117: 印奇’s 14 Years in AI Entrepreneurship: All Brilliance That Cannot Close the Loop Is Temporary
Summary
印奇 defines Qianli Technology as a second startup, not a capital transaction: using Geely’s vehicle and industrial-chain scenarios as a foothold, he is continuing Megvii’s unfinished “AI in physical—software-hardware integration—robots” path while building a data loop. In the short term, the focus is “AI+cars,” integrating intelligent driving and intelligent cabins; robots remain the endgame, but he estimates that integrating the industrial chain into mature products will take about 5 years. In China’s commercial environment, long-term goals must be paired with a strategy capable of closing a loop within 3 years.
Megvii’s shift from Hong Kong toward an A-share listing, its prolonged listing process and inability to raise more capital led 印奇 to reorder his core principle from “technology conviction, value pragmatism” to “value pragmatism, technology conviction.” He repeatedly emphasizes the loop: “business model is the best model,” and “brilliance that cannot close the loop is only temporary.” Abandoning the original listing path was not abandoning the loop; it meant being forced to pursue a bigger, longer loop.
Qianli’s intelligent-driving competition is not a battle of concepts, but a scale game built around data, vertical software-hardware coordination and an alliance of leading customers. 印奇 plans to serve roughly 3-4 automaker groups in depth, with no more than 1-2 customers of each type, using Geely to establish the first data loop before exporting it outward. He expects no more than 4 Tier 1 suppliers—and more likely only 3—to remain; Qianli’s near-term goal is first to “stay at the table.”
On technology, 印奇 has chosen a Tesla-like data-driven, model-driven path, but believes the industry has “absolutely oversold” VLA. The shift from map-based to mapless driving is a fundamental leap; once mapless, the priority is to raise the share of modeling while strengthening large perception, large-scale planning and control, data systems, and software-hardware coordination. General-purpose recognition by VLM “may be necessary”; VLA is a trend, but not the core problem in intelligent driving over the next 1-2 years. “The ingredients may not be fresh, and adding a lot of seasoning at the end” cannot replace the fundamentals.
Qianli is betting on an open supply chain to differentiate itself from Huawei-style vertical closure: it must own the algorithms, overall solution, domain-controller design and some core-chip capabilities, while sensor manufacturing and peripheral hardware should not become strategic priorities. 印奇 believes vertical integration is more efficient in the early stage of an industry, professional division of labor and open systems are better during the middle stage, and integration may return at the end. Qianli’s bet is to use Geely as its base, partner with the best suppliers across chips, lidar and other links, and build brand recognition as a supplier.
印奇 sees this large-model wave as the “finals” of the deep-learning supercycle that began around 2015, but believes pure search-style Chatbots cannot cover model development and inference costs; the next phase must move from information services to closed-loop services. Agents can be divided into work, creation, life and emotional-companionship categories. He expects the work segment to be absorbed more readily into existing software systems such as Office and Feishu, while startups still have opportunities to produce breakout products in creation. Life has the largest potential market but depends on operating systems and hardware. The first 3 categories could see successive inflection points over the next 18-24 months.
“Super model + Super App” is not an either-or choice: any genuine Super App must compete at the frontier on both performance and cost, while merely calling open-source weights or performing shallow fine-tuning is unlikely to create a moat. 印奇 still believes fundamental technological change will produce new giants, but startups must concentrate resources on one scenario, control deep training, modification and personalization of the model, and own a software or hardware channel directly to consumers. His common-sense judgment is: “As long as the transformation is fundamental enough, a new company will definitely win.”
Deep dive
1. A 400-Person Cadre Meeting First Turned “Capital Changes” into a Second Startup
On his first day as chairman of Qianli Technology, 印奇 convened roughly 400 cadres. He wanted the former Lifan team to see a real person rather than a name in an announcement, while giving the organization an accurate expectation of its new chairman.
His second message was more direct: changes at the capital level did not mean the company had truly started. “I will treat this as my second startup.” It was both expectation management for the organization and a statement of resolve.
For the former Lifan organization, which had gone through bankruptcy restructuring and then spent 3 years integrating with Geely, the state of the people mattered more than the capital structure. 印奇 wanted his physical presence to signal commitment, while acknowledging that Qianli was still a new organization and industrial system for him.
2. From Megvii to Qianli, the Platform Changed, Not the AI Endgame
印奇 summarizes 3 things that have not changed: AI is his lifelong事业; AI must find scenarios and physical carriers and move toward software-hardware integration; the endgame remains robots. “In terms of the starting point, the endpoint and the path, the essence has not changed.”
The formal change is that he left the old Megvii system while having to integrate Megvii’s accumulated strengths and resources into Qianli. More importantly, Geely provides the industrial scenarios and hardware carrier that AI companies have historically lacked.
He does not see this as starting from scratch. Megvii had been filling gaps in carriers, supply chains and customer scenarios; Qianli is trying to make the AI team and the industrial partner complementary. The change visible from outside is, in his mind, “less than what everyone sees externally.”
3. A Frozen Listing and AI’s Reacceleration Pushed Megvii to the Decision Point
At the end of 2023, Megvii was still stuck in a long A-share listing process and could not raise new financing during the process. As its funding chain tightened, the external large-model wave simultaneously demanded more compute, talent and capital, creating a direct collision.
程曼祺 listed 3 options: seize the large-model startup window, continue pushing Megvii’s listing, or join Qianli. 印奇 ruled out leaving the original team behind to start another business, as well as a capital solution that could provide short-term relief but not support the company’s long-term development.
He had set multiple deadlines for waiting on the listing, with the latest roughly in mid-2024. Without the Qianli plan, the alternative was to withdraw the listing application and raise capital again. Important decisions should “try everything” and gather as much information as possible, but the final answer still depends on cash, the team’s ability to absorb pressure and the decision-maker’s threshold.
4. Qianli Became the Answer Through “Hope,” Not “Fear”
The Qianli plan was initially proposed by Vice Chairman Shu. 印奇’s first reaction was that it was a great idea, but he did not commit immediately because he still wanted to complete Megvii’s original capital path. By Q1 2024, he concluded that it might be both “the only choice” and “the best choice.”
His own explanation was: “You make a choice not because of fear, but because of hope.” The decision was driven not by fear of a failed listing, but by months of conversations that convinced him Qianli could be a better platform for taking AI to its next level.
After restructuring, much of Lifan’s historical baggage had been cleared and its base business was relatively clear. The decisive question was Geely: did its management truly see AI as the future of automobiles and intend to invest for the long term, rather than treat intelligence as another configuration item?
印奇 saw “technology conviction” in Geely’s long history of studying technological frontiers and repeatedly backing technical paths. In his view, one decision is not enough to represent a value system; staying with one path for 20 years is what shows that a company believes it represents the future.
5. AI in Physical Is Both a Value Choice and a Commercial Differentiator
印奇 believes large models are breaking down the boundary between digital and physical, and that almost all future AI applications will require the full capabilities of a Super Model. Without them, to B products easily decay into solutions, while to C products struggle to become pure killer apps.
His preference for AI in physical begins with values. When choosing between AR and VR, he leaned toward AR because he did not want technology to pull people away from their natural state. Highly immersive VR may offer an excellent experience, “but this is not the better state of human life as I imagine it.” Robots, by contrast, can work in physical space and become companions.
His second judgment is that purely virtual data cannot cultivate the AGI he imagines; physical-world data, interaction and conditions for evolution are indispensable, yet still extremely scarce. The third layer is competition: the digital world already has powerful giants, while physical AI is newer and Megvii’s software-hardware integration DNA is more differentiated.
Robots remain “the original calling and the endgame,” but 印奇 estimates that even if the main ingredients are basically in place today, integrating the industrial chain into a mature product could still take about 5 years. That is why Qianli is choosing the clear carrier of cars for now while pursuing a long-term goal alongside a strategy that can close a loop within 3 years.
6. DeepSeek Proved Research Firepower, Not an Automatic Commercial Loop
印奇 first rejects treating DeepSeek as an ordinary startup. It is, above all, a large company—or at least a company ranked near the top by scale—and cannot be viewed simply as a research institute or scientific organization, because organizational character ultimately determines commercial strategy.
He recognizes that DeepSeek released impressive results through a relatively non-commercial open-source approach and achieved mass-market reach. “This is a very good example.” But research results, open-source influence and commercial sustainability are separate matters; the step itself does not mean the loop has formed.
By comparison with Megvii, he says the team did build “the best research institute in the Eastern Hemisphere,” achieved many global firsts and supplied key personnel to DeepSeek, OpenAI and other institutions. The difference today is that AI’s explosive power and public attention are both greater, making research achievements more likely to break into the mainstream directly.
That experience has made him insist on distinguishing commercial companies from research institutions. Ultimately, resource allocation must be anchored to customer value; otherwise, even dazzling research influence may be only temporary momentum.
7. The “Genius Youth” Model Creates Talent; Commercial Value Keeps It
Megvii built an early training system around students with strong mathematics-competition, informatics-competition and quantitative foundations. After 2-3 years, they could become extremely strong by the time they graduated. 印奇 believes DeepSeek’s talent structure resembles this today.
His reasoning is that AI is changing too quickly for deep PhDs in legacy fields to be the scarcest resource. The industry needs “the smartest people,” combining algorithm design, programming, experimentation and tuning. 唐文斌’s informatics-competition background gave Megvii an entry point, while Sogou had used a similar model earlier.
This kind of talent cares about compensation, but also about being surrounded by peers, receiving rapid feedback and having impact through the work. Megvii retained its underlying talent pool during its downturn. Once the platform regained commercial prospects, technical frontier status and resources, people would return: “Talent follows commercial value.”
That is why 印奇 changed the old formula of “technology conviction, value pragmatism” to “value pragmatism, technology conviction.” 唐文斌 put it more sharply: “business model is the best model.”
8. Abandoning the Original Listing Loop Only Made the Loop Longer
程曼祺 asked whether terminating the listing meant abandoning a principle of his life. 印奇 replied: “Actually, I didn’t want it to become a bigger loop, but it had to become a bigger loop.” He had originally hoped to complete a certain loop first and move the company into a period of lower operating pressure.
The common model for the previous generation of AI companies was multiple financing rounds out of proportion to early commercial results, with investors believing that explosive commercial outcomes would appear after crossing a threshold. 印奇 still agrees with the broad logic, but the longer the investment period and the heavier the resources, the greater the risk and competition attached to the final monetization.
“The longer the loop, the longer the chain, and the easier it is for the chain not to connect back.” When he started a company at 23, he did not foresee that the path would be so long. His “revolutionary optimism” at the time was the belief that a group of smart people and a clear vision were enough to get the job done.
He now believes AI is difficult because 3 uncertainties coexist: foundational technology, productization and commercialization; an extremely rapid iteration cycle; and simultaneous competition among the world’s smartest and best-funded people. Relying on a grand mission alone to drive a commercial organization is one correction he has made to his early convictions.
9. ByteDance’s “Close One Loop, Then Attack” Path May Also Be Closer to AGI
Facing ByteDance’s path of completing one loop before starting the next, 印奇 admits it “could” succeed. After a company has operated for 10 years, its founder, core executives, talent density and basic assumptions settle into organizational genes that are difficult to change.
If an organization has taken the previous generation’s business to the extreme, it is highly adapted to the previous paradigm. If the new and old tasks are closely related, it may extend its advantage; if not, transformation is difficult. The saying that “each generation has its own god” leaves the real uncertainty in how long a given version lasts.
If AI 1.0, 2.0 and 3.0 are merely minor versions within the same major cycle, one company may accumulate advantages and win enormous results throughout. If one change is fundamentally a different thing, it is more likely to be completed by another group of people.
10. Qianli’s First Phase Is to Build a “Car BU,” Not Another Ordinary Automaker
印奇 inherited a platform that was “relatively clean but thin at the base.” His first 100 days focused on setting strategy, building the leadership team and organizing the workforce. The overall strategy was defined as “AI+cars,” split into two wheels: terminal products and technology.
The terminal wheel was described as “2 wheels, 4 wheels and 2 legs.” Motorcycles must become electric and intelligent; 4-wheel vehicles should seek categories outside the middle ground of ordinary passenger cars; the 2 legs ultimately point to robots, though they are not the external focus for now.
He believes “China no longer lacks another ordinary automaker.” Entering the already crowded new-energy passenger-car segment would add limited value. For mainstream passenger cars, Qianli would rather act as a Car BU, empowering automakers through cost, performance and standardized delivery.
The technology wheel currently focuses on intelligent driving and intelligent cabins. Megvii contributes its intelligent-driving capabilities, while the Geely system provides cabin resources including Flyme. The core task over the next 2 years is to integrate the 2 into a complete experience that can be exported externally.
11. An Open Supply Chain Is Qianli’s Differentiated Bet Against Vertical Integration
Qianli’s 2 strategic keywords are “open” and “international.” Geely is the most important starting point and foothold, but the solution will not serve Geely alone; international customers will leverage Geely’s existing overseas experience.
Open does not mean simply bringing in external shareholders. 程曼祺 invoked Huawei’s Yinwang and Changan’s investment, and 印奇 responded that the key question is whether the solution and supply chain are open: is everything from chips to systems proprietary, or is the best partner in the industrial chain chosen at each link?
He does not see an absolute winner between openness and vertical integration. When the supply chain is immature, vertical integration is more controllable; during rapid mid-stage expansion, professional division of labor and risk diversification are more effective; at the end, highly homogenized products may prompt reintegration to cut costs. “Long separation must be followed by unity, and long unity by separation.”
Qianli must own its algorithms, overall solution and domain-controller design, while retaining some independent capability in its core-chip matrix. Domain-controller manufacturing and mainstream CMOS are not strategic footholds. Cameras, main chips, core models and the data system that may extend to cloud supercomputing determine the bulk of cost and performance.
12. Independent Governance and a Supplier Brand to Address Qianli’s “Geely Affiliate” Problem
On concerns that other automakers may see Qianli as part of Geely’s internal system, 印奇’s first answer is governance. Qianli has a more independent, third-party governance structure and should draw talent and resources from the entire AI industry, not from a single automaker.
The joint venture established by Geely, Qianli and other partners is “Qianli Intelligent Driving.” Geely’s “Qianli Haohan” will be used across its product lineup, but external output in the future will still be led by Qianli. Putting “Qianli” first in the name is also a choice to build shared supplier recognition.
印奇 uses the smartphone industry as an analogy: Qualcomm chips and Sony sensors add value to a phone brand, while automotive-intelligence supply chains have yet to develop an equally mature division of brand roles. Qianli wants to become a core supplier that guarantees quality rather than weakens the automaker’s brand.
The current debate has 2 layers. Outside Huawei, it remains unproven whether intelligent-driving suppliers can deliver consistently over the long term; if AI’s core capabilities come from outside, where is the automaker’s differentiation? Qianli’s deep partnership with Geely is a joint experiment addressing both questions.
13. 3-4 Deep Customers Are Closer to a Data Flywheel Than Broad Deployment
Qianli will not chase a large customer count, but selectively serve leading automakers. 印奇 expects roughly 3-4 automaker groups to be deeply integrated over the long term, with no more than 1-2 customers of each type to avoid direct conflicts in the customer portfolio.
The choice begins with the auto industry’s large-customer structure. Concentration will continue, while properly serving one automaker requires deep engineering, mass-production and operating investment; too many customers can reduce delivery quality.
More important is data. Without deep binding to an automaker, there is no commercial base and no loop of terminal data feedback, training, validation and redeployment. For intelligent driving and large-model cabins, that loop is the foundation of sustained technical leadership.
印奇 distinguishes between 2 organizations. A delivery-oriented supplier’s core capability is completing projects; a data-driven organization builds models and systems around a continuous loop. If customers do not open their data, the former will struggle to remain technically competitive over the long term.
14. The Real Intelligent-Driving Metric Is “Trust,” Not Takeover Counts
On the “parking spot to parking spot” function, 印奇 first asks whether services such as parking-lot payment are actually integrated. If the underlying services have not closed the loop, the conceptual feature must be broken down into finer product units.
He believes intelligent driving today is still not trustworthy for the broad population. Takeover counts are not a sufficient metric: users have different risk thresholds. More important indicators include willingness to take over, ride comfort and whether the system creates a stable sense of safety.
Tesla’s advantage, in his view, is that its driving is more “human-like.” The closer a vehicle’s judgment and responsiveness in complex conditions are to human intuition, the less ordinary users suspect that the system has lost control. 程曼祺 cited the bizarre performance of FSD after entering China and asked whether it used only internet-type Chinese data; 印奇 did not confirm the premise, instead comparing the difficulty of cross-cultural data adaptation to “having an Indian person learn Chinese by watching videos.”
He still credits Tesla’s fundamentals: a higher proportion of end-to-end modeling and a lower share of white-box rules and controls, like “a smart brain, rather than someone who memorized more exam answers.” Huawei combines white-box and black-box methods for deep optimization. Both paths have strengths; Qianli has clearly chosen the Tesla-like route.
15. Map-Based to Mapless Is a Generational Shift; VLA Is Still a Later Concept
印奇 criticizes automakers for “absolutely overselling” concepts: “They may be doing 1.0 while talking about 2.0 and 3.0.” The shift from map-based to mapless is a fundamental technical leap. After mapless, whether the system is end-to-end or VLA, it is broadly a process of modeling.
Intelligent driving must handle combinations of vehicle models, compute platforms and regions. The ideal foundation model should transfer across platforms and avoid basic errors in different scenarios like a human. Many domestic solutions instead pile on rules and customization before the foundation model is solid enough.
His analogy is: “The ingredients may not be fresh, and then a lot of seasoning is added at the end to cook an end-to-end system.” Rules can improve local experience, but weaken generality and make long-term technical iteration increasingly difficult.
What needs strengthening is the modeling fundamentals, as well as backend data systems, cloud iteration and cross-platform capabilities. VLM provides general-purpose multimodal recognition and “may be necessary”; VLA is a trend, but not the core problem in current intelligent driving.
16. Cars Have a Limited Action Space; VLA Is Not a Necessary Definition for Intelligent Driving
印奇 breaks VLA into Vision, Language and Action: vision handles inputs, Language represents logical judgment, and Action handles outputs. This multi-input, multi-task and multi-actuator framework is more naturally suited to robots with hands and legs.
Driving, by contrast, has extreme requirements for reliability and latency, with a relatively limited action space centered on the steering wheel, accelerator and brake. Under a broad definition, any end-to-end driving model can be called VLA; the name itself does not create a technical breakthrough.
He believes the real gaps today are 2 foundation models: a more reliable, faster large-perception model and a large planning-and-control model that reduces the share of rules. Hardware, software architecture and latency also need to be optimized vertically.
The main theme of the next 1-2 years will not be a full-scale chase after VLA, but making perception, planning and control, data, and hardware coordination scalable in cost and reliability before deciding whether a more unified model structure is worthwhile.
17. Embodied VLA Is Still at the “Toy” Stage; the Brain and Cerebellum Are Not Connected
On embodied intelligence, 印奇 believes the broad direction of VLA “will be,” but scaling laws have not yet been established. Both real-world data and simulation data leave major unresolved questions about whether they can support continuous scaling.
程曼祺 used “going to the airport” as an example: the system must first decide the mode of transport, airport, flight and time, then reduce the task to whether to move the left or right foot. 印奇 believes this reaches the core issue: the full space of thought—from macro planning to meso-level representation to micro actions—has no clear definition.
Many robot demonstrations today operate inside human-defined action spaces and remain “toy-level experiments and preliminary results.” When the representations of both the brain and cerebellum are unclear, it is difficult to connect high-dimensional reasoning with low-level action.
That is why cars remain the clearest and best carrier for robots. “The best robot for mobility services” has not yet been fully solved; if even Tesla has not finished the job, it is too early to extrapolate a commercial explosion in more robots.
18. There Is No Magic in the Technology Path; Execution and Business Model Matter More
印奇 understands technical taste as an instinct built through long accumulation. Technology practitioners develop an intuitive sense of what is more ultimate and elegant, but the judgment is continuous and can be improved through papers, expert exchanges and industry practice into a level of craft sufficient to support commercial decisions.
He warns against deifying technology paths. Information is highly transparent under the current paradigm, and domestic companies are largely learning and following. The key choice may have been whether to take the Waymo or Tesla path 3-5 years ago, not whether a company possesses magic unknown to outsiders.
After choosing a path, 4 capabilities are still required: a strong engineering team, integration of resources and capital, the right technical leader, and basic judgment about what business models can work and what definitely cannot. Most differences come from execution speed and completeness.
The intelligent-driving business model must allow software to iterate continuously. 印奇 favors to B-to-C models such as subscriptions, licenses or one-time purchases: end users pay, automakers and suppliers share the revenue, and suppliers retain the incentive to continue OTA updates. If AI is treated only as a free cost item, the model is unsustainable over the long term.
19. 5 Years in the Trough Was High-Pressure Training for the Final Stretch of the Loop
印奇 uses “fitting the seam” to describe a loop. The closer a board gets to final compression, the greater the pressure. The difficulty does not rise linearly from 80% to 90%, 95% and 99%; the final segment often carries the most information and the fastest growth.
He therefore does not view 2019-2024 simply as a period of silence. Megvii had financed and expanded quickly in its early years, but the AI industry had not accumulated enough and was closer to a physical industry. It needed these 5 years to “let the seedlings take root” and “press down the soil,” making the team more solid.
The listing process pushed the training to an extreme. During the A-share application period, the company could not raise additional financing; cash flow was extremely tight, yet it still had to develop new technology, retain core talent and deal with the capital market. He admits that multiple points came “close to being crushed,” but the company ultimately survived.
20. AI Has Yet to Produce Absolute Profits Commensurate with Investment
印奇 applies a blunt test to the industry: if AI is truly as powerful as advertised, it should at least lift GDP by 10%. But as of the interview, he was unsure whether any pure-AI application globally had created profits on the order of $100M, and asserted that no AI-led application had created $1B in profits.
Revenue is not the same as value. An item that costs $5 and sells for $2 can still generate large revenue; profit better tests whether customers are willing to pay for net value. “In the long run, profit is most tightly linked to value.”
The internet could use low marginal expansion costs and network effects to form de facto monopolies and monetize later. AI is still driven by intensive R&D and has not produced a similar concentration, so the narrative of Amazon losing money for years before harvesting profits cannot simply be copied.
The best example he has heard is Midjourney, whose cumulative profits may be around $100M or slightly higher. Its investment was relatively controllable, the product simple and the timing favorable. But image and video generation depend heavily on underlying language models; once foundational capabilities make a step change, application models and user loyalty may quickly lose their moats.
21. ROI Does Not Weaken Technology Conviction; It Keeps It Alive
The most fundamental change over the 5 years was that the team began using ROI to frame every commercial and R&D project: approximate investment, potential return and time to reach a steady state. The cycle can be long, but it cannot lack a value anchor entirely.
印奇 recalls encountering 2 types of committee members during the listing hearing. One asked why Chinese AI had not built a “SpaceX-style rocket”; another asked why the financial statements showed such severe losses. He believes both sides were right: the company must look up at the stars while keeping its feet on the ground.
His sustainable research model is to feed itself first, then invest a high proportion of the money it earns into long-term research. The value created by the first loop becomes fuel for a second, larger loop, rather than requiring external capital to keep paying for conviction.
Projects should not be judged only by current profit, but by steady-state ROI. Some revenue may not be profitable in the short term but can convert to profit after maturity; other revenue may look attractive today but lack long-term development potential. Neither is the loop he wants.
22. Moving R&D and Marketing from 4:1 Toward 1:1 Made Customer Orientation Real
“When money is tight, you look at how every penny is spent, and your thinking changes.” 印奇 began checking finances, processes and productivity item by item, discovering substantial unnecessary spending and organizational waste built up during periods of easy financing.
Company size also exposed management thresholds. Around 100 people is one type of organization; 400 is a hurdle; 1,000-2,000 is another. When Megvii reached roughly 2,000 employees, the founding team’s operating and management capabilities had not kept pace.
In the past, many businesses had R&D-to-marketing expenses of 3:1 or even 4:1. Over the past 3 years, the ratio gradually moved toward roughly 1:1. Counterintuitively, once the ratio approached balance, the business often began moving toward profitability or healthier operations.
Mechanisms matter more than slogans. When R&D resources are overly abundant, the next-generation product starts before the previous one has sold. Only when the market becomes the locomotive of the value chain, with R&D budgets pulled by customer demand and sales results, does “customer orientation” become more than internal messaging.
23. Management Has No New Continent, Only Old Problems Matched to the Stage
印奇 opposes chasing new management concepts because “people are an eternal subject.” The basic tensions among supply chain, R&D, market, incentives and authority recur; the real difficulty is matching established principles to the company’s strategy and development stage.
Learning from Huawei does not mean copying today’s Huawei. One must know what organizational systems it used at different revenue scales and under different competitive conditions. If a startup directly applies the management system of a company with hundreds of billions in revenue, the method may be correct while the stage is completely mismatched.
Reconstructing a company’s history is almost “mission impossible.” Startups have no time to codify methodology; after maturity, key people leave and memories are beautified by results. Managers can only combine fragments of information into their own systems, without imagining that they have created a theory that never existed before.
This fits 印奇’s underlying worldview: “The world is continuous.” What looks disruptive often reflects failure to see another axis of evolution. Understanding causal continuity makes people more grounded, focusing effort on causes rather than demanding too much from any single outcome.
24. Goals and Performance Are Basic Management Skills Everyone Uses but Few Practice Deeply
In 印奇’s view, the 2 most important management tasks are goal management and performance management. Almost every company has processes for them, but “99% of companies are not doing them seriously.”
The core of goal management is not writing down a number, but deeply examining risks, competitors, resources and assumptions before setting the goal. Many mistakes that look obvious in hindsight were not unavoidable at the time; the decision process simply missed information that already existed.
Performance management is relatively simple for profitable businesses: “create value, share value.” During exploration, however, the organization needs its strongest, most expensive and most innovative people, while most explorations may fail, making investment, feedback and final value difficult to match one to one.
Organizations therefore need both management methods and core leaders capable of executing them. Without people who have experienced the hardest stages and still chosen to stay, theory cannot become real fighting strength.
25. Knowing the Right Thing Does Not Mean You Can Do It; “How” Often Matters More Than “What”
程曼祺 noted that Megvii had already identified software-hardware integration as early as 2016-2017 but failed to fully deliver it because of external factors. 印奇 corrected him directly: “It was not uncontrollable; it was still a capability issue.”
Many teams can reach similar levels of understanding; execution is the scarce resource. Traditional companies repeatedly talk about customer centricity yet still cannot learn Haidilao-style service, showing the huge distance between knowing a principle and implementing it at every point.
In the internet era, the upstream traffic, downstream advertising and middle-product paradigms were relatively fixed, allowing an exceptional strength to close a loop quickly. Physical industries and AI depend more on cross-link coordination. “In the internet era, people thought the what was important; in most business situations, the how may be more important.”
Understanding also changes with business scale. The problems created by 1M, 10M and 100M users are completely different. An organization stuck at 1M will not automatically gain the understanding required for the next stage; understanding must drive the business, while business scale generates new understanding in return.
26. 印奇 No Longer Obsessing Over “the Smartest”; Value and Results Come First
Over the past 5 years, his view of talent has undergone a “fundamental change.” Intelligence remains important, especially in R&D organizations, but it cannot replace execution, collaboration, resilience or customer value.
He now sees his past pursuit of cognitive superiority as partly ego-driven—the desire to see what others could not. After setbacks, he cares more about outcomes and value: “In the end, results matter, value matters, and creating value matters.”
Value is not limited to product revenue. It includes operating the company well, building an effective organization and helping the industry enter its next stage. Being results-oriented means accepting tedious work and making short-term sacrifices for critical outcomes.
Colleagues describe him as harsher and more decisive, and he largely accepts the observation. The change is not that he values talent less, but that he is more willing to make decisions on resource allocation and timing. Truly “super-level” talent should be moved to new directions where possible, rather than lost with a project.
27. Pace Determines When to Build Reserves and When to Sprint at Full Saturation
印奇 has always believed that “pace may matter more than direction.” Many people can see the broad direction; the difference lies in when to launch the sprint and whether resources, team and technology maturity are ready at the same time.
Technical accumulation has a non-compressible timeline. Some fields do not need 1,000 people for 1 year, but do need 100 people working continuously for 3-5 years to build understanding, a talent pipeline and foundational R&D. This requires early positioning with ROI under control.
Once technology approaches the commercialization threshold, the strategy must switch to “saturated resources” and focused attacks. First movers do not necessarily reach the finish line first; later entrants can only catch up by accurately judging the decisive moment.
This does not contradict his earlier view that companies should “maintain the slowest development pace possible when survival is not at stake.” Moving slowly early preserves resources and builds capabilities; moving fast in the decisive battle concentrates execution. “Smart people use clumsy methods”: the essence is doing the fundamentals that cannot be bypassed over the long term.
28. Intelligent Driving Has Entered the Sprint; Three Systems Will Decide the Winner
Megvii had followed intelligent driving since 2017-2018, but believed both the technical path and business model were immature. By 2021, the BEV-based end-to-end framework had at least been established, while Tesla, Li Auto, NIO, XPeng and Huawei pushed the industry out of early exploration and into large-scale mass production. Only then did Megvii decide to increase investment.
印奇 sees “Li Auto building end-to-end last year” as an obvious “sprint signal.” The capital market has never been the objective. As long as sustained investment does not fundamentally interfere with a listing, the company should follow the pace of the industry.
He expects the Tier 1 landscape to converge significantly by “next year” as described in the interview: no more than 4 players, and more likely 3. Qianli’s near-term definition of victory is not monopoly, but becoming one of the few remaining players—“staying at the table.”
Winning depends on 3 systems: a large-model data system driven by terminal feedback; vertical coordination from solutions and algorithms to models and chips; and a customer alliance capable of delivering differentiated value to leading automakers. L3 is not the sole determinant; cost, scale and commercial alliances matter as well.
29. Resource Integration Is a Means; Organizational Resilience Is the Long-Term Asset
印奇 understands Geely’s capital integration as serving an asset-heavy, multi-brand and international automotive strategy rather than pursuing transactions for their own sake. High-quality technology and brand assets are difficult to replicate from zero; open cooperation can accelerate access, but “resource integration must never become the objective.”
Founders with technical backgrounds are more accustomed to quietly building products without overpromoting them, a limitation 印奇 acknowledges. China’s market has long been driven by users, channels and marketing. When product supply is ample, “whoever controls the channel” often has the more certain commercial opportunity. He admires auto founders who have learned Lei Jun-style communication, but only wants to make that capability sufficient for his own needs.
Over the past 1-2 years, the organization moved from “many ideas, with things completed to 70% or 80%” to requiring completion at critical nodes. “Life-or-death battles,” including mass production of Geely intelligent driving on a compute-constrained platform and capital-market processes, trained a stronger ability to execute precisely against a target.
印奇 sees resilience as an equivalent exchange. If the goal is a peak, the required effort and cost must be paid: “If input and output can be one-to-one, that is already the best case.” Pressure must be high enough to train execution, but not so high that it actually crushes the organization.
30. AI’s Calling Began in Adolescence; Entrepreneurship Taught Him to Find the “Real Problem”
Before entering university, 印奇 already saw AI as something worth dedicating his life to. AI at the time was closer to the imagination of robots and was influenced by films such as The Machine Butler. His studies in Tsinghua’s Department of Automation and Yao Class, followed by research into 3D vision and sensors at Columbia University, all centered on machines in the physical world.
He originally planned to pursue a PhD for roughly 3 years before returning to China, because entrepreneurship was “a lifelong thing” and he needed to strengthen his hardware capabilities. Face++ and Facebook’s acquisition of Face.com accelerated the timeline, bringing him back early to form a complementary team with 唐文斌, who leaned toward talent and product, and 杨沐, who leaned toward engineering.
The research values shaped by teachers including 姚期智 and 孙剑 had 2 parts: research must have real impact and cannot be filler; more importantly, one must define the problem first rather than merely solve a problem assigned by others. Choosing the right entry point often requires 2-3 years, or even 3-5 years, of understanding the field.
Research, business and organizations share the same task of finding the “real problem”: research finds the right hole, business finds the customer’s real pain point, and organizations find the people or production relationships behind what is not working. If a judgment cannot survive 3 consecutive “so what” questions, it usually has not reached the root cause.
31. Physical AGI, Art and Luck All Fit into the Same Continuous Worldview
印奇 believes that if the human brain consists only of neurons and electrical signals, without an unknown high-dimensional or “magic” mechanism, it is in principle a simulable computational system. Future AI systems may therefore simulate some human capabilities, but to become partners capable of completing real tasks, they still need physical inputs, outputs and interaction with the environment.
He acknowledges that virtual immersion will continue to accelerate. After robots replace people at work, humans will need more entertainment to kill time. But he believes survival, reproduction and dependence on nature will push humanity to self-correct rather than drift indefinitely into a purely virtual world.
Art has no natural exemption either. Established artistic languages can be structured and turned into craft and engineering, allowing AI to participate. Humans will lose the superiority of believing that creativity cannot be replicated while continuing to “peel the onion” and ask what core remains to be formalized.
He increasingly attributes less of life to accidental luck: “Surviving one battlefield is luck; surviving 1,000 times means there is something more fundamental.” Habits of choice shape long-term luck. More important than favorable or adverse circumstances are growth and “seeing the world, seeing all beings, seeing oneself.”
32. GPT Pushed the 10-Year Deep-Learning Cycle into the Finals
印奇 sees the current AI wave as part of a deep-learning supercycle that began around 2015 and has now entered the “finals.” Earlier battles in vision and speech were more like local campaigns around the brain; GPT was the first to reach a more central and general learning paradigm.
The signal of the finals has 2 components: a general mechanism resembling imitation learning and an expansion path capable of scaling massively. Only after both the learning paradigm and the upper bound of scale were unlocked did the industry enter a concentrated sprint.
A new company may have been founded in 2023, but he believes its core team is unlikely to consist entirely of newcomers. This is an industry built on accumulated knowledge, and the eventual competitors will probably have gone through the technical, organizational and product tests of the previous 10-year cycle.
Existing giants each have distinct chips. Huawei has an upward-extending hardware-software system built around chips and balanced marketing, covering To G, To B, To C and domestic and overseas markets. ByteDance is growing downward from applications toward the cloud and potentially an operating system, with high talent density but limited monetization networks both domestically and overseas. Alibaba spans chips, cloud and models and is strong in open source. Tencent controls WeChat and the content ecosystem, giving it enormous room for an AI strategy.
33. Agents Must Upgrade Information Services into Closed-Loop Services Worth at Least 10x More
The first Super App phase had 2 main visions: new-search-style conversation and Character.AI-style companionship. ChatGPT was the main product to break through. Companionship demands more from emotion, memory and model capability, while early foundation models were not ready.
印奇 gives Super Apps 2 standards: a sufficiently high long-term commercial ceiling to support continuous model iteration, and strong user growth in the short term. General search Chatbots are constrained by the baseline economics of mature search advertising, making it difficult for related revenue to cover model R&D and serving costs.
In the second phase, Agents should not merely provide information, but complete closed-loop tasks in work, creation or life, with users willing to pay for the result. He estimates that compared with “chatting and obtaining information,” such services could create a value-chain market at least an order of magnitude larger.
He divides the opportunity into 4 categories. Work will integrate deeply with existing software such as Office and Feishu; creation combines big-tech advantages with opportunities for startup breakout products; life has the largest potential scale but depends on mobile, operating-system and hardware entry points; emotional companionship is more ultimate and may require a physical carrier. Work, creation and life could see successive breakouts over the next 18-24 months.
34. Super Apps Must Take Control of the Model to Give New Giants a Path Through the Incumbents
Agents require foundation-model capabilities beyond the o1 paradigm, and early products rely heavily on leading models such as Anthropic Claude. Work-oriented wrapper applications can quickly create a sense of value, but their high model dependence and competition from existing office-software entry points make 印奇 skeptical that they are a durable startup category.
Open-source weights solve early access but do not automatically solve continuous training, personalization or adaptation to core scenarios. Fine-tuning alone may also be insufficient. A genuine Super App must optimize performance and cost simultaneously; without deeply modifying the model, it will struggle to win in scenarios where the strongest competitors are also fighting.
“Super model + Super App” must therefore be done together. Startups should concentrate resources on one vertical—work, creation or life—build generational differentiation, and reach consumers directly through a software or hardware outlet. The short-term breakout of products such as DeepSeek and Manus at least shows that “the giants’ blockade is not without gaps.”
印奇 ultimately continues to bet on new companies. The organizational capabilities, product forms and business models required by large models are changing too much for incumbent giants to dominate by default. Entrepreneurs must stay focused and close the loop, and his common-sense judgment remains: “As long as a huge technological transformation is fundamental enough, a new company will definitely win.”