it can be divided into several dimensions. first of all, the dynamic part, the driving habits of people in different places are also different. some places may drive more intensely, while others may not, depending on some local driving habits.

in addition, the rules for giving way to pedestrians in traffic regulations vary in different places. some places clearly state that you must give way to pedestrians, while some places do not seem to state that you must give way to pedestrians, and you just need to ensure safety. however, in some places, you must give way to pedestrians and stop completely, while in some places, it is not stated and you can give way to pedestrians at your discretion.

also, there is the issue of traffic lights. in beijing, you can turn around directly when the traffic light is red. for example, at an intersection in beijing, there is no sign that says "no turning around" when you exit the leftmost lane, and there is a round red light in front. sometimes you can turn around directly, even if it is not written. but it is not the case in guangzhou, so it is very interesting. the waiting area is also very special. in some areas, the waiting area is implemented when the green light is on, but in some areas, it is not the case. in the process of turning left, the front suddenly stops for you, which is also very different.

no matter what method is used, rules or end-to-end methods, we have to face various rules. this is also the advantage of the oem, because we have native user data, and we can infer the understanding of traffic rules through user behavior, and reflect it in our particularly agile and generalized models. for example, in the u-turn scenario, the gameplay is very different. i think in general, some smart city traffic understanding will be a little more difficult, and more models are needed. this is also the need and necessity for our xbrain to understand the scene.

for drivers, traffic rules are important, but safety is more important. safety is always the most important thing. for safety, the most important thing is definitely the safety of vulnerable traffic participants, so i think the biggest challenge is to deal with electric vehicle pedestrians. in some cities without dedicated non-motorized vehicles, the challenge will be even greater.

xiaopeng motors has nationwide data. xngp is based in guangzhou. there are small roads in the urban villages, where there are many small battery vehicles, couriers, and takeaway drivers. this is one of the reasons why we challenged the high-difficulty mode at the beginning. if we want to face the challenge of being able to drive and use it all over the country, this is a challenge we have to face.

end-to-end is just the tip of the iceberg. the 95% below the surface is more important.

tencent news "high beam": all companies are talking about end-to-end, but can the user experience level perceive the differences between models? what is the differentiated competitiveness of xpeng motors?

li liyun:i won’t directly compare ourselves with our competitors.

from xiaopeng motors' perspective, what is the most important thing that end-to-end brings to us? the first is that r&d efficiency has increased. for example, u-turns, roundabouts, and u-turns were previously only available in a few cities with high-precision maps, but now they can be handled in almost all cities. because we embrace the end-to-end capability, for invisible places, we only vaguely know that there is a u-turn here, because you can't see it clearly at all. for you, a u-turn is like seeing a vertical front, and you can't see anything.

from the appearance, or from the difference between what users can feel and what we can feel, we have a strong ability to imagine invisible places or to walk and look at them, just like humans, naturally integrating and walking and looking at them. we are one of the few who can turn around anywhere in the country, and users have experienced it.

it doesn't mean that u-turns and roundabouts cannot be done based on rules. based on rules, these can be done, but it takes a lot of energy and time. we jokingly say that you need to identify the scene first, then build the road, then make the path, then make the prediction, and then make the speed. this chain may take half a year in the past, but now it may take one or two months. in our opinion, the efficiency improvement is huge.

of course, there is a very important point here. efficiency improvement is also based on the internal strength and the matching of the middle and back-end systems. this is also the advantage and confidence of xiaopeng's long-term investment. because we have been firmly investing in this area for many years, the data closed loop and system, including the entire r&d process and system, are very different.

let me give you an interesting example. i think this is also a point that we can attract talents. when our engineers write code, the first step of evaluating the code is not the teachers, but ai. ai will point out that the style is not good here, there is duplication here, it is not efficient here, and it is wrong here, such as forgetting to return, the type is wrong, the runtime is wrong, or there is a risk of memory stampede here.

in our entire system, ai is not only reflected in the capabilities of the vehicle and the cloud, but also in the empowerment of code engineers. all codes must first be checked by the ai ​​teacher. if they fail the ai ​​teacher's check, they will not enter the process of the tech leader (technical director) or be integrated into the vehicle. ai has penetrated into every corner of r&d.

tencent news "high beam": when did we start using ai to evaluate code?

li liyun:we started planning very early, starting in mid-2022. our embrace of ai is not just about connecting a few models on the vehicle side, or end-to-end. these are just the tip of the iceberg. what you can see is the part above the water, but what is more important is the 95% below the water. that is, the entire system must change, and your system must be built based on model iteration and data mining.

more importantly, you need to change your habits and embrace ai at work to improve your efficiency, so we are comprehensive.

tencent news "high beam": what does the 95% below the water surface include? xiaopeng previously said that there are very few companies in china with more than 10,000 cars, and there are only a handful of companies with 5,000 to 10,000 cars.

li liyun:we cannot compare with those who work on general artificial intelligence. among car companies, our cards should be at the top level, which is also a manifestation of our firm investment. because if you want to attract an excellent ai engineer, it is not enough to have a large number of people. he will ask you how many resources and cards you have for him to use. more importantly, we provide training services to engineers around these cards, including whether the data flow is efficient and whether the training infrastructure is complete. these are the results of our long-term, continuous and firm investment.

in fact, the network or things on the car require more patience, resources and investment. because the car-side network has some similarities in the end, or the big logic is similar. after all, the tasks it solves are the same, but it cannot be said to be completely convergent. in addition, the car side is just the floating part. how you iterate and upgrade the car side in the long term is actually a more important part. this includes not only the data flywheel and closed loop in the middle and back-end, but also the working methods of all employees.

tencent news "high beam": my understanding is that to verify whether a system is easy to use, or whether it has reached a specific target value, first, it is necessary to compare it horizontally with competitors so that users can have a sense of it. second, there must be a systematic verification logic and method internally. how do you do this internally?

li liyun:comparing with our competitors is just the final experience, but from the time we send out a software or any software changes, we have a complete full-chain process. it starts with ai looking at your code. in the first step, ai looks at the code without human intervention. later, humans will start to review the plan. later, we will have a complete sil, hil, and full-chain process of actual vehicle verification.

one of the most important is simulation, which we sometimes call software-in-the-loop. in fact, this is also a term used by car companies. autonomous driving companies like to call it simulation, which we do very well. we have collected tens of thousands of cut-ins. you can imagine playing games in the cloud or replaying this parallel world in the cloud to see what the ending of each story is.

if these comprehensive special game scenes are better than before, you have passed this level. of course, this is just an example of cut-in. in addition, we have a large amount of data to simulate in the cloud, which is about tens of millions to hundreds of millions of kilometers every day. how big is the cloud? in addition, we will also have a hardware-in-the-loop (hardware in the loop test), which is a bench simulation. this is closer to the real car, but its world is a virtual world derived from the previous real world. the methodology is similar. from a closed-loop perspective, we can see whether the evolution of these worlds meets our expectations.

after this stage is passed, we will eventually conduct a small-scale test with our actual vehicle products, allowing r&d to see it for themselves and see how it feels. this is a complete iterative chain, and when it reaches a certain stage, it will be packaged into a software and delivered to our nationwide large-scale actual vehicle testing, and finally to the user's car after internal testing and public testing.

tencent news "high beam": how long does it take to verify a major version iteration?

li liyun:previously, we would release a version every month, and users would receive our official version updates every month. but internally, basically every week we could allow internal car buyers to directly experience the latest version, and they could also ask questions and give feedback.

tencent news "high beam": i think the characteristics of each company's products are actually different. xiaopeng is a relatively aggressive company in terms of technology investment. what is its positioning in terms of intelligent driving product strategy? for example, radical or conservative?

li liyun:first of all, technically we start from the end in mind, because we believe in this direction, so we see it.

in terms of products, we are determined and persistent that the user experience of the software we promote must be continuously improved. i think it is difficult to directly define the style at present, but i can ensure that for each of our software, if it is really defined as progressive, our experience must be gradually upward. this is a monotonically increasing curve, and we can absolutely guarantee that it will not decline. this is our commitment and persistence to our users.

"if you make a big model when you have nothing, then you will have no limit"

tencent news "high beam": what do you think about the idea of ​​putting large models on cars?

li liyun:i think that for car companies, they may not have an all-knowing and all-powerful model in the car, such as a model that can answer questions like "who are you and where are you from?" i think it depends on the application direction. do you want to do driving or language question answering, or do you want to be all-encompassing? the parameters of the language model are all hundreds of b (billion), and our car-side autonomous driving model is at the scale of tens of b, which is already a very good situation.

tencent news "high beam": in the end-to-end trend, can companies that do not have a mid- and back-end layout in the early stage overtake others? how do you build your own moat?

li liyun:as i just said, the vehicle side cannot be said to be completely convergent, but it has some similarities because the tasks it solves are the same.

xiaopeng and i think that there should be many car companies that will eventually have intelligent driving capabilities, but most of them will probably only score around 60 or 70 points. we believe that the car companies that can truly achieve 80 or 90 points must be very determined, self-developed, have strong mid- and back-end capabilities, and are truly oriented towards ai reforms.

we believe there may be only a few companies like this in the world, and xiaopeng will be one of them.

tencent news “high beam”: do you mean globally?

li liyun:yes, global car companies. first, we believe that whether you evolve from rules to models, or go directly to models without anything, the car side is only a part. the more important 90% is the closed loop of data that serves the car side, including data acquisition, collection, cleaning, labeling, storage, and the effectiveness of the training framework. for example, the utilization rate of a training cluster is 10% and 90%, which is different. for example, everyone has a thousand cards, your card utilization rate is 10%, and my utilization rate is 60%, which is different. it is also important to have enough high-quality data to feed the model for training. at present, the amount of data for our model training has reached 20 million clips (video clips).

the second point is that our engineers’ way of thinking and the way they solve problems must also become data-driven. so in terms of moat, firm investment and talent are the most important. in addition, the entire mid- and back-end systems and data systems must also be transformed towards ai.

the third point is that it doesn’t mean that i embrace ai and throw away all my burdens, and then i will become ai. in fact, ai is a very long process. it is a set of working methods and working mechanisms and a set of concepts and behaviors, and even a process of infiltration of changes in your organizational structure. we believe that this process can even take years. so, when ai is not done well and cannot solve your problems, is your original foundation still there? can your original set of capabilities help you to cover a bottom line of 70 points, or even a bottom line of 80 points? i think this is also very important. therefore, i think both accumulation and change are needed.

tencent news "high beam": is it a gradual and mutually integrated process?

li liyun:yes, so if you build a big model without anything, then you have no lower limit. to some extent, ai can bring about a rapid improvement in your average ability and a substantial increase in the ceiling, but what about the bottom line? what about the areas that it does not cover? you still need the previous accumulation to cover it.

tencent news "high beam": how long does this long-term investment process need to last to reach the industry average? for example, xiaopeng motors has been investing for 10 years. i believe that the first five years will definitely not be as much as the last five years. when will this investment stabilize?

li liyun:we will invest with confidence, but the investment will definitely not grow indefinitely. intelligent driving has gone through so many twists and turns, including from rules to ai, from map-based to map-free, including various debates, robotaxi and mass-produced assisted driving, etc. in my opinion, assisted driving has entered the final sprint today, so i think there are basically two important points:

1. in my opinion, on the car side, it will reach a wider range of users by the end of this year or early next year, and the penetration rate of ai will be higher than it is now. this is the first moment of change we hope to see.

second, more importantly, after the middle of next year, we will see that everything from product quality to iteration and evolution will show different speeds and accelerations driven by ai, and will get faster and faster. we are gradually reaching this stage today.

tencent news "high beam": someone has already proposed a one model, what do you think?

li liyun:the essence of end-to-end is an extreme imitation learning, which requires the extreme use of data without loss. to achieve this goal, our basic strategy is to simplify the model, open up the data channel, and build a more powerful ai architecture. the goal of end-to-end is not end-to-end, but to effectively use data. behind it is the ultimate and strict engineering capabilities, which must be based on a strong infrastructure to make data flow. so end-to-end is just the beginning, not the end.

we have changed the relationship between the three major modules of perception, planning, and control from upstream and downstream series to deep integration, allowing data to flow freely and efficiently in an integrated model, thereby enabling learning and generation from input (scene images) to output (vehicle actions), and achieving driving similar to human experience and intuition.

tencent news "high beam": in the end-to-end era, what do you think of the competition between pure vision and lidar routes?

li liyun:i think end-to-end downplays the hardware area. the core is the network. whether you use cameras, lidar, or millimeter-wave radar may not be so crucial, and the difference will be downplayed.

tencent news "high beam": some people say that the day when end-to-end implementation matures is also the day when the industry polarizes. solution providers can only do low-level, low-cost assisted driving, and the leading oems will dominate the mass production of higher-level autonomous driving. what do you think?

li liyun:we believe that end-to-end will enable us to achieve l4 autonomous driving, but the cost should be described as follows: with the development of end-to-end, the hardware cost of a single vehicle will decrease, but the demand for the entire large model deployment, computing power, and data volume will be huge.

xpeng has been firmly committed to ai for many years. whether it is the advancement of models, the liquidity of data, or the investment in computing power, it is at the leading level in the industry. ai system capabilities are the key to making the data flywheel run efficiently. currently, our end-to-end can achieve an average of one iteration every two days, and it will be faster and faster in the future. our goal is to achieve l4 autonomous driving capabilities through software and hardware upgrades in 2026.