Li Xiang: Cars Are the Biggest Application of AI in the Physical World, but the L4 Ticket Requires Three Things
He divides AI products into three stages: enhancing my abilities, becoming my assistant, becoming family. Today we haven't even done the first stage well, so the real competition hasn't started yet.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The foundation model is the operating system of the AI era
Li Xiang believes that the reason hundreds of electric vehicle companies can exist today is because China has a complete supply chain; but in the future, very few companies will be able to build foundation models. He defines the foundation model as ‘the operating system plus programming language of the AI era’, a new-generation entry point that sits above all devices and all services. So his requirement for the team is: in the coming years, they must ensure the large language model foundation model is in the industry's top three, and to that end he is willing to invest in computing power and compete with the leading companies. He sees this as a dividing line—not all companies can build operating systems and large-scale cloud services.
— Li XiangAI products have three stages, and today we haven't even done the first well
Li Xiang describes the evolution of AI products in three ways: the first stage, ‘enhance my abilities’, where AI is an assistant and responsibility lies with the human, such as using MindGPT to draw a picture and then still needing Photoshop to edit it; the second stage, ‘become my assistant’, where it can take continuous tasks and independently bear responsibility, such as having an L4 car go pick up the kid by itself; the third stage, ‘silicon-based family member’, where it doesn't need instructions, proactively manages the household, and even continues memories. He thinks that today we may not have even reached L3, and if L3 is achieved it's more like the BlackBerry stage—the steering wheel is still there, just like the keyboard was still there. The real iPhone 4 stage will have to wait for L4.
— Li XiangEnd-to-end solves capability, not counter cases
When Li Xiang persuaded the Langbo team to switch to end-to-end, he used his wife as an example: after graduating from driving school, she drove a BMW X6 and frequently scraped it; switching to a Golf GTI, she still scraped it. Solving counter cases didn't help. Later she went to the BMW driving training school's beginner class, where in one day they only learned two things—where to look while driving and how to slam the brake to the floor—and after that she basically said goodbye to scrapes. He says this is capability, not a feature. So what end-to-end demonstrates for L3 and L4 is capability, not solving one counter case after another. Some companies do solve counter cases better than Li Auto because they hire five or ten times as many people and fix every intersection, but that is endless.
— Li XiangThe Mega failure was two problems: market judgment and understanding of pure electric
Li Xiang divides the Mega review into two stages. The surface problem was misjudging the market size: they thought Mega could capture users across all passenger cars above 500,000, like the L9 which captured sedans, MPVs, and SUVs, but the Mega is 5.3 meters long, narrowing the user base; in reality it could only capture users among MPVs above 500,000, and that market is only 4,000 units a month, so achieving 1,000 units would already be a 25% share. The deeper problem was insufficient understanding of pure electric: they thought building charging stations on highways was enough, ignoring that owners in first-tier cities also need to charge within their social circles and are unwilling to squeeze with ride-hailing drivers and wait over an hour. Early Mega charging stations had an NPS of just over 30, later optimized to nearly 90.
— Li XiangAfter 100 billion in revenue, operational capability couldn't keep up at all
Li Xiang says that from having revenue in 2020, it took about three years to exceed 100 billion in annual revenue, but operational capability was still stuck at the scale of Autohome's era of a few billion, close to 10 billion. He gives an example: a colleague hired from a big company asked ‘how do we do this, is there a process’, and then said how our company used to manage it—the description was ‘you remember to turn left at the fourth tree ahead, and turn right when you hit a manhole cover’. He realized that to support a 100-billion scale, they first had to build the roads, not rely on everyone memorizing manhole covers and trees. So he moved Liu Jie, Fan Haoyu, and other strongest product people to build roads in various domains, at the cost of losing a group of great generals in product creation.
— Li XiangA top investor says the common trait in picking people is seeing the essence at critical moments
Li Xiang relays an observation from a top investor: although he doesn't know a method that guarantees picking the right people, in successful cases these people share a common trait—regardless of their aura or expressiveness, whenever the toughest and most critical moment comes, they can always see the essence and then make a choice. This has nothing to do with background, education, or experience, and this trait can continue; they will make such choices multiple times. Li Xiang adds: 2019 was not the hardest time; the hardest was in May 2008 when several small shareholders wanted to kick him and Fan Zheng out of the company, they couldn't raise money, cash was completely cut off, and internal shareholders challenged them. That was the hardest time to date, and also the time of greatest growth.
— Li XiangBeing kicked out in 2008 taught him not to shoulder everything alone
Li Xiang reviews 2008: Shao Zhen was one of the three partners at the time, brought him to Beijing, helped him build the business system, but later became the leader trying to kick him out. During reconciliation, Shao Zhen said the most painful thing was that when the company had financing and operational difficulties, Li Xiang shouldered it all alone without telling anyone; if he had spoken up, everyone would have been willing to mortgage their houses to support the company. Li Xiang says he could see from his eyes that it was true. At that moment he did a huge reflection: before 2008 he was extremely harsh on himself, driving a Polo while others bought BMW 5 Series, working 14 to 16 hours a day without vacation. Afterward he learned two things—to be good to himself and accept his strengths and weaknesses; and not to shoulder everything alone, because problems you can't solve are where everyone's abilities come in.
— Li XiangThe sales system's ladder comes from games
Li Xiang says many systems come from games, including the sales system's ladder method which is entirely based on games, as well as the sales system's operations and incentives. The port rank, career, and economic systems he proposed to HR all come from games. His understanding of games is: games are the fairest, society is a ladder, and learning, exams, and choosing companies are all ladder methods, except the ladder is very fair because you can go up and down. Games have built up the professions, interpersonal relationships, operations, distribution methods, and incentives within the ladder, because if they weren't good, people wouldn't play. He also posts sales rankings because consumers choose the ladder, and companies ranked higher are chosen first.
— Li XiangIn their own words · checked verbatim
So that day it only taught two things: one is how to look at the road, the second is how to step on the brake. But what it taught you is capability.
所以它那一天只教了两件事情 一个是怎么看路 第二是怎么踩刹车 但它教的你是能力
Li Xiang1:04:27
Whenever the toughest and most critical moment comes, he can always see the essence and then make a choice. It has nothing to do with his background, his education, his experience—it's that he has this trait, and this trait can continue.
每当最艰难和最关键的时刻 他总能看透本质 然后做出选择 跟他的背景 跟他的学历 跟他的经历没有关系 是他具备这个特质 而这个特质是可以延续的
Li Xiang1:43:48
Because games are the fairest. One understanding of mine: society is a ladder.
因为游戏是最公平的 我的一个理解 社会就是个天梯
Li Xiang2:32:15
Many people treat the company as a home, and when they go back home, they make demands and play the ladder game—that's backwards.
很多人 会把这个公司当成家 又回到家里去的时候 去提条件 去搞天梯 就搞反了
Li Xiang2:34:15
Figures
| Monthly sales in the MPV market above 500,000 where Mega competes | 4,000 units | 1:24:39 |
| Mega charging station NPS (early) | just over 30 | 1:26:41 |
| Mega charging station NPS (after optimization) | nearly 90 | 2:22:08 |
| Time for Changzhou No. 2 factory from start to full capacity | 15 days | 49:24 |
| Time for traditional automakers from start to full capacity | 6 to 12 months | 49:24 |
| Li Xiang user family base | over 1 million users, 3 to 5 million people including families | 30:18 |
| End-to-end team size | 200 people | 1:12:31 |
| Campus recruitment team size | over 3,000 people | 1:58:52 |
| Li Xiang data filtering ratio | top 3% | 1:15:35 |
Glossary
- VLA / Vision Language Action model
- Vision Language Action, combining vision, language, and action into one model; Li Xiang believes L4 must use VLA.
- VLM / Vision Language Model
- Vision Language Model, a visual understanding model built on large language models, processing two-dimensional images.
- MPI / Miles Per Intervention
- Average miles driven between two human interventions; Li Auto's 2025 goal is to increase from 50 km to 500 km.
- BLM / Business Leading Model
- A strategic tool started by IBM; once strategic intent is set, it automatically generates organization, incentives, and culture.
- DT / Data Technology
- Data Technology, used by Li Xiang to describe the transition from the IT era to the DT era, where data must be customer-closed-loop, atomic, and include finance.
How to listen
Suitable for founders and investors focused on judging AI product stages, automotive intelligence routes, organizational change, and founder retrospectives.
The extended version after 2:08:42 leans personal life; skip it if you only want strategy.