87. A 3-Hour Interview with Li Xiang (Podcast Edition): Geeks, AI, Family, Games and the Leaderboard

Summary

• Li Xiang defines Li Auto’s end state as an AI company that “connects the physical world and the digital world,” with AI representing “the entirety of the company’s future.” He says this is not a post-ChatGPT pivot: the September 2022 Yanqi Lake strategy meeting had already locked in AI and autonomous driving, while ChatGPT merely prompted him in early 2023 to turn a “hidden strategy” into an “open conspiracy.” Cars will remain the company’s hardware pillar, revenue source and real-world AI application.
• The real decisive battle in autos is not electrification but L4, and Li Auto has given itself 3 years to remove the steering wheel. Li Xiang believes end-to-end systems plus VLMs can at most deliver L3, with one takeover every 500 to 1,000 kilometers; general-purpose L4 will require a VLA combining vision, language and action. Reaching the final requires more than 5 million vehicles on the road, its own VLA foundation model, top talent, sufficient compute and capital.
• In Li Xiang’s framework, the foundation model is “the operating system plus programming language of the AI era,” and the true dividing line among EV makers. Li Auto wants its language foundation model to move from outside the top 10 into fifth place, then the top 3 in China, while keeping spatial intelligence No. 1 in China. The autonomous-driving team’s 2025 goal was to raise MPI to 500 kilometers, roughly 10x the level at the time, while expanding the model from its current 100-billion-parameter scale toward 500 billion: “If capability is insufficient, everything else is pointless.”
• Li Auto’s “Lixiang Tongxue” starts with more than 1 million owner households, or roughly 3 million to 5 million people, and ultimately evolves from assistant and Agent into a “silicon-based family member.” In phase one, humans make decisions and bear responsibility; in phase two, the system independently handles continuous tasks such as picking up children and preparing weekly reports; in phase three, it organizes the family without instructions and preserves personal memory. Li Xiang calls Agent the “iPhone 4 moment” ordinary people can actually use, and believes a fully capable Agent could emerge in China within 3 years.
• Li Xiang says OpenAI’s lead comes not only from models but from the combined force of standards, research, R&D, productization, commercialization and distribution. Citing data available at the time, he said ChatGPT held roughly 80% of the global AI-chat product market with 3.6 billion monthly visits, versus roughly 200 million to 300 million for the then-recent Gemini, and compared Microsoft and Apple’s distribution positions to Google winning the AOL homepage. His blunt conclusion: “How hard is it for everyone to respect facts?”
• MEGA’s failure exposed three gaps—market judgment, pure-EV infrastructure and the operating system required for a 100-billion-yuan company—not merely a styling dispute. At 5.3 meters long, the vehicle narrowed demand to the 50万元-plus MPV market, which sells roughly 4,000 units a month; 1,000 monthly sales already represented 25% share, and even 50% share would mean only 2,000 units. Meanwhile, charging NPS rose from the 30s to nearly 90, but the strongest product executives were reassigned to “build the roads,” reducing MEGA’s review panel from VPs to directors and senior managers.
• The organizational issue investors should watch is not Li Xiang’s product instinct but whether he can turn his personal ceiling into repeatable capability. He acknowledges, “I am Li Auto’s product ceiling,” and therefore also its bottleneck. By rebuilding the product lines, bringing back “generals” who can argue with him for 1 or 2 months, and rotating more than 3,000 campus hires through factories and stores, he is trying to move the organization from founder dependence toward a system coordinated by “AI professors, AI coaches and compute infrastructure.”

Deep dive

1. AI Is Not an Add-On to Cars but Li Auto’s Entire Future

• Asked what AI means for Li Auto, Li Xiang gave an unequivocal answer: “It means the entirety of the future.” The second half of the intelligent-vehicle race is not conventional software intelligence but genuine AI—the next necessary stage in building cars.
• He defines the automobile as “the largest application of AI in the physical world,” which is why Li Auto should be an AI company. “Li Auto” is merely the external communications name; the logo has never included the word “automobile.”
• The fuller vision is to “connect the physical world and the digital world,” with cars serving as the hardware pillar, revenue source and real-world application. Zhang Xiaojun repeatedly asked whether this was simply a bigger story; Li Xiang pointed to 10B yuan in annual R&D spending, a foundation model, autonomous driving, and intelligent-commerce and intelligent-industry teams.

2. ChatGPT Did Not Create the Strategy; It Turned the Hidden Line into an “Open Conspiracy”

• Li Xiang says AI and autonomous driving were established as strategic priorities at the September 2022 annual strategy meeting at Yanqi Lake. The early-2023 announcement that Li Auto would become a globally leading AI company by 2030 was not a short-term reaction two months after ChatGPT appeared.
• OpenAI nevertheless played a major role: it made a new AI era visible, allowing Li Auto to stop hiding its long-term ambitions, speak openly about its goals and attract enough talent.
• The thread goes back to the company’s founding. Autohome succeeded online but failed to change the physical links of manufacturing, warehousing, inventory, logistics and user experience. “Apart from spending another dollar online, we had not effectively changed the physical world,” leaving a gap that the third venture had to address.

3. The Third Venture Had to Choose a Forest, Not Guard a Single Tree

• Autohome was early with Web, Windows Mobile, iOS and Android products, capturing the mobile-internet window. But Li Xiang believes competition among automotive websites had already ended in 2009; the following years were easy, yet they trapped the team’s energy in a vertical niche.
• His retrospective is blunt: “Maybe we missed a forest for the sake of one tree.” That is why the third venture had to enter a sufficiently large forest, however difficult, and aim to become the largest company in it.
• Zhang Xiaojun still doubts that Li Xiang had the AI end state fully mapped out on day one of founding the automaker. Li Xiang admits it was initially “more of a feeling,” but he already viewed Li ONE’s single-model strategy as an exercise in human data and software integration rather than a conventional auto project.

4. Autonomous Driving Started Late Mainly Because of Capital Constraints, Not Lack of Conviction

• Li Xiang says serial entrepreneurs understand corporate sequencing: from zero to one, solve for product and survival; from one to ten, once revenue exists, build the platform. Li Auto had the weakest financing ability and raised the least money in its early years, so it had to make the product work first.
• After the product gained market acceptance, Li Auto IPO’d in the US and Hong Kong in 2020 and 2021, respectively. Only then could it systematically invest from 2020 onward in platforms including AD Max, AD Pro, the SS cockpit and the XCU vehicle-domain controller, gradually becoming a Tier-1 supplier of critical controllers.
• During the pandemic, an autonomous-driving supplier took 2 weeks to respond to an email and concluded that Li Auto could not build the system. That experience of being “bullied” pushed Li Auto to take autonomous-driving R&D in-house, starting with the domain controller.

5. Li Xiang Understands His Own Growth as an AI Training Pipeline

• In middle school, without a computer, he spent his pocket money on computer magazines and books and used Telnet to access information. Once he finally had a computer, he surpassed the classmate who had said he was “not qualified to talk about computers” within 1 month, because 3 years of reading had already prepared him.
• He summarizes his method as “learning, verification and growth”: read and talk broadly, set goals to solve problems, then review the process and turn it into capability. In AI terms, these roughly correspond to pretraining, post-training and reinforcement learning.
• That is also why large models feel intuitive to him, while he has long found rule-based algorithms and knowledge graphs awkward. He believes the internet achieved equality of information, while AI will promote “equality of knowledge, cognition and capability”—though he remains skeptical about “technological equality.”

6. OpenAI’s Lead Comes from Five Capabilities and Two Super-Entrances

• Li Xiang says OpenAI is “even more remarkable than Google was to the internet,” and argues that the company must be judged across five layers: defining industry standards, research, technology development, productization and commercialization.
• Distribution is the often-overlooked sixth layer. Google taking the AOL homepage and Baidu acquiring hao123 each turned search into an exclusive entrance; in Li Xiang’s view, OpenAI securing Microsoft and Apple is similarly critical.
• Citing data available at the time, he said ChatGPT held roughly 80% of the global AI-chat product market with 3.6 billion monthly visits, while the recently launched Gemini had roughly 200 million to 300 million. To those who still refuse to acknowledge its lead, he asked: “How hard is it for everyone to respect facts?”

7. Agent Is AI’s Real “iPhone 4 Moment”

• If he were OpenAI CEO, Li Xiang first rejected the idea that he would do a better job than Sam Altman. He uses OpenAI’s tiering: L1 is a chatbot, L2 is a reasoner for professional users, and ChatGPT Pro’s $200 price tag shows it is not a mass-market product.
• At L3, Agent, the system can independently and continuously complete tasks without dense prompting. Li Xiang calls this the “real iPhone 4 moment,” because ordinary people can finally use it directly.
• He believes the US may get there first, while a sufficiently capable Agent in China should arrive within 3 years. The unresolved issue is not direction but interaction and product form, which every leading company must rethink.

8. The Foundation Model Is the Operating System Plus Programming Language of the AI Era

• Li Xiang rejects defining companies by a single hardware category. Apple does not only sell Macs; Huawei is not merely a carrier-equipment company; Xiaomi expanded from phones into IoT, ecosystems and cars. Hardware is the pillar business and starting point, not the boundary of capability.
• Smartphones eliminated many touchscreen-phone makers. The key question was not whether the supply chain could build hardware, but who owned the operating system, app store, cloud services and large-scale software capabilities. In the AI era, the equivalent dividing line is whether a company can build its own foundation model.
• His definition is direct: “The foundation model is the operating system plus programming language of the AI era.” The super-products built on it will sit above every device and service, becoming the next-generation entrance.

9. Lixiang Tongxue Starts with 1 Million Households and Extends to Every Device

• Li Auto already has more than 1 million owner households. Counting people rather than households, Li Xiang wants to serve roughly 3 million to 5 million people first, with the same AI delivering a consistent experience in cars, phones, computers and eventually glasses.
• The demand is already visible. Children discuss homework with Lixiang Tongxue, draw pictures and generate comics, then continue calling out “Lixiang Tongxue” after leaving the car. The mobile app is initially designed to capture these existing behaviors.
• In the long run, an AI with its own foundation model must be able to use every device and service autonomously. In the short run, the strategy is not to fight for the entire market immediately, but to use 1 million households to create a stronger cold start than an ordinary startup could achieve.

10. The Assistant Race Is Still About Securing a Ticket, with Ranking Goals Set in Stages

• Zhang Xiaojun argues that personal assistants have become highly homogeneous. Li Xiang disagrees: the true mass-market To C inflection point will come with Agent, and every company is still trying to secure a ticket to AGI L3 and autonomous-driving L4.
• Li Auto admits it is a follower in language models and products, so its goals must be climbable: “How do we move from outside the top 10 into fifth, then from fifth into third?” Spatial intelligence, by contrast, must remain No. 1 in China.
• On privacy, Li Xiang’s answer was categorical: “No.” His explanation is that pretraining mainly uses public data, while personalized memory is converted into Tokens rather than Bits or conventional text and audio records. This is his technical judgment; Zhang Xiaojun followed by asking whether that meant Li Xiang himself would have no privacy.

11. Language and Spatial Intelligence Will Ultimately Converge in VLA

• Internally, Li Auto calls the language model represented by Mind GPT “language intelligence” and autonomous driving “behavioral intelligence”; borrowing Li Fei-Fei’s terminology, the latter can also be called spatial intelligence. The two tracks are separate today, but Li Xiang believes they must connect.
• Language and cognition alone can produce an intelligent person who cannot act; behavior and spatial capability alone resemble a single-purpose worker. A complete human combines vision, language and action, which is why Li Auto calls its end-state model VLA: Vision, Language, Action.
• Li Xiang believes the timing of spatial intelligence reaching L4 and language intelligence reaching Agent may be close. At that point, the two models will “most likely become one model.” L4 likewise requires an agent capable of understanding the physical world.
• Resource allocation therefore cannot be limited to horizontal comparisons among automakers. The language foundation model must enter China’s top 3, spatial research must be No. 1 in China, and the company will invest in whatever training compute the teams require to reach those goals.

12. Autonomous-Driving Data Must Give Language Models a 3D World

• Even when a VLM is fed billions of images, its foundation remains primarily 2D. It may recognize bus lanes, tidal lanes, traffic police and their gestures, yet still fail to know an object’s precise position in real 3D space.
• The reason is that end-to-end systems and VLMs use different foundations. They can interact but cannot use end-to-end learning to determine an object’s exact position. Images and diffusion generation alone also cannot reconstruct the real physical world.
• Li Auto has experimented with putting 3D-vector Tokens into language-model pretraining so models like Mind GPT can acquire spatial capability. Li Xiang says future papers will show the work, though it is uncertain whether future models will still be called Mind GPT.

13. AI Products Will Move from Capability Enhancement to Independent Assistant to Silicon-Based Family

• The first phase is “enhance my capabilities.” After MidJourney generates an image, a person still edits it in Photoshop; AI-written articles still require human organization; L3 autonomous driving still needs supervision. Capability and responsibility remain with the human, so AI merely improves efficiency.
• The second phase is “become my assistant.” The user hands over a task, and AI completes it independently while taking responsibility for the result. An L4 vehicle can go to school, identify and pick up a child, then take the child to swimming or Lego; an office Agent can prepare a weekly report and send it to subordinates.
• The third phase is the “silicon-based family”: users no longer issue instructions, and AI becomes a family member and organizer that understands the person, children and friends and actively manages the household. After the human body disappears, memory can persist, allowing descendants to speak with it much as they would speak with the person.
• Zhang Xiaojun asked whether this implied the end of privacy. Li Xiang said that by then AI would not merely be a collection of capabilities but would also need to aggregate humanity’s best wisdom. Before reaching the genuine family-member stage, he does not advocate forcing a family-style name onto the product.

14. AI Sells Probabilistic Capability, Not Deterministic Function

• Traditional hardware and software provide functions: they accomplish a specific purpose in a specific scenario, and failure is a quality problem or bug. Competition centers on functions, experience and brand.
• AI provides capability, whose core is probability. Li Xiang uses Liu Xiang as an analogy: “If I raced him 10,000 times, I might win once or twice,” but Liu Xiang’s probability of winning would be far higher. Demanding 100% success every time is applying a functional standard to capability.
• Product value comes from the intersection of user needs and technical capability. The larger the intersection, the higher the value. If teams focus only on product form without understanding model capability, they waste resources and create internal friction.
• This also changes testing. Whether a vehicle has an accelerator and brakes is a function; whether it drives well, quickly and safely is capability. That must be tested through a world model and real-user experience panels, not accepted item by item under legacy software standards.

15. Doing AI Products Without Research Is “Carving Flowers into Stone”

• In the functional era, companies could gain understanding by buying, experiencing suppliers’ solutions and acting as “guinea pigs.” Li Xiang genuinely learned about software entrances from the iPhone 1 and about range-extender shortcomings from the Chevrolet Volt.
• In the capability era, external observation is insufficient. A company must conduct research first, then technology development and only afterward productization. Research also includes the relationship between models and human memory, not merely engineering metrics.
• He agrees with Yang Zhilin’s view that “the model is the product”: search, dialogue and other applications are “eggs laid along the way” by model capability. OpenAI will also enter vertical fields such as search because once a capability exists, it can extend to every problem it can solve.

16. Of Timing, Place and People, Technology Must Come First

• Li Xiang’s lesson from Bubble Network—“we got up early but arrived late”—is that timing, place and people have a clear order. In business, timing means technology and whether a company can use it effectively.
• A latecomer might build a 10-speed gasoline car and still sell less than one-tenth of the volume of Mercedes-Benz or BMW in China. Range extension combined with software and some AI experience allowed Li Auto to catch the German luxury brands in the short term and potentially surpass them within 1 or 2 years.
• Place means where the company is founded. An AI company outside China or the US may lack sufficient language, market, talent and industrial accumulation. People—users, organizations and partnerships—come after technology and market.

17. The Auto Industry Is Moving from BT to IT to DT, with Control Becoming Closed-Loop Cognition

• Li Xiang calls Ford’s assembly line, later further optimized by Toyota, BT. It turned carmaking from a workshop craft into a reproducible process, allowing ordinary families to own cars at a fraction of the previous price.
• The core of the IT phase is control. When building Changzhou’s first factory, he initially preferred flexible Oracle, but a consultant told him that Chinese automakers usually switched back after using it to SAP—“inhuman, with nothing allowed to be changed.” He realized that industrial software’s value was precisely preventing people from bypassing processes.
• DT means Data Technology. It does not merely move processes into software; it captures an end-to-end customer-facing loop, atomic-level causes, processes and outcomes, cross-business data, and financial information such as revenue and cost.
• Only when finance enters the same loop can each function stop looking only at its own “small patch of land.” The company sees not isolated control points but the full picture of a customer and a business.

18. DT Turns Experts’ Tacit Knowledge into Reproducible Best Practice

• Li Auto initially recruited veterans from Toyota quality, Hyundai cost management and GM manufacturing, but found that they struggled to write down their know-how. The model was in their heads and could only be activated when a problem appeared.
• A complete data system solves that bottleneck. If a business process succeeds at high rates and low cost, it becomes a best practice that can be retained; the way a problem was solved becomes organizational capital rather than remaining personal experience.
• Changzhou’s second complete factory took roughly 15 days from start of production to full capacity. Employees from traditional automakers said their previous companies typically needed 6 to 12 months. Li Xiang attributes the gap to systematic replication of best practice.
• Direct stores benefit in the same way. Employees receive centralized training and software is deployed directly; a correctly located new store can typically exceed 100 monthly sales in 3 to 6 months, sometimes faster than finding a franchisee.

19. User Behavior Data Is Closer to the True Business Cause than Interviews

• At Autohome, Li Xiang did not rely on conversations with users. He studied the full journey: why they arrived, why they left, and how they behaved after entering through different channels. The moment of departure was often the real cause.
• Making users turn a page after every 10 images exhausted them; switching to 50 images per page materially extended time spent. If a text link required entering a second-level page, only about 20% continued; taking users directly to the content could bring that figure close to 100%.
• The same data drove acquisition, retention and revenue, while also measuring how much revenue and profit a dealer generated. Customers complained when prices rose, but the company could use data to assess the revenue and profit contributed by each store.

20. Proprietary Post-Training Data Is Li Auto’s Model Currency

• A general foundation model can look helpless in a specialized field because professional best practice is not published online. Only leading companies know how to collect it, and they must first make the underlying business good enough.
• Autonomous-driving training selects the top 3% of driving segments while requiring compliance, safety and high traffic efficiency. When weights are wrong, the team cannot simply patch the issue; it must adjust the data and retrain. That is precisely why post-training Scaling Law favors Li Auto.

21. Li Xiang Learns AI through Research Meetings, Direct Use and Dense Dialogue

• He attends roughly 4 to 5 AI meetings a week: a small-group meeting every other day plus a company-wide meeting on Wednesday. The agenda is mainly internal analysis of the latest papers and even the smallest best practices from different teams.
• Li Xiang admits he is poor at reading papers directly, so he relies on research teams to explain them. The second learning path is using models and products himself, gaining impressions that cannot be replaced by secondhand accounts.
• The third path is interviews and podcasts. A monologue mixes Why, What and How together, making it hard for listeners to absorb. A questioner first structures intent and meaning, and dialogue improves the transmission efficiency of complex new knowledge.

22. Removing the Steering Wheel in 3 Years Is a Multi-Condition Goal, Not a Single Technology Forecast

• Li Xiang uses assisted driving for roughly 80% of his driving. The remaining 20% is mainly because he is “in a hurry.” His goal is to give the team another 3 years to launch a genuine L4 product that removes the steering wheel.
• The timetable depends simultaneously on technology, product definition, environmental policy and consumer trust. It will not materialize automatically when the model is complete. That is why he repeatedly says, “Give me 3 years,” preserving the conditional nature of the goal.
• The autonomous-driving team’s 2025 target under BLM was to raise MPI to 500 kilometers, roughly 10x the level at the time. This is a milestone, not the L4 end point.

23. One Month of End-to-End Training Outpaced 3 Years of Rule Iteration

• At the beginning of the year, Li Xiang sent the team to experience FSD V12 in different US cities. When they returned, his demand was severe: “Either you build this, or we stop doing autonomous driving.” Continuing with rule-based algorithms, in his view, was no different in substance from using a supplier’s solution.
• Li Auto assembled a team of roughly 200 people and the necessary resources for end-to-end training, producing the first model brave enough to put on a car within 1 month. Li Xiang’s impression was that the progress in this 1 month exceeded the previous 3 years.
• During a test drive, Zhang Ying found the car’s response to a pedestrian being avoided by a neighboring vehicle “very human-like,” while reaction speed had improved by several multiples because end-to-end no longer passed through 4 serial steps. Li Xiang used this to distinguish capability growth from old-style functional patching.

24. Pure Vision Drives; LiDAR Remains the Family Seat Belt

• Li Auto removed corner millimeter-wave radars and made end-to-end driving vision-only, but kept forward millimeter-wave radar and LiDAR. Li Xiang emphasizes that this is not because the model route is impure, but because of additional safety redundancy.
• Chinese roads at night may contain trucks with broken taillights, vehicles stopped in the main lane and poorly marked construction. At the time, cameras in no-light conditions could see only slightly more than 100 meters, while LiDAR could see roughly 200 meters—enough to support AEB at 130 km/h.
• Li Auto is also developing AES and two-stage evasive maneuvers, aiming to eliminate more than 90% and eventually all major injury and fatality accidents. Scrapes may remain, but the lives of family users must come first.
• Li Xiang even believes that if Musk drove extensively on Chinese highways late at night, he too would retain a forward-facing LiDAR: Tesla values safety as well, but the answer must be judged in the specific environment.

25. End-to-End Is Enough for L3; L4 Needs VLA-Level Cognitive Generalization

• Li Xiang believes end-to-end plus VLM can deliver L3, with roughly 500 to 1,000 kilometers between takeovers and a much easier in-car experience. But generalization remains far too weak for unsupervised L4.
• Rule systems can generate 3 new corner cases for every one they solve because patches fit only specific scenarios. “External brains” such as Wang Xing and Lu Qi may not provide direct answers, but they continually push the team to ask how humans work.
• He uses his wife’s driving as an example. Moving from a BMW X6 to a smaller Golf GTI did not prevent scrapes; the real improvement came from 1 day of beginner training in a BMW, focused only on where to look and how to press the brake pedal fully.
• Some competitors really do fix corner cases better because they deploy 5x or 10x the manpower to rewrite rules or maps street by street. Li Xiang believes that no matter how large this effort becomes, it cannot replace learning capability.

26. Robotaxis May Not Eliminate Private Cars; L4 Could Make Space More Valuable to Own

• Li Auto’s mission remains “create a mobile home, create a happy home.” The former culminates in L4, the latter in the silicon-based family member, but this does not mean the company must make Robotaxi its only business model.
• Li Xiang counters the claim that autonomous taxis will eliminate car ownership with housing. Renting is often cheaper than buying, and even deposit interest may cover rent, yet people still buy homes because they want stability, safety, comfort and high-quality companionship.
• L4 may lower the cost of family mobility and encourage more people to own a mobile space they can share with family and friends. Whether Robotaxi or private L4 becomes dominant must be reassessed over the next 5 to 10 years.

27. Tesla and Waymo Will Converge; Car Sellers Can Monetize Capability First

• Li Xiang believes constrained-area systems and rule-based algorithms can deliver L4, but general-purpose L4 will ultimately move toward VLA or a superior architecture of the same kind. Tesla, Waymo and every company serious about L4 will not reject effective technology, just as the industry ultimately adopted Transformer.
• He links Tesla V13’s larger improvement to the language-model understanding developed after xAI was founded, including multimodality, RAG and MoE. Google’s paper on putting 3D spatial Tokens into Transformer pretraining likewise shows the routes converging.
• On-device inference cost and compute remain constraints, but once the path works and terminal compute improves, capability can be pushed down to devices. Li Xiang’s view is that convergence is inevitable because “they are all the smartest people in the world.”
• Commercially, car sellers need not wait for L4 to monetize. If Li Auto raises MPI from 50 kilometers to 500 while a rival reaches only 200, the experience gap translates directly into sales and cash flow. He also acknowledges that Waymo’s share of rides in San Francisco is rising.

28. Humanoid Robots Are “100% Certain,” but Must Come after L4

• Asked whether Li Auto will build humanoid robots, Li Xiang put the probability at “100%,” but says the timing is definitely not now. If the company cannot solve L4 cars, it has no basis for solving more complex robots.
• Cars are contactless robots. Roads, signals, traffic participants and rules are highly standardized, making the automobile the simplest embodied environment; physical contact will make everything more complex and difficult.
• Whether Li Auto manufactures or partners depends on supply. It will partner if others do the job well and build in-house only if the market cannot meet its requirements. Phones, computers and browsers are already mature, so Lixiang Tongxue can enter terminals through Chrome or Safari without rebuilding the hardware.

29. Li Auto Is Really Betting on Only Language, Spatial Intelligence and Compute

• Faced with the question of whether Lixiang Tongxue, Mind GPT, intelligent driving and cars are too scattered, Li Xiang says the company is actually doing only 3 things: language intelligence, spatial intelligence and the compute supporting both.
• The CEO’s job likewise compresses into 3 responsibilities: ensure the team is solving the right problems, ensure the people and organization are right, and provide resources matched to the goal rather than chosen arbitrarily.
• He does not want to build a Chinese xAI because his time and financing conditions are inferior to Musk’s. After a spinout, he would have to spend time fundraising; it is better for Li Auto to generate cash and fund model development.
• If forced to choose between Lixiang Tongxue and autonomous driving, his answer is: “I’ll go pick up something else; I won’t give up either.” Models should not rush to monetize: once they reach the 100-billion-parameter scale, they should continue toward 500 billion.

30. World Models, Real-User Panels and Top-3% Data Form a New Validation System

• End-to-end driving did not make driving expert Wang Jiajia irrelevant. It turned him from a rule writer into a coach who defines what an experienced driver is, how to select data, and how to validate safety, efficiency and comfort.
• Li Auto has built 2 validation systems. One is a World Model that can reconstruct and generate traffic scenarios down to individual cities; the other consists of real-user panels of 1,000 and 10,000 people. The results from the two systems are then fitted together.
• Training data is selected from the top 3%, but not simply from the fastest drivers. It must satisfy compliance, safety and high traffic efficiency simultaneously. When a new problem appears, the team adjusts data weights; there is no traditional patch to apply to the model.
• Li Xiang calls this an exam. AI demonstrates capability and must be placed in an environment to compare results, rather than tested like old software by checking whether each button works.

31. Products Should Not Please People; They Should Remove Obstacles Blocking Value

• Zhang Xiaojun asked which users he most wanted to please. Li Xiang answered: “I don’t want to please anyone.” The product should identify value that ought to belong to the user but has not been realized because of an obstacle.
• EVs have value but charging is difficult. Li Auto’s answer was range extension, 5C charging and a self-built charging network. Small experiences are not hard; the hard part is whether the team has personally experienced enough genuinely good products and services.
• He encourages product managers to use bonuses to buy good cars and stay in good hotels when traveling, because experience is product “pretraining.” They must also remain sensitive and perceptive rather than becoming numb to crookedness, shaking and discomfort.
• Finally, there must be discipline and trade-offs. Google constrained itself into a search box; OpenAI into a dialogue box. That is not functional poverty but the insight that “simplicity creates richness; complexity makes things rigid.”

32. Ferrari Is Brand Pretraining; the AI Supercar Retains a 50% Probability

• Li Xiang bought a Ferrari not because it had AI but to understand brand elevation, emotional value and scarcity. Years of experiencing the BMW X7, Mercedes-Benz GLS and Tesla Model X informed his judgment that the L9 should be a family 6-seater; brand capability likewise cannot be learned from articles alone.
• Zhang Xiaojun asked whether buying a 3-million-yuan training card would not be better. Li Xiang said the two are compatible: functional experience, model training, brand and emotional value can all constitute a product.
• Before buying the car, he would have said Lixiang Tongxue would never appear in Ferrari; afterward, he admitted it was possible. Even if L4 turns most cars into space-efficient “boxes,” the pleasure of driving retains an independent value, like high heels.
• His conditional forecast is that by 2030 Li Auto has a 50% probability of making a “very interesting supercar,” but it must be an AI car. Ferrari itself should remain rare and freely designed, becoming a better Ferrari rather than an ordinary technology company.

33. Apple Abandoned Cars Because Perfect Organizations Resist Change and Hold an Old View of Privacy

• Li Xiang does not believe automotive competition will ever permanently end. After traditional automakers came Tesla, and after new-energy startups came Huawei and Xiaomi; when cars no longer have steering wheels, Apple may still re-enter.
• He says leaked Apple materials he saw included Jony Ive saying Apple had no need to design a car with a steering wheel. But actually building a car would force the organization to change. Cars are more complex than phones, and the more successful and perfect the governance system, the harder it is to persuade it to rebuild itself.
• When Apple was worth roughly $2T and Tesla only several hundred billion dollars, it was easy for internal teams to ask why they should learn from a smaller company whose valuation was already considered excessive. Without a clear vision, organizational inertia wins.
• Li Xiang also links Apple’s AI delay to its privacy values. In the rule-based algorithm era, data directly mapped to privacy; after large models convert information into Tokens, he believes there is a new solution. Once Apple understands that, its catch-up speed could exceed expectations.

34. Lei Jun’s Hardware Instinct Is Real; Xiaomi’s Only Advice Is “All in”

• Li Xiang’s assessment of Lei Jun is direct: Xiaomi does not just make good cars; it makes good TVs, air conditioners and other hardware, with an enthusiast’s mindset. It can spot features that a family user may not care about but that attract another group, such as the SU7’s acceleration in just over 2 seconds.
• During the pandemic, Lei Jun spent a long time asking about Tesla, BYD and Huawei, then requested only one piece of advice on building cars. Li Xiang’s sole answer was: “If Xiaomi cars are going to succeed, you have to go All in.”
• The two companies are not only competitors. When MEGA ran into problems, the Xiaomi team offered help and Lei Jun publicly supported it. Li Xiang has repeatedly expressed gratitude.

35. MEGA’s First Mistake Was Treating a 5.3-Meter Car as a Cross-Category Product

• Li Auto initially expected MEGA to take customers from 50万元-plus sedans, SUVs and MPVs, as the L9 had done. After launch, it found that the 5.3-meter length sharply constrained self-driving, parking and daily convenience; the cross-category logic did not hold.
• The actual buyers who were highly satisfied were mostly long-time luxury-MPV users. Sedan and SUV owners might appreciate the product but still preferred to drive a more conveniently sized vehicle themselves. Li Xiang is explicit that the core issue was not simply styling.
• The 50万元-plus MPV market sells roughly 4,000 units a month. MEGA’s 1,000 monthly sales already represented about 25% share; even 50% share would mean only 2,000 units. No marketing campaign could repair that market-size misjudgment.

36. Charging NPS Rose from the 30s to Nearly 90, Turning Failure into a Pure-EV Foundation

• Li Auto initially assumed it only needed to build superchargers on highways. It overlooked the fact that owners in second-tier cities also needed to recharge after entering first-tier cities, and that users’ time was valuable: they did not want to queue for 1 hour alongside ride-hailing cars.
• Early charging-station numbers, locations, app and vehicle-interface design all inherited range-extender assumptions, leaving MEGA users with charging NPS in the 30s. After adding city stations, improving software and certifying high-quality third-party chargers, the score rose to nearly 90.
• Li Xiang observed that many Beijing MEGA owners with multiple cars no longer drove gasoline vehicles to Aranya, ski resorts or nearby destinations, because the phone, vehicle interface and en-route charging network together created confidence.
• Stores also shifted from counting units to counting “shelf space.” MEGA required a larger display area and pure-EV training for sales staff. Li Auto then expanded shelf space and built 5C stations at qualified stores, laying the sales foundation for future pure-electric SUVs.

37. 100B Yuan in Revenue Exposed Missing Operating Roads and Drained the Product Generals

• The L9’s “ceiling” came from platformization that began in 2019 and firm investment after 2020. Li Auto surpassed 100B yuan in annual revenue in just over 3 years, with product growth far outpacing the operating system.
• When new employees from large companies asked about processes, internal directions still sounded like “turn left at the fourth tree, then right at the manhole cover.” Li Xiang realized the company needed to build real roads rather than making everyone carry their own trees and manhole covers.
• Its strongest product leaders—Fan Haoyu, Liu Jie and Tang Jing—were reassigned to build R&D, sales and management systems. As a result, L9 reviews had VPs in the room, while MEGA reviews relied more on directors and senior managers, many of whom had not experienced the previous success.
• Li Xiang does not regret filling the operating gap required by a 100-billion-yuan company, but admits the cost was real: “MEGA was not the ceiling in every respect, but L9 was.”

38. Bringing Back the Product Generals Turned Conflict into a Quality Grindstone

• Li Auto re-established product lines and reassigned Liu Jie, Fan Haoyu and Tang Jing, returning product leadership to people capable of arguing with Li Xiang for 1 or 2 months.
• One dispute was whether the R7 and R8 should be combined into a single model with 5-seat and 6-seat versions. Li Xiang insisted on one model; the team demonstrated through ingress, egress and styling tests that the experience would be compromised, and he ultimately accepted splitting them into 2 models.
• The compromise came not from hierarchy or voting but from evidence that “we cannot make a makeshift product.” Li Xiang cites Steve Jobs’ story of rough stones being put into a machine and polished into beautiful round stones. The strategy committee meets every Saturday and people can slam the table and swear: “Argue however you need to argue.”

39. There Is No Secret Weapon in Pure EVs; L4 Is the Decisive Battle

• Li Xiang offers no mysterious formula for the coming pure-EV war. At each price point, whether the company reaches the ceiling on every value users care about will depend on the combined execution of technology, product, value communication and the commercial system.
• Electrification is merely the ticket to the intelligent-vehicle finals. Samsung entered the smartphone era through hardware; Apple and Google entered through operating systems and large-scale software capability. The auto industry will likewise admit latecomers with different sources of strength.
• The real winner will be L4. Today’s sales, cash flow and model investment are all aimed at securing a future qualification that cannot be purchased temporarily with today’s resources.

40. 5 Million Vehicles, VLA, Talent and Compute Form the L4 Ticket

• Li Xiang’s first hard requirement is at least 5 million vehicles operating on the road, generating enough real-world data with sufficient complexity.
• The second is control of its own VLA foundation model rather than simply purchasing a function. The third is enough capital to recruit top talent and buy training and inference compute.
• These conditions must be achieved at an excellent level, not merely possessed. If all are met, he believes Li Auto can become a company like Apple.
• Conversely, without L4, Li Auto cannot become a trillion-yuan company. A later entrant must deliver the revolutionary experience and business model in which “people do not buy cars in order to drive.”

41. Li Xiang Is Both Li Auto’s Product Ceiling and Its Organizational Bottleneck

• Asked whether he is Li Auto’s product ceiling, he answered: “I think I am Li Auto’s product ceiling.” From zero to one, he had to lead because his long-term hands-on experience driving 3-row SUVs exceeded that of almost any peer CEO.
• He never uses a permanent driver as a substitute for personal experience. He often drives himself to the airport and has the driver take the car back. Deep experience with the X7, GLS and Model X formed the product judgment equivalent of “pretraining.”
• But at the current stage, his personal ceiling is also the company’s bottleneck. The task is to build training and professional systems, have Tang Jing, Xin Yang and others drive competitors over the long term, and eventually give every product owner capabilities exceeding those of the Li Auto team several years ago.

42. A Good Product Manager Must Sense Across Domains, Reject Compromise and Follow Discipline

• The first capability is cross-domain sensitivity. The main inspiration for Li Auto’s range-extender architecture was not the Chevrolet Volt but the Mac’s Fusion Drive: a fast, expensive SSD for programs and a large HDD for storage, which suggested the range-extender concept.
• The Volt mainly helped the team identify a technical issue—that a large battery and a small battery had to be connected. Li Xiang believes the car itself was “made for the sake of making it,” without a clear user-value proposition, so Li Auto did not copy it.
• The second capability is sensitivity. The L9’s lower-body contact when exiting was wrong and could not be excused because Benz and BMW also rubbed people’s legs; MEGA’s suspension wobble likewise could not be tolerated because other MPVs wobbled.
• The third is professional discipline. Experience, definition, validation and correction after problems are found must all follow a complete system. Sensitivity provides inspiration, not permission to act on impulse; product experience is inherently subjective and need not be disguised as objective.

43. AI R&D Is Being Reorganized around Professors, Coaches and Compute

• The first category is the “AI professor,” responsible for pretraining by assembling human knowledge expressed in every form. With MoE, this resembles a knowledge system containing multiple specialties and degrees.
• The second is the “AI coach,” using post-training and reinforcement learning to turn best practice into capability. Such people must understand BT, IT, DT, business environments, data collection and quality evaluation.
• The third provides compute for AI. Former software engineers, product managers and business experts must all find their positions among these 3 new professions.
• Li Xiang believes he cannot do pretraining and is not suited to compute infrastructure. He is more like an AI coach; if he builds a To C assistant, the coach must also construct DT data around human dialogue and memory.

44. A Founder’s Scarce Ability Is Seeing through the Essence and Choosing at the Hardest Moments

• A top investor told Li Xiang that successful founders have no common template for extroversion, presence, education or communication style. Their true commonality is that at the most difficult and critical crossroads, they can see through the essence and make the best choice for the team.
• This ability appears repeatedly rather than by accident, and future results can validate it. Li Xiang also sees “choice” as the CEO’s job, rather than working hard on every operating detail.
• He regards his low tolerance as both strength and weakness. When something feels wrong, he solves it instead of carrying it indefinitely. What looks like anger is often a deliberate lever to amplify the seriousness of a problem; internally he is calm, and the frequency has roughly halved every year.

45. The 2008 Crisis Taught Li Xiang to Stop Carrying Everything Alone and Treat Himself Better

• His hardest moment was not 2019 but May 2008: he could not raise money, cash flow had completely stopped, and several small shareholders tried to remove him and Fan Zheng from the company. The pain was so great that he would start crying just thinking about it.
• Shao Zhen, the partner who led the team to Beijing and built its commercial system, had initiated the challenge. During reconciliation, his most painful accusation was: “You were carrying everything alone and did not tell us.” Had they known about the cash-flow problem, the partners would even have mortgaged their homes to support the company.
• Li Xiang learned to treat himself better and accept his strengths and weaknesses: “If I have never tasted sugar, how can I give sweetness to others?” He then bought a BMW X6, a McLaren MP4-12C and a better home, and began taking vacations. His family and team relationships improved.
• In one financing round, he told Zhang Ying about the risk when cash still covered roughly 6 months and raised about 1.9B yuan. In the next, more difficult round, he met more than 150 investors, became seriously ill with allergies and ultimately went directly to Zhang Yiming and Wang Xing, both of whom invested. In early 2019, he also told the team about cash pressure, and everyone cut costs together.

46. Employees Are the CEO’s First Customer; Time Goes Only to People, Products and AI

• Li Xiang’s first priority is people: recruiting, training, organizational systems, and interviewing all level-18-plus candidates and some level-17 candidates. If the Guangzhou auto show conflicts with employee culture training, he chooses the training.
• The second priority is product, including product reviews, styling reviews and the product-capability system. The third is AI, mainly through study sessions, expert exchanges and internal sharing of best practice.
• A typical day handles product in the morning, recruiting and training around midday, and AI later. He usually arrives at the office around 10 a.m. and does not go out to meet customers or users.
• Major decisions are reserved for the Saturday-morning strategy committee, when daily business does not interfere and the team can debate at length. Li Xiang believes the meetings are high quality precisely because the team is “arguing every day.”

47. “Trillion” Is Not Personal Desire but a Resource-Allocation Tool

• Asked whether he craved a trillion-yuan market cap, Li Xiang answered: “I don’t have much desire. Growth is my desire.” What excites him is delivering long-term meaningful products such as range extension, L4 and Agent.
• He treats the goal as a management tool. Only by setting an inspiring and meaningful target will the organization accept the corresponding cost, allocate initiatives, talent and resources, and keep iterating when the path diverges.
• A trillion-yuan valuation reflects user value and industry standing. The point is not the slogan itself but whether the goal forces the organization to build the required capabilities.
• For Li Auto, failure to achieve L4 means it cannot enter the trillion-yuan club, because only “not needing to drive” creates an experience fundamentally different from existing products.

48. More Than 3,000 Campus Hires Demonstrate the New Management Formula: Care, Recognition and Support

• Li Auto did not formalize campus recruiting at scale until 2022, previously believing that automotive employees needed experience with at least 2 products. With 2 generations of products and internal knowledge in place, the company could finally train newcomers; the campus-hire team now exceeds 3,000 people.
• Beyond 1 to 2 weeks of classroom training, new hires enter the motor plant, vehicle factory and stores. They experience industrial work, frontline workers’ attitudes toward products, family-user decisions and delivery service before returning to class to deepen their understanding of the process.
• Results exceeded expectations. Many level-13 and level-14 campus hires produced value at level 17 or 18. More than half of the projects shown in internal AI demonstrations came from campus-hire teams, including Lixiang Tongxue.
• Li Xiang summarizes the new management approach as caring and listening first, recognizing the solutions people bring, and then providing the necessary resources. The principle applies to people born in the 1970s, 1980s and 1990s in a low-growth era, not only those born after 2000.

49. China’s AI Path Should First Win Domestic Rankings, Then Wait for VLA Convergence

• Given geopolitical constraints, Li Xiang does not advocate shouting “No. 1 globally” from day one. The path is to become No. 1 in spatial intelligence in China, reach the top 3 in language models and services, then combine them into VLA and enter the full Agent and L4 stage.
• This is a growth path requiring research, organization and investment to be determined in advance, not a direct jump from today to global leadership. The US may be stronger, but he insists that “as long as Chinese companies do not give up, anything is possible.”
• He compares it with the US eventually beating Japan in processors and Chinese smart cars challenging German-brand dominance. These examples express possibility, not a guarantee.
• If Li Auto fails, the cause will be a missing link in research, technology development, product, commercialization or internal cash generation. Li Xiang will give everything to the effort, but would also applaud any Chinese company that reaches the fifth stage of AI he describes.

50. AI Should Not Replace Frontline Labor but Redesign the Distribution of Benefits

• Li Xiang says the more one understands AI, the more carefully one must decide what value it should create. Li Auto “will never replace workers” and does not want to replace frontline salespeople.
• The better mechanism is to raise factory workers’ efficiency and income, while AI automatically finds opportunities from large volumes of leads. Salespeople still handle human contact, but can sell more cars, earn more money, buy homes in their cities and send their children to good schools.
• The structural AI question is therefore not only the number of jobs but how benefits are redistributed. Companies need capability growth to have a positive relationship with existing employees’ income.
• Organizations may also be generated from strategic intent. In 2024, Li Auto promoted BLM, or Business Leading Model: once key initiatives are set, organization, incentives and cultural traits follow, rather than departments being fixed first and tasks sought afterward.

51. Games Wrote the “Fair Leaderboard” into Li Auto’s Institutions

• Li Xiang has played games for decades and still spends 1 to 2 hours a day playing Overwatch. He prefers team games with high coordination demands and is less interested in single-player or pure action titles.
• Li Auto’s sales rankings, operating incentives, roles, levels, careers and economic systems all draw inspiration from game mechanics. Games are attractive, he believes, because “games are the fairest.”
• Society itself is a leaderboard: education, career choice, company rankings and income distribution all allow people to rise and fall. Consumers tend to buy brands near the top, and talent tends to join highly ranked companies.
• The responsibility of management is not to abolish the leaderboard but to make the rules fair and transparent, with the possibility of moving up or down. When a large game misallocates rewards and incentives, players leave; companies work the same way.

52. Reversing the Roles of Company and Family Causes Many Relationship Problems

• Li Xiang keeps work society and private life separate. The company is a fair leaderboard; family and close friends follow “contribution equals return.” The two systems cannot be governed by the same exchange logic.
• He observes that many people “treat the company as home, then go home and start making demands and building a leaderboard,” exactly reversing the two systems.
• At the company, the CEO’s first customer is the employee. At home, the most important customer is the partner with whom one spends the most time, followed by the children. Family gives him energy and high-quality companionship, not an organizational identity.
• Having 6 children has also shown him how different personalities grow in their own ways. Family trains acceptance of strengths and shortcomings and builds patience, rather than forcing everyone toward the same answer.

53. A Cat and a Dog Taught Li Xiang That the World Is Made of People and Relationships

• The game that most influenced his worldview was not a mainstream hit but Dāyǔ’s A Cat and a Dog. It was the first time Li Xiang understood that the world is made not only of things but also of people and relationships between people.
• Real relationship ability still comes from “suffering huge losses again and again” in the real world. The hardest relationship to handle is the one with oneself; once that is put in order, external relationships become easier.
• Treating oneself well also means protecting oneself in dealings with others: first reduce the harm caused by their shortcomings, then absorb their strengths, rather than letting their weaknesses keep hurting you in order to prove that they are foolish.
• He uses investors as an example. Listen seriously to valuable views and put them into practice, but do not compete to prove yourself right against amateur or even “pseudoscientific” automotive opinions. That way, you gain resources and support without turning pointless arguments into mutual harm.

54. Intimacy Is About Being Understood and Supported, Not Given Answers

• When speaking with his partner, Li Xiang listens carefully rather than rushing to provide a solution. He finds that once she has fully expressed herself, he usually already has the complete answer; what she needs is attention, understanding, recognition and for him “not to make things worse.”
• His partner also responds to his growth drive. She studies auto-company financials, products and competitive dynamics and shares updates on the children. Both continue learning, turning the relationship into mutual support rather than one-sided sacrifice.
• From “when the family is harmonious, everything prospers,” he extracts the sequence “when the father is diligent, the mother respects him; when the mother respects him, the children are secure.” The emphasis is on the man treating his partner well first, not merely demanding that the children or wife preserve family stability.
• On having many children, he sees no necessary connection with building cars. He simply says that watching different lives grow and accompany one another is “a very happy thing.”

55. Choosing Proactively Creates More Responsibility and Growth than Judging Right and Wrong

• Li Xiang interprets “proactivity” in The 7 Habits of Highly Effective People as actively choosing while remaining positive, not rushing in to help without understanding the situation.
• When he dropped out of high school, he compared 2 paths: university, which might lead to a second-tier degree and a job editing computer magazines, or entrepreneurship, which was already generating close to 20,000 yuan a month and might scale 10x, but could also leave him unable to find work if he failed without a degree.
• If the question is simply whether dropping out was right or wrong, parents and society turn it into a binary judgment. Comparing the returns and risks of 2 paths makes the person responsible, encourages continuous iteration when problems arise, and makes it easier to explain the choice to family and win support.
• The same applies to AI companies. A company that does not do pretraining should not be dismissed for its post-training, nor a weak productization effort used to dismiss its RAG. People who choose absorb usable capability; people who turn everything into a right-or-wrong question often only prove themselves right.

56. Reducing Public Exposure Was an Environmental Choice, Not Retreat from Labels

• Li Xiang says he is “not rebellious,” merely someone who keeps making active choices. The label of not having attended university does not bother him; even when external harm occurs, he can adjust in a short time.
• But he believes the current environment is no longer suited to frequent output as “the founder who never attended university,” delivering black-and-white methodologies. Negative feedback now exceeds positive feedback and the entrepreneur audience is shrinking, so he chose to appear less often on Weibo.
• He does not want to comfort people by repeatedly recounting hardship and is not good at light jokes. Rather than becoming a communicator he dislikes, he would rather let the team speak and “shut up” himself.
• In the rapid-fire questions, he said his MBTI is “probably INTJ,” his favorite food is hotpot, and his current favorite city is Chengdu. He prefers highly collaborative team games; his children like Genshin Impact, while he is not fond of action games such as Black Myth: Wukong.

57. AI Memory May Be China’s Most Attractive To C Startup Bet

• Li Xiang believes current assistants are still at the AOL stage of the mobile internet or the Palm stage of smartphones. The industry is eager to discuss Agent and AGI, but has not deeply built the first-stage assistant: useful functions, help with growth and high-quality companionship.
• Genuine companionship means listening carefully and understanding the user over time, even helping create a diary, annual review, autobiography or book. Genuine help with growth means continuously digging into a problem rather than stopping halfway through a discussion of the Riemann hypothesis.
• He believes China’s To B market is difficult, while overseas already has data companies such as Snowflake worth tens of billions of dollars. To C offers more room; the biggest challenge is memory, and a company able to build long-term memory could become an important startup.
• This capability could also support overseas expansion: AI could directly call each country’s APIs and generate music and required services, potentially bypassing the difficulty of configuring maps and content applications country by country. Li Xiang reduces his own 3 questions to: “What capability to build, what organization to match it with, and how much capital to prepare.”