153. 曾鸣: OAI and Anthropic Likely Not Big Winners of the Native Era

Summary

• 曾鸣’s core judgment: OpenAI and Anthropic are “very likely not to become major players or big winners in the native-application phase,” and may not even survive to the end. The basis is industrial history, not sentiment: Yahoo had a $120B market cap in 2000, equivalent to roughly $1T today at an annualized return of about 8%; AOL peaked at $220B, and BlackBerry was the fastest company to break $100B. The first-wave winners of every general-purpose technology cycle were once equally staggering, but “companies from the first phase rarely survive into the second.” His analogy is a refinery and a chemical plant: “You never hear of a car company that was transformed from an oil company.”
• The three-stage view of industrial history puts the present in context: 2026 marks the end of the first stage—infrastructure—and the opening of the second, an explosion of agent applications. Token becoming the standard unit of measurement—Jensen Huang’s “token factory” and Sam’s statement that OpenAI’s core business is selling tokens—signals infrastructure maturity. OpenClaw, or “Lobster,” showed him the structural parallel with the 1992 website-building movement: moving from sharing information to sharing capabilities. The biggest contrarian opportunity is that “we are still waiting for a browser”: a unified Agent standard and an entry point that sharply lowers the barrier to invocation could be “the most exciting thing of the next two years.”
• The end state for model companies is AI cloud, with only one viable business model: oligopoly plus heavy government regulation. In the US, xAI and Meta have “already exited and fallen behind,” primarily because Anthropic began using large models to train large models, creating an intelligent-compounding flywheel that, combined with organizational advantages, is producing a black-hole effect. China is an elimination tournament: “By year-end, any company that cannot reach a 3T model will probably be out.” The window for traditional foundation-model startups has closed. Models will become “a very good mature business”—but not the throne of the next era.
• No incumbent is safe in front of a giant wave, and there are almost no historical examples of companies successfully crossing eras: it is not a question of how many, but whether there are any—basically none. Only IBM and Microsoft can each claim to have done it once, while Microsoft “did not invest enough in models and has effectively exited the model business.” ByteDance has a strong chance of becoming an AI cloud company, but Doubao “definitely is not” that future—“otherwise it would be pouring money into it without regard to cost.” Tencent’s social relationships and Alibaba’s people-and-goods relationships will both “definitely be restructured.” AI is a productivity revolution that overturns the Industrial Revolution, not the “1.5 innovation” of mobile internet.
• The company is an institutional product of the industrial era and will disappear; New Labs are early prototypes of a new organizational form. The basic unit of organization is shifting from the job to the task; hierarchy and middle management will disappear; incentives will come from self-direction, or High Agency, and transparent contribution. “The transfer fees for top researchers are so high because their contributions are so transparent.” Future companies will increasingly resemble partnerships or sports teams. The first New Lab was OpenAI, or perhaps OpenAI plus Anthropic. “How many people do you think are happy going to work at companies today?”
• The most brutal truth for founders is a basic law of economics: “As long as supply is homogeneous, you cannot earn high profits; you are simply not a very valuable company.” That is the source of last week’s market panic over commoditized model competition and the resulting doubts about capex. Technical founders must evolve from “my technology is so powerful” to understanding how the commercial world works. Good is not the same as great: “Greatness always comes from going against consensus,” and is recognized only in hindsight. The common trait of great people is a small ego—they attribute success to luck and to others. Only mission-driven people can reach greatness.
• This year’s application freeze is precisely the contrarian phase for application founders, not the end state. Last year’s Silicon Valley Agent startups were “children playing house,” and were submerged as soon as model capabilities improved. The right posture today is to assume intelligence is sufficient and tackle work 10x more complex, building an intelligence loop of your own rather than scaffolding. Models will not consume everything: Edison lit Manhattan in 1892, washing machines and refrigerators appeared around 1910, air conditioning arrived in 1925, and General Electric ultimately won through consumer appliances, not power generation.
• Strategy is “infinitely important” in the AI era, but the methodology is shifting from strategic planning to strategic generation. “Look 10 years out, think 3 years out, execute 1 year out” derives its creative power from the tension around the 3-year milestones: linear extrapolation “could never produce Anthropic’s 80x year-on-year growth this year.” Projecting ARR 12 months forward from today creates “the illusion of growth”; once competition multiplies, pricing falls from customer value toward marginal cost. Organization comes before strategy: “Only a new organizational form can generate a new strategy.” The renewed popularity of Context not Control is no accident.

Deep dive

1. Three existential decisions in Alibaba’s first 20 years—and all were contrarian

• 曾鸣’s review of Alibaba’s first 15 years, which he experienced firsthand: “The most important decisions were just three: founding Taobao, founding Alipay, and founding Alibaba Cloud.” None was accepted by society, and all triggered intense internal debate. When Taobao was launched, several senior executives strongly opposed it: the core business was not yet mature, and the company was taking on a new core business, with too much risk.
• Strategy does not exist at just one level. Alibaba Cloud’s second-order strategic questions—whether to serve one scenario or multiple scenarios, whether to serve Taobao or search, and whether to be open source or closed source—were equally contentious. Only further down did the company arrive at “de-IOE,” replacing IBM, Oracle, and EMC. “Making one major strategic decision is not enough. The further down you go, the tighter the integration between strategy and tactics.”

2. Good strategy cannot be validated at the time: “You see because you believe”

• Xiaojun asked how anyone knows, at the moment of decision, whether a strategy is good. 曾鸣’s candid answer: “You don’t. You see because you believe.” Every era-defining company is eventually proven right against consensus, but survival bias must be kept in check: “The most successful companies are always contrarian, but being contrarian does not guarantee success.”
• Can a company become great by becoming the strongest player in a consensus market? “That is excellence.” “Is it greatness? Greatness always comes from going against consensus.” This distinction ran through the entire conversation.

3. The three-stage history of general-purpose technology: infrastructure → application explosion → native applications

• 曾鸣 has studied the three Industrial Revolutions as well as the PC and mobile-internet eras, and distilled a common pattern: technology is first accepted by society and becomes infrastructure; then vast numbers of new applications emerge on top of that infrastructure, with many approaches flourishing; finally, people realize that a more fundamental method is needed to run the new world—the native-application phase.
• BlackBerry was the first product widely called a smartphone. In 2010 it still held more than 40% of the US market, versus just over 20% for the iPhone; the iPhone only took off after the iPhone 4. ByteDance first built Neihan Duanzi and Toutiao; Toutiao discovered recommendation algorithms, and once bandwidth and compute matured, short video emerged. Only the combination produced Douyin, a mass-market native application.
• Why can’t companies skip the second stage? “The explosion of applications gives technology its direction for the next stage.” Without the accumulation of the second stage, the third cannot be reached. And the second stage itself “may last 8 years or 10 years—it can be extraordinarily rich.”

4. The coordinates for 2026: tokens close the first stage, OpenClaw opens the second

• The clearest sign that the first stage is largely complete is token becoming the consensus unit of measurement. Jensen Huang said Nvidia’s future is a token factory, and Sam immediately declared that OpenAI’s core business is selling tokens. “Once a technology has a standard unit of measurement, it can be used at scale and standardized”—like a kilowatt-hour of electricity or a ton of water.
• That same Lunar New Year, OpenClaw, or “Lobster,” showed people “the unlimited potential of an Agent that can independently complete tasks,” producing a frenzy so intense that “people could not sleep through the holiday.” His conclusion: AI’s second stage is the Agent stage; “Agents are the new applications of the AI era.” We are at the beginning of stage two.

5. The essence of an Agent is sharing capability; we are still waiting for the “browser”

• 曾鸣 admits his thinking has moved forward. Two or 3 years ago he was still saying, “Let’s embrace Yahoo.” After seeing OpenClaw, his first reaction was: “How have we gone back to the browser stage?” The point is not that a browser has arrived; “we are waiting for a browser to arrive.”
• The structural parallel is clear: “The essence of an Agent is sharing capability”—packaging what it knows so that it can be shared and invoked. The 1992 website-building movement was about sharing information. “The innovation of this AGI cycle is moving us from the information age into the capability age.” Today’s model advances are built on the digitization of information from that earlier era.
• The browser unlocked two things: a unified standard for building websites that sharply lowered the creation barrier, and clicking, which reduced the browsing barrier to zero. That caused supply and demand to explode simultaneously. The methodological lesson is to “understand the underlying operating rules after abstraction and modeling, rather than getting stuck on the technology’s outward form.”

6. The next biggest opportunity: Agent-era Yahoo has not even taken shape

• After websites began to proliferate, Yahoo emerged to build directories and categories. “It was reducing entropy, not increasing it; it lowered everyone’s cognitive burden.” Today’s explosion of Skills is exactly the opposite—entropy increasing. “Ask anyone around you, and no one can casually name 2 or 3 Agents they use every day.”
• He has “seen absolutely no prototype of Yahoo.” His rough judgment is that the next major opportunity is something browser-like: a unified Agent standard and an ultra-simple way for users to invoke Agents, followed by an explosion of Agents. That could be “the most exciting thing of the next 2 years.” The interaction model is impossible to visualize: “We can summarize an abstract rule, but we can never derive a specific form from it.”
• Does this era have a Google? “I’m not sure there has to be a huge gap between Yahoo and Google in our era.” It is too early to tell. The historical footnote is that after Google was founded, it was a junior player providing Yahoo with B2B algorithm services. “In that sense, Yahoo gave birth to Google.”

7. Model companies are AI cloud companies: the refinery and AOL analogies

• Data is the new oil, and “a model is essentially a refinery.” From the access side, it also resembles the vanished AOL—a gateway that lets everyone access intelligence easily. The blunt conclusion: OpenAI, Anthropic, Kimi, and DeepSeek “are the AI cloud companies of the future.”
• This explains the complex bargaining between US cloud providers and model companies. Cloud providers believe they are paying too much and getting too little; model companies are likely to build their own clouds and develop their own chips. The two ecosystems will collide and be reconfigured. “Several relatively stable AI cloud companies will emerge. That is certain.”
• Can model companies become successful application companies? “It is an open question. If I had to give a subjective answer, the probability would not be very high.” Xiaojun cited Codex and Claude Code as counterexamples. 曾鸣 rejected the argument: “That means Coding will become part of foundational intelligence.” Without Coding capability, the token business itself would not work.

8. A trillion-dollar Anthropic is not “different this time”: Yahoo, AOL, and BlackBerry did it too

• On the claim that Anthropic reaching a trillion-dollar valuation in less than a decade is historically rare, 曾鸣 was direct: “That description itself is an illusion and a misunderstanding… This is the classic ‘this time is different.’” He checked the history: Yahoo was founded in 1995 and reached a $120B market cap in 2000, equivalent to a $1T company today at roughly an 8% annual return; AOL peaked at $220B in 2000 and exceeded $300B after acquiring Warner, making it the highest-valued company at the time; Google was founded in 1998 and reached $100B in 2005; BlackBerry was the fastest hardware company to cross $100B.
• The conclusion: “It may very well survive and do well, but it is very likely not to become a major player or big winner in the native-application phase.” Chemical plants later became much more valuable than refineries. “A car cannot run without oil, but an oil company certainly cannot build a car.”
• Xiaojun cited 陆奇’s 2024 statement that OpenAI would definitely become bigger than Google—“the only question is whether it will be 5x or 10x bigger.” 曾鸣 did not yield: “I believe there will definitely be a $10T company in the future, but it may not be either of the 2 current leaders.” He describes himself as a researcher too, “only researching the laws of motion of the commercial world.” His theoretical framework says companies from the first stage rarely survive into the second.

9. Why companies fail at stage transitions: technology changes tracks rather than growing linearly

• The first stage is about whether technology works, driven by original technical innovation. The second begins precisely when first-stage technology slows and becomes callable by everyone. Applications require product originality, empathy for ordinary users, and a new stack of application-side technologies—context management, for example. “Whether you define it as a large model or not is actually not so clear.”
• Some people in New Labs will shift toward the technologies needed for the application stage. Combined with insight, they “may become the kings of the second stage.” Technology will continue to advance, “but it may not grow linearly along its original trajectory.”

10. The end state: oligopoly plus heavy regulation; xAI and Meta have already dropped out

• Public-infrastructure industries have only one viable business model: “oligopoly plus heavy government regulation.” Society depends on them, so they must be regulated; supply is too important to leave to a single provider; and investment is too large to support unfettered competition. “Can you raise water or electricity prices whenever you want? I know how brutal this reality is for many people starting companies today.”
• The end state does not negate the value of current competition. This year, “Anthropic took the lead, OpenAI is catching up, and Google was temporarily left behind—there were gaps in organizational and technical capabilities. xAI and Meta simply exited the race.” In an oligopoly race, some players have already fallen out.

11. Intelligent compounding and the black-hole effect: leaders start using large models to train large models

• The explanation for xAI and Meta falling behind is “intelligent compounding.” Since the second half of last year, Anthropic has made it clear that “they have started using large models to train large models.” AI has entered its own core business, sharply improving training capability. It has not yet reached fully autonomous Self-Learning; “it is only being used locally, but it has already generated enormous value.”
• Organizational advantages combined with new operating methods for technology “have really begun to push weaker players backward.” As OpenAI and Anthropic continue along the Self-Improving path, the gap will widen further. “It is beginning to form something of a black-hole effect, which is also why US cloud providers are starting to panic.”

12. China is an elimination tournament: a 3T cutoff by year-end, and the window is closed

• “The competition among Chinese model companies is an elimination tournament. By year-end, any company that cannot get to a 3T model will probably be out. It is like the World Cup: 8 become 4, then 4 become 2.” Xiaojun reported the scores: Kimi K3 had just reached 2.8T. 曾鸣 said, “Kimi was the first to submit on 3T, and the result is basically satisfactory.” Next comes MiniMax, while Tongyi Qianwen will certainly release its model soon.
• The landscape may not stabilize this year. Only once the top 2 or 3 players have widened their lead over the rest will they become difficult to overtake—the black-hole effect. Should a new company still build a model? “The window for the kind of foundation-model company we define as infrastructure is definitely closed.” The intelligence flywheel is already running; a new entrant would have “no accumulated advantage at all.”
• The shape of the oligopoly is clear: “Homogenization guarantees substitutability; differentiation is what leaves everyone some money to make.” Is a model a good business? “It is a very good mature business.” But in 3 to 5 years, if it has not entered the next stage of competition, it will become a service everyone calls. “People will only complain when you are unstable. They will not praise you every day.”

13. The second-stage question: who will be Netscape, and who will be Yahoo?

• The companies of the next era will be judged on 2 questions: “Who establishes the standard for Agents? Who captures the user entry point for Agents?” That runs against today’s consensus. At Web Summit, application companies occupied only a small section on the basement level; the large companies on the second and third floors were all working on what exists today.
• An Agent entry point could emerge in 3 years because a proliferation of Agents will create trust problems: “Would you hand your wallet to an Agent? Would you hand over your personal privacy? How do you know the fee it charges is legitimate?” A new third-party service provider will be needed to establish trust, set standards, and make recommendations—a classic two-sided market with “a moderate degree of network effects.”

14. The Agent playbook: tackle complex work, not scaffolding

• At a Silicon Valley Demo Day last summer, 3 to 5 of the roughly 10 projects were doing advertising—creative generation and campaign placement. “These are such simple things. I immediately felt they could not have long-term value.” Growth companies looked more promising because growth is complex. “You have to handle complex tasks before you can build your own intelligence loop.”
• The key warning: “When intelligence is insufficient, building an application is largely building scaffolding, and then waiting for the model’s next iteration to drown your scaffolding.” The entry point today must be completely different: assume intelligence is fully sufficient and tackle complex work. “The simpler work is, the more easily it will be swallowed by complex work.”
• Manus had sharp market instincts at the time, catching the window opened by Coding capability and reaching Prosumer productivity demand. But for an Agent to establish itself today, “you need to find an entry point at least 10x larger.” Making presentations or producing reports is not big enough. Education and health are large verticals. In B2B, “organizational brain” Agents will be a major theme this year; in B2C, emotional companionship and problem-solving applications will re-emerge.

15. The data flywheel becomes an intelligence flywheel; the black-hole effect upgrades network effects

• The data flywheel is evolving into “intelligent compounding”: invoke foundational intelligence, give an Agent greater intelligence on a specific task, collect real-world feedback—provided the Agent can work independently without human participation—and combine it with context management, memory, and other algorithmic innovations. The result is compounding intelligence within a scenario that can then penetrate more complex tasks. “The people who find, or get lucky enough to stumble into, the most scalable and generalizable scenario are the most likely to create a black-hole effect.”
• Network effects will remain standard, but the upgraded version is the black-hole effect: “The more application scenarios you have, and the more complex your connections become, the more likely you are to expand the boundaries of your intelligence.” This is the upgraded version of the data-intelligence and network-collaboration double helix described in Intelligent Business. “AI is still data intelligence; it has not entered the symbolic-AI stage.”

16. Native products spread by word of mouth: a firsthand view from UIUC in 1993

• “The best things in the second stage will definitely spread through word of mouth.” Lobster spread on its own, without paid distribution; Yahoo also traveled by word of mouth. 曾鸣 was studying at UIUC in 1993, “possibly one of the first 10,000 people in the world to use a browser. The entire campus went crazy.” Soon someone realized they could build a website and upload content; someone else realized it could make money, and the flywheel began turning.
• Why has OpenClaw’s heat faded? “First, its users are niche—mostly geeks and CEOs. Second, the technology is genuinely immature. The CEOs willing to try it have been tortured beyond belief.” It was a valuable rehearsal, but the product itself is an early-stage geek experiment in the technology-adoption cycle.
• A good B2C Agent has still not appeared, but he is “extremely optimistic; it will definitely appear.” Investor pessimism is simply the result of being hurt in the last cycle and seeing nothing convincing in this one. “Think about it: if there is no good B2C application, why do we need AI?” OpenAI has always wanted to build applications and definitely will. “Whether it can ultimately succeed is another matter.”

17. New platforms and the Agentic OS: judgment replaces information asymmetry

• There will certainly be platforms in the new era: “The entry point is the platform.” The previous generation solved information efficiency. Today, “we can use judgment to replace information asymmetry, but judgment is also scarce.” Judging an Agent is harder and more costly than judging a website, so people will likely delegate that judgment to a trusted third party. Can old platforms evolve? “Probably not.” Old platforms matched simple information; new platforms match capabilities with needs—a new species of matching. Getting there will likely take another 3 to 5 years.
• No one can define the Agentic OS yet. 曾鸣 places it in the third stage, when the OS begins to allocate resources: “You only have to tell it an intent, and I directly call every Agent in the market to finish the work for you. You never even see the Agent.” That requires Agents to proliferate and mature enough to be called simply, so “it may be farther away than we think.” Who will win? “Under my theoretical framework, it will definitely be a company that no one knows what it looks like. That is what major innovation means.”

18. The application trough is the contrarian phase for founders, not the end state

• Agent-application financing has been poor this year. 曾鸣’s diagnosis: “We are now in the contrarian phase for founders in the application field.” No good application has yet shown people the potential or the playbook. AI short dramas and animation have barriers that are too low: “Everyone will have finished them in 6 months. Once model capability catches up, the work is basically done.”
• The industry is between harvests: model capability has spilled into simple tasks that models can now handle on their own, but no one has yet found a sufficiently large scenario in which an Agent can do work that is complex enough and valuable enough, fully use cost-effective intelligence, and build differentiated technical accumulation. “Most people have actually underestimated how hard Agents are.”
• His review of last year is unsparing: it was “children playing house,” with ARR as the only KPI; “the entire Silicon Valley looked like a paradise for startups, and everyone was having fun.” Then a major model-capability breakthrough arrived this year, and “the children who had been playing house were submerged.” The real turning point is when “everyone feels intelligence can genuinely do work and can be trusted, and someone seriously asks whether it can be used for something important. That is the true sign that the application stage has begun.”

19. Models will not consume everything: a century of electricity applications

• “The premise for models to consume everything is that everything lies within the model’s range.” Once intelligence is sufficient, a range of new technologies will be needed for different scenarios. Models cannot swallow everything.
• His favorite intuitive example: Edison’s electric light illuminated one square kilometer of Manhattan in 1892. Around 1910, grids covered the major US cities, and consumer applications beyond lighting—washing machines, refrigerators, and radios—appeared. Air conditioning arrived in 1925 and television in 1930. “The development of applications clearly has its own pattern.”
• The corporate choices are even more revealing. Edison believed power generation was more important and founded General Electric, without recognizing the importance of the grid; his assistant left and founded the first power-transmission company in the US. “General Electric’s real success came from producing the earliest refrigerators and washing machines.” Once consumer applications broke through, the grid became a B2B infrastructure nobody cared to brand, while GE prospered for 60 years. IBM followed the same pattern: founded in 1910 to manage information, it truly took off after computers appeared.

20. Application companies do not need to train traditional models: Cursor became fuel for model iteration

• The concept needs clarifying. If Agents are the startup subjects of the new era, their boundary with models will gradually become clear. Competition among large models will commoditize their capabilities, and “companies building Agents do not really need to develop these large models anymore.” They can call models reliably without worrying that a model company will cut them off, because a reasonable substitute can always be connected.
• Cursor is not the kind of principal entity he has in mind. “It was testing the boundary of model capability when Coding capability was insufficient. As models advanced, it was Fold in, absorbed into the model, and became fuel for model iteration.” Manus sits in the middle: “It has to explore where that boundary is.” These are phase-specific questions; the ecosystem will clarify.

21. Three reasons New Labs emerged—and why they are prototypes of a new organization

• They arise from 3 sources. One group believes the existing path is fundamentally misdirected and has a limited ceiling. Another sees diminishing marginal returns from Scaling Law. The third exists precisely because foundational model capability is now in place, enabling different but equally fundamental innovation outside the current path, such as AI for Science.
• These are not transitional phenomena but “early prototypes of a new organizational form.” They call themselves Labs because they are dissatisfied with the company form. “The real first New Lab was OpenAI plus Anthropic.” The current wave should be called New New Labs. In 5 years there may be a new term—but it may not be called a company, and “it definitely will not be a company.”

22. Companies will disappear: from jobs to tasks, the end of bureaucracy

• The logic is complete. Companies emerged with the Industrial Age; before that there were only workshops. The company was the Industrial Age’s greatest institutional innovation. “If many of the Industrial Age’s basic economic laws are challenged, its organizational form will certainly be challenged as well.” At a deeper level, humans have always been proud of being the only intelligent animals on Earth, with language as a defining marker. “AI can speak too… Things we take for granted could completely disappear 20 years from now.”
• The deepest change is that “jobs may no longer exist. The basic unit of organization in the AI era is the task.” Reporting lines, hierarchy, and compensation are all attached to jobs. Reorient them around the flow of tasks and the organization becomes entirely different. Finding work becomes claiming tasks: “These major Labs are all publishing tasks. If you think you can do one, you take it. If you deliver the result, your work goes live.”
• Why does hierarchy disappear? “Hierarchy exists because no one knows who has the final say… In the future, whoever has the stronger capability gets called.” Once the chain of command is broken, “middle management in the traditional sense will certainly disappear.” 曾鸣 feels no sadness about it: “People are born, age, get sick, and die; companies do the same. Nothing is more normal. That is why new companies have opportunities and young people can have dreams.”

23. Incentives come from self-direction and transparency: High Agency, transfer fees, and team culture

• Without climbing the ladder, where does motivation come from? “At root, it is self-direction.” High Agency has been Silicon Valley’s most popular concept over the past 6 months: look inward for motivation, understand what capabilities to build, find the environment best suited to those capabilities, and grow with it. When people take on tasks and deliver results, everything is transparent. “You will receive the corresponding reward.”
• “Why are the transfer fees for these top researchers so high? Because their contributions are so transparent… They do not even need to write VP or similar titles anymore; everyone knows their standing in the field.” The analogy these people like is a sports team. The star player on an NBA championship team is the highest achievement.
• Is there a bubble in New Labs? “Probably. But bubbles are good things; it takes heavy rewards to attract brave people.” The market has no God’s-eye view, so resources have to be deployed to let people experiment. The valuation anchor is OpenAI and Anthropic: “You still have a company that went from zero to $1T in 10 years.” Investors in the first 2 companies care only about whether revenue can grow from just over 100M to at least 300M in the next 3 years, or preferably 500M. Investors in New Labs are betting that one of them becomes the next OpenAI.

24. Organization before strategy: strategy is generated, not planned

• Why does the book put organization before strategy? “The way the future strategy you want is generated is incompatible with the existing organization. Only a new organizational form can generate a new strategy. The 2 are yin and yang.” The native challenge for a new organization is that “decision frequency keeps rising, while the quality bar keeps rising too.”
• He deliberately coined “strategic generation” to replace strategic planning: “Everyone is used to video generation and image-text generation… Greatness cannot be planned.” The real purpose of organization design is to build a system in which strategic insight can emerge. “In practice, people and AI are training a small model for the organization together.” The ultimate test is whether collective intelligence is evolving.

25. Building an AI-native organization: the CEO rewrites the company’s neural network

• Founders who have gone far enough “have basically rewritten the operating system inside their companies themselves.” The 李志飞 story is illustrative: “I locked myself at home for 3 days and basically wrote the framework.” The dividing line is that most people settle for upgrading their company with Feishu’s AI, because migration is costly and they lack the ability to judge someone else’s system.
• Why must it be rewritten? “That is the neural network of the organization of the future.” The old ERP is “a fixed road, necessarily rigid and limited; it cannot be a real-time responsive network… A neural network naturally corresponds to tasks.” Any strategy discussion must include organization: how many people are in the company, and who participates in strategic discussions?
• Organizations in the second stage will have fewer people. “The AI leverage is large enough, while the demands on people are relatively high.” The first to adapt will be a small group that is highly proactive and has exceptional learning ability. Once they discover methods for human-machine collaboration, those methods can be opened to more people.

26. Everyone becomes a partner: one-person companies are transitional, and scientist-founders will fade

• A key judgment in the book: “The entire enterprise will increasingly resemble a partnership.” Parallel institutions have always existed—law firms, consulting firms, and early investment banks. Their defining feature is that each person’s contribution matters; remove one person and the system does not work. You are hired because your contribution is indispensable. “Otherwise, an Agent would replace you.” The rise of co-founders is a clear feature of AI-native organizations.
• A one-person company is not the direction of travel. “The work one person can do is always limited, so there will certainly be organizations.” But “in the future, everyone will be a one-person company, a six-sided generalist, and then collaborate with other six-sided generalists.”
• Scientist-led startups are a feature of the current phase, not a permanent trend. “In the next stage, the share of scientist founders will fall sharply.” The current generation already shows the shape of what is coming: one co-founder with a large-model background and another focused on applications and business. Product managers will “become enormously important.”

27. The brutal truth: homogeneous supply guarantees that you are not valuable

• Put “a little more sharply,” the hurdle is this: “I have found that many technical people cannot accept a simple fact: my technology is so good, I created so much value, so why might my company be worth very little? Because of a basic law of economics: as long as supply is homogeneous, you cannot earn high profits. You are simply not a very valuable company.”
• This contradicts everything technical people have been trained to believe. “A very small number of technical people must complete this transformation.” As CEO, you have to understand economics and the operating laws of the commercial world to take a company far enough.

28. Good ≠ great: greatness is judged in hindsight, and ego must be small

• Can you identify a great founder at a glance? “No. Greatness is recognized in hindsight. It comes from overcoming negative feedback again and again. When no one believes in you, you ultimately prove through enormous success that you were the era’s earliest prophet. Excellence is a process of steady growth with positive feedback. The 2 paths are completely different.” Many people have crazy ideas; very few can be right and execute.
• Citing Built to Last and Good to Great, he says truly great people “attribute success to everyone else and believe they succeeded because of luck.” Why? “The mission is bigger than the individual. Your ego has to be small enough to go that far… If you are only trying to prove yourself, the world will eventually prove you are not as capable as you think.” People with large egos need short-term feedback, so “you remain in the positive-feedback loop,” following the path that is broadly correct.
• He divides companies into 4 levels: opportunity-driven, strategy-driven, vision-driven, and mission-driven. “But the mission has to come from within. It cannot be something you invent and put on the wall to motivate other people.”

29. The 10-year question: “I just want to build a $10B company” is a failing answer

• 曾鸣’s first question to founders often leaves them stumped: “What kind of person would you be satisfied with 10 years from now?” It best reveals their current expectations and the motivation driving the journey. “I am deeply interested in this, and I hope it succeeds”—the mission is larger than the person. “I am curious” is also more interesting than wanting to prove something. “I just want to build a $10B company” is a failing answer—external metrics are driving the person, so every valuation round can easily shake them.
• The best answer he has heard was Jack Ma’s “make it easy to do business anywhere,” which was “something he genuinely believed in.” Musk’s goal of going to Mars is also a mission: “He wants to take humanity into the era of outer space.”
• The mindset for separating signal from noise is to perceive more broadly and think more deeply. “Many people do not really process things. They follow whatever is fashionable… Someone else is already 2 steps ahead, but you see only the third step above the water. If you chase them, you lose on day one.”

30. Look 10 years out, think 3 years out, execute 1 year out: creativity lies in the middle

• Is strategy important today? “It is enormously important. Make one major decision incorrectly and you are out. This elimination tournament is measured in months. The more uncertain the environment, the more strategic thinking matters.” The historical counterexample is the Industrial Age, when growth was linear and strategy became so unimportant that “large companies outsourced strategy to McKinsey.” The explosion of McKinsey and BCG from the 1980s to 2000 was a product of that environment.
• Linear extrapolation no longer works: “Linear extrapolation could never produce Anthropic’s 80x year-on-year revenue growth this year.” Looking 10 years out does not help with the immediate move. “The real creativity lies in looking 3 years out”—thinking through 2 or 3 milestones and creating dynamic tension across short, medium, and long term. “It is like finding the optimum in 3-dimensional space, which is entirely different from optimizing on a 2-dimensional plane.” The cadence can change: look 5 years out, think 2 years out, execute 6 months; or look 3 years out, think 1 year out, execute 1 quarter.
• One example: an Agent company with rapidly rising ARR came to ask for advice last year. He asked only one question: “What happens after your ARR keeps rising for 2 more years? How quickly will the ceiling arrive?” The founders ran the numbers, found the ceiling was low, and immediately started something new. “Extrapolating 12 months from today easily creates the illusion of growth… Once there are more competitors, pricing is based on your marginal cost.”

31. Context not Control: alignment, OKRs, and the “old-timer” phenomenon

• Implementing OKRs demands too much from most organizations. They either abandon them or turn them into distorted KPIs. The AI era may finally make real OKRs possible: an organizational brain can track natural-language expressions and “will probably be able to tell us that we have all deviated from our OKRs.”
• “Why has Context not Control suddenly become popular again?” Five years ago, people heard the Netflix and ByteDance examples and thought they were impressive but irrelevant to them. “Today everyone knows that if you give enough Context, people can make the right decision.” Align goals, align understanding, and give everyone the full context. Humans cannot process massive amounts of information, “but they have instantaneous creativity and the ability to see connections that do not exist.”
• The “old-timer” phenomenon has a clear mechanism. People used to respect age because the experience of older people had value. “Experience that can be converted into knowledge is no longer as valuable, because large models have absorbed it. Smart young people who use AI leverage well can instantly become experts in any field.” What should older people do? “Either retire comfortably or become young again.” The new culture is open, transparent, shared, and collaborative. Old companies are unhappy because work is depleting and management is controlling. In the future, “happiness becomes a necessity, and emotional value becomes a necessity.” The CEO becomes a service function: “You cannot PUA people anymore.”

32. Stage-specific strategy and the capex inflection point

• Divide companies by development stage, not industry. Large-model companies have broadly entered the strategic-development stage, with strategy relatively converged: pursue organizational efficiency, execution efficiency, and an operating model that can close the loop; the core is scale. In the exploration stage, such as embodied AI, “the efficiency of innovation must be high, not the efficiency of execution.” The top leader must lead experiments and reduce the cost of failure, asking, “Can you let reality take precedence over your ego?” There are always exits outside consensus: “People who reject the consensus go start companies. One contrarian idea in a New Lab may become the new consensus in 3 years.”
• He saw through last week’s panic. Homogeneous competition “broke the expectation that large-model companies could enjoy high monopoly profits over the long term.” Once that expectation broke, people began questioning capex sustainability: “Infrastructure exists to serve model companies. If they create limited value, why should I invest in all this?”
• His view is that the growth runway for large models remains very long and infrastructure will probably continue. “But at some point, infrastructure investment may, because of cyclicality, exceed real demand from large models.” That would not necessarily be bad for the industry; it would sharply lower the entry barrier for application companies. But application companies cannot “hang around and wait”: “If you are not in this game, you cannot accumulate the relevant experience. When the real wave arrives, you will not be able to catch it.”

33. Robots: both the south and north slopes work; appliances are the better analogy than cars

• Embodied AI is still in a “very early” period of strategic exploration, with data as the core constraint. There are 2 paths: build the brain first and pursue generalization, or start from a scenario and use a closed loop to feed the model. “Logically, both paths work… It is like climbing a mountain: you can climb from the south slope or the north slope, and it is difficult to say in advance which will be faster.” The industry sees itself as being around 2020, before ChatGPT.
• The history of automobiles is the analogy. From 1900 to 1920, the US had thousands of car companies. In 1913, Ford built the world’s first moving assembly line, making cars affordable for mass consumers—the first B2C breakthrough. Ford became the largest US company by selling only black cars; once demand diversified, General Motors won with 7 or 8 models. “A strategy is effective only if it works in the period when it is applied.” “Whoever is first to genuinely sell 10,000 robots will already have achieved something remarkable.”
• Xiaojun prompted a better analogy: “Robots may be better compared with the development of appliances.” Electricity can power N types of appliances: companionship and housework at home, and widespread applications in industry and other external settings. “It is a basic interface through which AI interacts with the physical world, a bit like electricity. Everything can be done again with electricity.”

34. No incumbent is safe in front of a giant wave: cross-era precedents are “basically nonexistent”

• “I have never believed anyone is safe in front of a giant wave. Whoever feels safe is the least safe.” Going all-in on AI is only a necessary condition. Are there many companies that crossed into the next era and still performed exceptionally well? “It is not a question of how many, but whether there are any. Basically none.” IBM did it once. Microsoft counts once too: it missed mobile internet but kept a foothold in search, accumulated cloud computing, and retained its PC OS monopoly, giving it one more chance in AI. But “it did not invest enough in models; it has effectively exited the model business… It can no longer afford to buy OpenAI.”
• Google “may become a very good AI cloud company.” It has TPU and large models, and can survive as a cloud provider. “But whether it can become a major consumer application or an entry-point application is genuinely uncertain.” Is that depressing? “Not at all. If one company had monopolized everything until today, the world would be too boring.”

35. ByteDance, Tencent, Alibaba: Doubao definitely is not that future

• “Will ByteDance be the ByteDance of the next era? Likewise, probably not.” It has a strong chance of becoming an AI cloud company; its video generation and multimodal capabilities are world-class. “Building infrastructure is easy because there is accumulated advantage and inertia. But the major B2C application has enormous uncertainty.” What about Doubao? “Doubao definitely is not it. Otherwise it would be pouring money into it without regard to cost. The fact that it is not pouring money in means Doubao is not future-facing enough. Doubao is a first-generation AI product, the chatbot—it has no agency.”
• Xiaojun asked whether it would be viable to embed Agents into Doubao, then offered the following mechanism as a framing: everyone wants to build the OS of the AI era, but if the previous generation’s entry point is not the factual standard for Agents, it will struggle to become the next generation’s entry point; services from the previous generation can be disrupted by services from the next. This was Xiaojun’s framing and follow-up judgment, not a standalone conclusion from 曾鸣.
• Tencent and Alibaba face the same challenge. “Social relationships will definitely be restructured. Why has Tencent, supposedly the safest company, become so anxious recently? Because it may not be safe.” QQ did not truly carry mobile internet forward; or, more precisely, WeChat did. “The relationship between people and goods will be restructured in the same way. Alibaba faces that challenge too.” Existing products plus AI plus Agents? “I think all of them are transitional products.”

36. AI is disruptive innovation, not mobile internet’s “1.5”

• He has long considered whether mobile internet disrupted or extended the PC internet, and concluded it was “1.5 innovation.” Continuity—the internet’s basic architecture and network effects—allowed leading companies to transition successfully, while the disruptive part ultimately produced native players such as Douyin and Pinduoduo. “But I increasingly see AI as a disruptive technology. It is not disrupting the internet; it is disrupting the Industrial Revolution. It is a productivity revolution, while the internet was essentially a revolution in production relations. The continuity is much weaker.”
• That is why “hoping the next-generation product will also emerge inside my company is very difficult.” A 100,000-person company transforming into an AI-native organization faces a different challenge from a 10-person company growing into a 1,000-person AI-native one. And “if you have not built the first truly successful Agent, you probably do not understand how Agents work. You cannot become the factual standard or the entry point.” A successful Agent must come first; only then can the black-hole effect emerge.

37. Organization determines success: xAI’s exhaustion, Meta’s bind, and the Sam-Ilya combination

• Firsthand observation from xAI: “Whenever I encounter a group of engineers in Silicon Valley, the xAI person is always the most exhausted one… This is no longer ordinary overwork; he has been consumed. When people are too tired, there is no innovation.” The deeper issue is that researchers do not feel Elon truly understands large models. His hardware-oriented first-principles methodology cannot produce the answer, and he cannot use a co-creative process to let the organization produce it either. “xAI is definitely stuck on organization. It has not used organization to generate strategy; it is all his own will.”
• Meta faces a different version of the same problem: “You are young, but you entered the game early, so you were depleted early.” It is difficult to allocate resources internally, and most researchers do not think Zuckerberg understands models, so “he cannot really make the most important decisions.” The contrast is Sam, who genuinely relies on Ilya, a person capable of driving the entire effort, while also being a strong CEO. Anthropic’s organizational capability is exceptional, and Dario has a clear grip on mission and vision. “Both companies are fundamentally startups of the new era, with cultures that fit better.”
• China has parallels. Kimi may be more native, but because it started early and has the resources of a large company, it has continued moving forward. Once 张一鸣 realized he had to go all-in and personally enter the field, ByteDance also moved quickly. The structural reason China converges more slowly is that early movers did not enjoy the 2 or 3 years of lead time available in the US; they were fast followers of US prototypes, and the starting gun inevitably triggered a hundred-company war, with everyone recruiting researchers from Silicon Valley. Even so, 杨植麟, 闫俊杰, and Zhipu had deeper understanding and were 6 to 12 months early, so “they are clearly leading this round.” Xiaohongshu and miHoYo are not competing to become API factories by building models; they are “reserving a foundational capability for the future.” “That chessboard is still in the opening.”

38. Timeline: no short-term forecasts; native mega-applications are 8 to 10 years away

• “I never make short-term forecasts. I only offer an understanding of general patterns.” Within 5 years, “an Agent entry point could certainly emerge, and the second stage will probably be clear to everyone.” Native applications on the scale of Douyin or Pinduoduo belong to the third stage, “perhaps 8 or 10 years from now.” Xiaojun sighed at the length of the timeline. 曾鸣 reminded him: “Don’t forget, OpenAI is already 11 years old.” Society will need “a generation—at least 20 years—to digest this technological revolution, because that generation will be the natives.”

39. The age of creativity: creating from nothing and the education system’s collision

• Drawing on Drucker’s framework of a productivity revolution followed by a management revolution and then a knowledge revolution, he defines the AI era as the age of creativity. “All knowledge work that can be structured will be taken over by AI.” Humans will be pushed—and freed—to develop their potential. Creativity needs to be redefined: “It is not simply being able to write a song or draw a picture. It is more the ability to define complex problems originally.” Just as people in the 19th century could not have imagined industrial workers or white-collar workers. Xiaojun’s summary—AI does everything that already exists; humans must create from nothing—won his agreement.
• Education will be the next major conflict. “Education used to be fundamentally about transmitting knowledge… From the earliest moment, all of us were pinned to the fixed patterns of the old world so that we would not lose at the starting line.” Most high-school students probably know that graduating and finding a job through the current system will not work, while young parents know a new method is needed but do not know what it is. Business schools? “They are already failing. They are all in decline.” The share of the virtual world will continue to rise. “The direction of the Metaverse was not wrong; it may simply have been 5 years early.”
• The mindset has to reverse. “We used to favor reducing uncertainty, but this phase requires embracing uncertainty. Uncertainty is the biggest opportunity… We have been worrying unnecessarily. We have worried too much on behalf of young founders.”

40. Happy researchers: mission, reading list, and ByteDance’s destiny

• 曾鸣 describes his mission as “being a happy researcher.” Will AI make people happier? The philosophical answer is unclear, but “the progress of civilization is technology constantly creating new possibilities. If people master new technology, they enter a new stage of civilizational development.” The Industrial Age solved subsistence; the consumer age solved insufficient consumption; the internet solved information asymmetry. Whether an organization makes people happier will become part of how it is judged. Google in 2007 and 2008 “was like a god in Silicon Valley,” offering an early glimpse that the new era needed a new culture and a new organizational model.
• A rapid-fire archive: he does not read AI technical papers; the most important current BET is “the emergence of a sufficiently useful B2C Agent”; on hearing the name “Language Is the World,” he said, “You may need to change it to Agent Is the World”; the books of his life are the Built to Last series, Drucker, and a book by Friedman on the basic operating mechanisms of markets that he copied out almost in full as an undergraduate.
• The conversation ends with ByteDance’s destiny. 张一鸣 is “at least vision-driven,” was among the first to propose company-level products, and was among the first to adopt Netflix culture. “But ByteDance ultimately became an efficiency factory… Its AI strategy requires 张一鸣 to come back into the arena and redefine it.” The broader rule is universal: “Every company ultimately pursues efficiency… The destiny of an organization is that the era will inevitably change. Either you transform completely, or the era eliminates you. It is fair to everyone—the new era.”