🛑直播回放：Ask Me Anything--小鹏在线解答VLA 2.0 | Ask Me Anything—Xpeng Answers Questions About VLA 2.0

Hello everyone! I'm very happy today. This should be my first time using such a vivid, interesting, and relaxed livestream to chat with everyone. In the past, they would do very formal press conferences, but I think today I'm also very happy. I feel we can chat more relaxedly. Ask me anything. I think this is a very, very good start.

Yes, actually, after our stores opened the second-generation VLA test drive, the market feedback has been very good, far exceeding our expectations. Xianming's ugly daughter-in-law will soon meet her in-laws, so very, very many friends are saying, 'You're already test driving it,' including media and our store sales staff. 'So when can I get it myself?' Many people are asking. Let me report to everyone first. Xianming hasn't rested since the Spring Festival, right? But our US team doesn't celebrate the Spring Festival, right? So because it's the China and US teams working together on this, the schedule has been very tight recently. And the effect of this launch is also very good. I can share with everyone that from March 2nd, after our media sharing session, to around March 11th when test drives started, the test drive traffic has nearly doubled. And the Ultra's sales volume, because previously the Ultra's proportion... actually, at the beginning, Ultra hadn't been launched yet, right? The proportion was not as high as the Max between Pro and Max at the beginning. But now there's no Pro, right? Now it's upgraded to Max plus Ultra, but the Ultra's proportion has increased significantly. It's more than doubled, actually. So I see a very happy thing, which is that many of XPeng's users are tech-savvy. They really... not only do they come themselves, but they also bring their kids, bring their moms. Because I have several friends who bring the older generation and the younger generation together to truly experience it. I think they all give me a very different feeling. They say the elderly or children think it drives well. I think this is something I'm really, really happy about. Let me share a small case. Recently, because I've been traveling, going to Beijing, Anhui, etc., I have drivers picking me up locally. Recently, I reminded a driver on both sides, I said, 'You don't need to drive so aggressively. You can change lanes more smoothly.' Later I realized, oh, because I've been driving the VLA so much recently, I feel that driving smoothly is a matter of course. But actually, when a professional driver changes lanes or overtakes, they will first slightly accelerate, then slightly straighten the steering wheel when they reach the spot, and then maintain a constant speed. So you feel a slight jerk. But the VLA feels... I really think compared to a professional driver, in some situations, it's even a little better. Maybe after driving the second-generation VLA so much, you start to be a bit more critical of even experienced drivers. You used to think they drove well, but after experiencing the second-generation VLA, when you switch back to human driving, you feel there's still some jerkiness here, it could be smoother here.

So you see we made a board today. The entire KPI... wow, this board, I find it's not easy to straighten. Really. I can share with everyone, today is March 16th, right? So on March 19th, the day after tomorrow, we will start pushing. We will first push the P7's OTA. Then we will push the G7 and X9. So also within this month, after the P7, P7, G7, X9 users will receive the version successively this month. For other models, we will start pushing in April, which is also very soon. Including the P7 Plus, including the G9, including the G6, and including the G7's extended-range version, we will push them all. Actually, yes.

This question, I'll answer first. Thomas already sold me out. Then Thomas said on Weibo that I said it won't be watered down. Indeed, we definitely won't be watered down. To put it more simply and directly, the version we push to media, the store test drive version, and the internal test version don't have much difference. We won't distinguish that this version is specially tuned for media, or that version is specially for store test drives. So the version you see will be further updated and iterated upon the current basis when it goes into production. In the past period, I'll just say, roughly from when we decided to start pushing to production, which was around the Spring Festival, our media test drive and store version was actually our second major version, with many small branches, about the tenth-something branch. And the production version we are about to push to everyone is the fourth major version, the 28th small version. The reason behind this is that we are constantly iterating quickly, the iteration speed is fast enough, and each version undergoes strict testing and simulation to verify its safety. We won't cut certain features just because we are going into production. But we do need to specially explain to everyone that for campus and underground garage roaming, parking garage roaming, these two capabilities, we currently think they haven't reached the amazing level we want, like in urban areas. So for everyone's experience and overall usability, we still decided to temporarily stop pushing them in the first version, not to push campus and underground garage roaming. But we will continue to push them out in the next version not long from now. So the capabilities we are pushing this time are actually very responsible and rigorous. Yes, the core is safety first.

Right, from another angle, Xianming knows that at the end of last year, I was hoping that in this year's first version, we would release campus and parking garage. But to be honest, if the previous L2-level assisted driving for campus and parking garage, I think might be a score of 10 or within 10. I think we've achieved 60 today, but I expect to achieve 85 or 90 before truly pushing it to users. Because users have already seen such good experiences on highways and in urban areas, they will expect the same experience in campus and parking garages. Yes. At this time, if one experience is 90-something points and the other is 60 points, I think it's hard for them to accept. So from our perspective, we hope everyone gives us a little more time to refine it, to do better in campus and parking garages. Like the question you just asked, that many people think it might be watered down. I think in the past rule-based era, many people joked that wherever a company's R&D team is, the roads nearby drive a bit better because you use it and test it every day, writing rules there. I think in the past, including XPeng's XNGP, we wrote rules. So you find that writing rules, from first-tier cities to fourth, fifth, sixth-tier cities, from China to overseas, especially to many places without traffic road signs, because China's roads are very good. At this time, rules are a point that both makes China drive well and makes the world drive poorly. So I think it's a very happy thing that this time, I can't say there are no rules, we are working hard to eliminate rules. Right? Yes. Only in this way can we have the ability to go global. Only in this way can we jump from L2 to L4. So everyone can rest assured, we are pushing out with exactly the same version. So all friends can be completely at ease.

First, this version is definitely our first production version. The first version means that in the future, it will definitely be the version with the lowest score, because we are still continuously iterating. So every subsequent version will keep going up. This is also what we've been saying, we are actually building a system and an infrastructure behind the scenes to make the entire iteration speed fast enough, including it becoming more and more scalable, to expand. This is a core, core issue. This is why we say we hope to build not a specific software or specific version, we hope to build a complete system. Now, if... I often look at the team's comments or daily reports, and I know the issues people are concerned about now. We do have areas that are not done well, for example, navigation. Navigation is something people often say, 'Navigation took a wrong turn' or 'route selection is incorrect.' Why? Previous navigation was basically done with rules. It's like you have a god's hand, you draw a line first. For example, the earliest high-precision maps, or lane-level navigation very close to high-precision maps, or at specific intersections, if you can't see, I add some pre-drawn routes. We joke that it's like driving on rails. You definitely won't be wrong. But why didn't we do that? We still hope the entire thing is solved through model iteration. Because we don't want it to only solve specific scenarios, we want its navigation to be... as long as the navigation information is valid, I can drive underground, on the ground, on village roads, mountain roads, dirt roads, anything. Including XPeng is not just aiming to be a local enterprise, our goal is definitely to go global. When there is no such high-precision information, what should you do? The model's generalization ability must be strong enough to drive on various roads. So this is why, although our navigation sometimes makes mistakes now, we still haven't used rules and high-precision maps, or any such targeted solutions. We are also constantly trying new methods and new data internally to solve the problem. We also see it continuously moving forward. So although the model's navigation is occasionally inaccurate, our capability will become stronger and stronger. We also hope that everyone will soon see that navigation becomes seamless, you can basically trust the car to it.

I just saw, actually, a Max owner... I think there are actually many points here that I myself think all our friends should pay attention to. I think VLA is XPeng's first version towards L4. The version we are about to release, I think in the past, we spent a long time solving offline issues, making safety good. Right? I believe today there are still many places, for example, what was just said, why we removed the underground parking garage, campus, including roaming, which we had already achieved 60 points in my opinion. Maybe in the next version, we expect to score higher, meaning we expect to be safer. Actually, we still have many capabilities not done well, right? For example, our start from standstill, not all scenarios can start from standstill. For example, if you are in a dead end, your start from standstill requires several maneuvers to get out of this alley. This requires our software and model capabilities to further do these things. But actually, from another angle, why are we very confident this is facing L4? That is, in the past, we did many things related to algorithms, completely using AI to solve, not using rules to solve. You must remember, seeing problems is completely solved with AI. I think in this, doing the algorithm well, making the automation efficiency high, this is what we have been refining for the past half year, more than half year, a good half year. So I still want to say, today this version may be good, but for us, the next version, and the one after that, I expect them to be even better. So I still expect everyone to understand that this is assisted driving. I think we will definitely make peace of mind better. Yes, not only peace of mind, high efficiency, but also smoothness.

I want to use some not very accurate metaphors, because it's easier for everyone to understand. I think the Max version is facing L2, we expect to make the best L2. I believe through distillation, it will turn this best L2 into... on our current basis, it can also improve a lot, a very large proportion. I believe today we are already the T0 of Max. I think at that time, when our Max distillation comes out in the second half of the year, we will redefine what is the T0 of the Max version. But I still think Max is not facing L4, it's not solving all scenarios, it's not solving all... basically, the corner cases we encounter in main scenarios. So I think the Ultra version is facing L4, it is able to... what I expect is one takeover per thousand kilometers or less. While the Max version actually does best at hundreds of kilometers, I think it can achieve... a thousand kilometers is very, very difficult. I think a thousand or even ten thousand kilometers takeover is a true L4.

Actually, the internal Max version progress is quite smooth. The entire Max distillation version, we are also continuously pushing it forward. Relatively speaking, from an engineering implementation perspective, we will first push the single Turing chip Max, because its computing power is relatively more suitable for the current system. And we will also gradually push out all the past dual Orin Max models before the end of this year. So at this year's launch event, I also mentioned this. Of course, the big brother set OKR on the spot. Actually, let me tell you a very interesting story. We initially thought the Max version, before any testing, even with distillation, might be... assuming the current Max version capability is 100, it might be able to reach 120, 130, 150, maybe. But actually, after distillation, we found it did much more, far more. So soon everyone will see our P7. On March 18th, we will also bring everyone the single Turing Max version and dual Turing Ultra version. That is to say, we really see different friends, some may like Max, it might be assisted driving, the fun of driving, and some may like the Ultra version, facing this new generation, this stronger intelligent agent, or the fully autonomous driving system. So please also pay attention to our P7 on March 18th, our P7 will bring new colors and more versions.

Yes, we are on this path, and our current internal test version can give all models and all owners the same, or everyone feels amazed experience. Actually, from another angle, we launch P7 on March 19th, then a few days later launch G7, everyone will see why we have a few days between each version. Basically, on the basis of a base model, we use different models' data for some SFT, some RL and a series of things. So my own view, from a technical logic perspective, is completely the same, right? Yes, there is no change in the framework. We hope this process will become more and more robust, the generality stronger and stronger. This is also a technical direction we are constantly pushing. I see Master, so here I'll do a small advertisement. Now several friends' companies have come to us, saying they are interested in cooperating with our VLA. They keep asking about its generalization and robustness on other cars. I want to tell everyone, we have already done quite well on this, we expect to do better, to let more cars use our second-generation VLA. Everyone can look forward to it.

Yes, I believe pushing to more models, we will do more pushes within this year. Actually, today, everyone can see that at the beginning of this year, I merged XPeng's cockpit and autonomous driving into one, all handed over to Xianming. Xianming has been under a lot of pressure recently, managing a R&D team of several thousand people, a bit tired. I don't know if he'll age a bit after a year. But I see the current situation, I think it's very, very, very happy for me. Why? Because facing the AI era, from the original smart car, it may have become not just autonomous driving, I think it's an intelligent agent combined with a car. In this case, autonomous driving is just one capability of the intelligent agent, or a main capability. At this time, the future smart cockpit operating system, voice, the car's brain, the car's cerebellum, the car's capabilities, how they couple, actually this is being done in one team. And we are now considering more, starting to do powertrain, chassis, such things. So today XPeng's entire architecture system is doing architecture integration. Once this integration is done, basically in the second quarter of this year, when it's mostly done, in the third and fourth quarters, everyone will see our accelerated forward progress. Many car platforms will be done better and better. So besides driving safely, passengers will also be more comfortable, the entire car will be smarter.

I think if you compare the current second-generation VLA with V14, it can't be compared because it runs in a different country, different roads, different environment. I think I have indeed seen many media recently doing our comparison with Tesla's V13. I think just from V13, I think we have a clear advantage. But from another angle, because XPeng is in China, our data is in China, we are more familiar with Chinese road conditions, so we are very much looking forward to V14 entering China, or at the end of this year or early next year we enter Europe, enter more countries in Asia. Actually, only in the same country, same complex road, same time period, I think there is more comparability. But from another angle, my own feeling is that when I drive FSD in the US and our VLA in China, I think our VLA handles more human-vehicle interactions better, for example, many delivery guys and many pedestrians, and narrower roads. I think our handling effect is better. I think this is a characteristic of China, not just China, Europe has many small roads, Southeast Asia also has many small roads, because the differences between different cities are very large. So I think if VLA develops so rapidly today, I personally think when we enter many countries, in different road conditions, XPeng is very likely to have more advantages.

How do you see it?

Actually, I think first, Tesla is a very respectable opponent and company. We prefer to call them a friend, not an opponent. This is one thing. Then, many people comment on whether our technical route is the same as Tesla's. Actually, to be honest, I don't know how Tesla does it either. None of us know. Tesla has never released it. So it's more like we are crossing the river by feeling the stones, we've gone through many pitfalls, and wasted a lot of money, not a little, a very, very large amount of money. But we believe one thing, that if we look at this problem from the starting point, everyone's solutions may ultimately converge. But we indeed, no one has told us how Tesla does it, we ourselves cannot judge whether a good solution should be like that. Throughout the process, I have a consensus with the big brother, we hope not just China's autonomous driving is our own, XPeng alone pushing this forward, we hope to pave a way for everyone, or do a more cutting-edge exploration, and then push the entire China autonomous driving industry forward. We also hope everyone can communicate more with us, and we also hope we can see the entire autonomous driving industry enter everyone's lives together. Because only in this way does doing this have meaning.

I think why don't you say it, you are rational, I am emotional.

First, the big brother said 5 times, that is... okay, I'll continue. Actually, we have some evaluation standards internally, some evaluation criteria. We generally don't just use the takeover rate number to evaluate. Generally, as we've said, internally there is CCES, which stands for safety, compliance, comfort, and efficiency, several different dimensions to evaluate, and finally output a comprehensive score. If looking at the comprehensive score, basically the improvement is indeed around 4-point-something to 5-point-something. So this also basically matches this number. So the big brother didn't just say it casually. And另外, we also have different KPIs, including takeover numbers, balancing everyone's average driving scenarios, including evening rush hour, morning rush hour, commuting, and normal conditions, the takeover rate is also about five times. That video from the Nürburgring is basically this number. Our own statistics from a large amount of testing, the KPIs are also about this number. So if it were 100 times, I wouldn't dare to accept it.

Actually, many media teachers, when they drive for two hours during morning or evening rush hour, as long as the route includes both complex and basic road sections, you will see... this is what many people say. But for me, I used to be like Xianming, talking about safety, comfort, efficiency one by one, talking for a long time and people don't understand. And I myself think, including this time, why I proposed to make an autonomous driving that even moms like, because I think for most of our driving users, they actually don't understand your technology, what is MPI, OK. I think they actually just want to know, can I trust it? Is my driving emotion stable? So from another angle, why did we make that water test? That is, the water placed on the car, driving with autonomous driving, how much will it spill? If it spills very little, you drive well, if it spills a lot, you drive poorly, right? So I think in our second-generation VLA, why it's called peace of mind, because the first is safety, then comfort, then smoothness, and so you feel your emotion is very stable, and you dare to trust it. So the second-generation VLA, I think a very important point is that it has fewer hard brakes, and it won't bring an uncomfortable feeling during lane changes, acceleration, deceleration. So this feeling is an emotional stability index. I think it has improved 5 times. So indeed, I saw several media teachers evaluate that our second-generation VLA's hard brakes are zeroed out, because in the one or two hours they drove, they didn't encounter a single hard brake. So why last time I took many moms from the cafeteria to drive, they were very worried at first, but after getting on, they said, 'Wow, wow,' they said they felt it understands everything, and is smarter than me, smoother than me, and more attentive than me, looking everywhere, never distracted. So its emotion is very stable. I think emotional stability is very important. So what I just said, I am from a mom or a novice driving autonomous driving, their subjective judgment. And to be honest, although it's 5 times now, I gave him the goal to improve another 5 to 10 times by the end of this year. So this is a challenging OKR. But I think we really can do it. But I especially want to tell all friends that if you only watch videos, it's very difficult to judge which autonomous driving is better, which is more comfortable, which is smoother. I really think you must come to our store to test drive and experience it. Only through test driving can you feel, wow, it really makes you very calm, stable, comfortable, and finally gives you peace of mind, a 5-fold improvement. So it feels like now store test drives, before it was one person coming to test, 'I think this car is good, I'll buy it.' Now it's the whole family, old and young, not only do I need to drive well, my elderly and children also need to feel it's very good, very comfortable, very liked, and then they will buy, right?

Actually, when I saw this video, I was still a bit scared, my heart was quite complicated. Why did we, this group of people, come to do autonomous driving? From the beginning, doing it for so many years, including the big brother actually asked me the first question was, 'Can you handle doing autonomous driving? Can your heart handle it?' This is something that makes everyone very emotional. Actually, we do autonomous driving hoping to ensure people's safety, to ensure everyone's safety. Watching this video, actually let me tell everyone, after we saw this case, we looked at the backend data. The second-generation VLA did decelerate, but this deceleration was not enough to stop. This is a real situation. This also gives us motivation, that we are still far from perfection. This is also our driving force to keep moving forward to solve these problems. We hope to really achieve enough safety in any sudden situation. I think this is also a mission we this group of people should bear. So we hope through continuous iteration, we can make this matter better and better, and really one day, not just that it starts to decelerate, but it can truly, completely and safely handle these things well.

Right, actually, Xianming just raised a question, why XPeng initially thought autonomous driving must be the most core point of XPeng cars. Actually, the most, most, most important thing is definitely not what lets us free our hands, what improves our efficiency. I think these are all byproducts. I think the core is to greatly improve safety. Everyone may not know, globally in driving, this data is not accurate, but roughly speaking, the death rate is close to 200,000 globally every year.

This is the first data point. The second one, which many people might not know, is that as a National People's Congress representative from Guangdong, in recent years, over 40% of all criminal offenses in Guangdong have been drunk driving. That's because drunk driving is a criminal offense. So 40% of criminal cases in Guangdong are due to drunk driving. If all the cars in these incidents were fully autonomous, and cars could communicate with each other, and every car could, for example, see the data of those four children we just saw—a real case—if they could all brake automatically, or even see a pothole or a collapse in the road, we believe the global annual death toll of 200,000 could drop to 200. Think about how many lives you would save. I think this is something truly worth dedicating our lives to. I think for our generation, and for our children who have already gotten their driver's licenses, they might not need to get one. So that case just now is actually a very, very special corner case. It was completely... a colleague of ours was testing VLA, and because he noticed VLA had strangely slowed down, he looked at the road. At that moment, he went to... at that moment, to be honest, I don't think he knew there were children on the road. So he first braked. Seeing the car slow down and then try to detour, he hit the brakes. Later, he told us he was scared for days. So from another perspective, if we are driving ourselves and we're going very smoothly, we might just go over that speed bump, and the consequences would be unimaginable. So I think safety, safety, safety is the most important thing for us in doing autonomous driving, and it's also the most important thing for society. I think making everyone safer, whether it's the driver or people on the road, is what we all expect. So from another perspective, this time we must get safety right, get the technology right, and solve every abnormal case. I think that's what a truly perfect L4 is. Although today, as Xie Ming just said, we only slowed down and detoured, but it really didn't brake to a stop. So I think we should do even better.

I've seen this video. You've seen it too. Yes, it was quite surprising. First, what exactly happened during this process? We always say the model is doing chain-of-thought, doing COT. It will try to generate different routes and then evaluate the feasibility of each route. So what did it do in those 14 seconds? We don't know, but we mostly think it was reasoning, trying to reason about the feasibility of different paths. That's the first point. The second point is, as we just mentioned about the navigation issue, this was clearly caused by the navigation not updating in time, leading to a road closure. If we strictly required the model to use high-precision navigation information, then it wouldn't have gone against the navigation to detour and choose a new road. So this is also... after the second-generation VLA was launched, whether in testing, media test drives, or at dealerships, we often see videos sent back, and our own R&D team will say, 'Wow, this works too?' It really brings us a lot of surprises. So we also believe the model's progress will be fast enough to make people discover new things every day, some of which you yourselves hadn't anticipated, like, 'Wow, it's so smart, it's thinking.'

I was just self-criticizing, because when I saw this video, it was my first time watching it. I immediately thought, why wait 14 seconds? I think it should have taken 2 seconds. Because I think its reasoning speed is faster than mine. I personally feel that it shows we haven't done well enough yet. Although it does represent generalization—not using rules, not following navigation rules, and adapting based on actual situations—I still expect it to be faster. Maybe 1 to 2 seconds would feel smarter, right? OKR is here again. Xie Ming's face instantly shows a bit more exhaustion. You can see his hair has turned completely white since the launch event. Maybe this year a new term will be added: 'CEO sets OKR on the spot.' Yes, two have already been set, and I've noted them down.