Zeng Ming on AI’s New Era: Why Great Companies and New Organizations Emerge

Zhang Xiaojun introduces Zeng Ming as a scholar of corporate strategy who observed the internet's full cycle. Zeng previously served as Alibaba Group's Chief of Staff and helped establish Cheung Kong Graduate School of Business and Hupan University. Zhang says they will examine AI's new industrial paradigm through history and Zeng's unconventional views; Zeng opens by saying hierarchical companies will eventually disappear.

Zeng says there is no need to mourn the decline of companies: people are not happy in many current workplaces, and homogeneous supply makes high profits difficult. He doubts Google or ByteDance will necessarily lead the next era, even though ByteDance's models are strong; its consumer AI application remains uncertain.

Zeng says Tencent and Alibaba are also vulnerable as AI reshapes relationships between people and between people and goods. Asked whether today's giants can become era-defining companies by going all in on AI, he says the challenge is enormous, but transitions naturally create openings for new companies and young founders.

Asked why founders say his teaching influenced their strategic decisions, Zeng credits both luck and preparation. He joined Alibaba during the internet's rapid growth, and his academic training helped him connect theory with practical decisions and turn experience into frameworks.

Zeng keeps studying industry developments so he can discuss both strategic methods and what may happen next. Looking back at Alibaba's first 15 years, he names three consequential choices: founding Taobao, Alipay, and Alibaba Cloud, all of which drew intense internal and external controversy.

Taobao's launch was risky because Alibaba's original business was not mature and the new venture would challenge eBay. Zeng adds that strategy creates layers of follow-on decisions, such as whether Alibaba Cloud should support one workload or many, serve Taobao or search, and open-source its systems.

Alibaba also debated building its own database and replacing IBM, Oracle, and EMC through de-IOE. Zeng says strategy and tactics become intertwined in execution; the company did not know in advance whether a controversial choice would work, but acted on what it believed it could see.

Zeng warns against survivorship bias: winners often pursued contrarian views, but being contrarian does not guarantee success. He defines strategy as creating the future, not merely planning for an expected one. Excellence can come from doing a consensus strategy better, while greatness requires a contrarian path.

Zeng says the AI future is confusing because it is hard to distinguish small ripples from waves that will reshape industries. He had been thinking about this for years, and by the first half of the year, signals began to converge into a framework that helped him see the stage AI industrialization had reached.

Zeng explains that his book Intelligence emerged from reflections on how to make strategy amid rapid change. Once he understood the stages of AI industrialization and their different operating rules, he felt the framework could guide decisions over the next three to five years.

Drawing on the histories of mobile and PC internet, cloud computing, and earlier industrial revolutions, Zeng says transformative general-purpose technologies tend to pass through three stages. First, the technology becomes socially accepted and infrastructure; then applications emerge; finally, native applications reshape the new environment.

Zeng illustrates the transition with mobile internet. BlackBerry was an early smartphone leader, but the iPhone and App Store enabled a new generation of native applications. The iPhone only pulled ahead after the iPhone 4, while the App Store helped native apps emerge.

Early mobile apps such as Moji Weather could quickly reach more than 100 million daily users without representing the future. Zeng says the explosion of applications helps reveal which products meet core needs, as Toutiao and other early ByteDance products explored how to use mobile internet.

ByteDance's recommendation algorithm combined with better bandwidth and computing to make short video possible, and Douyin became a native application. Zeng calls the application boom a valuable middle stage rather than a mere transition, and believes AI could spend eight or ten exciting years in it.

Zeng says AI is at the opening of its second stage, with 2026 marking the symbolic end of the first. Token became a shared unit of measurement, as Jensen Huang called NVIDIA a token factory and Sam Altman described selling tokens as OpenAI's core business.

When users can relate token use to completed work and cost, AI becomes a standardized service society can use at scale. Zeng says infrastructure will keep improving in intelligence and cost. OpenClaw then excited people by showing agents that could complete tasks independently.

Zeng identifies agents as the new applications of the AI era and says the industry is just entering that stage. Asked for an internet analogy, he recalls his earlier prediction that Yahoo was coming, but says OpenClaw made him think the field is still waiting for a browser-like interface.

An agent packages and shares capabilities, Zeng says, much as a website once packaged and shared information. He compares the current shift from an information age to a capability age with the early web, when building a website let people publish information for anyone to access.

Zeng says today's large models can answer questions because the web digitized information decades ago. The useful historical analogy is structural, not literal: a browser established a simple standard for creating websites and made browsing as easy as clicking, sharply lowering barriers for creators and users.

As websites multiplied, people needed help finding what mattered. Yahoo began as a directory that selected and categorized sites, reducing users' cognitive load. Zeng sees a similar need for standards and discovery as agents and skills proliferate, before a new “Yahoo” for agents appears.

Zeng has not seen a Yahoo-like guide for agents yet. He thinks a browser-like product could standardize how agents are called and help users discover them, setting off a major wave of applications in the next couple of years.

Zeng says the broad evolutionary pattern is easier to see than the eventual product form. It is too early to know whether a new “Yahoo” will be followed by a “Google” with a similarly large advantage; he wants to see the new intermediary emerge first.

Zeng compares models to oil refineries: data is the raw material, and models process it into usable intelligence. He also compares model companies to AOL, which helped people get online, and says today's models provide a broad entry point to intelligence.

Zeng calls OpenAI, Anthropic, Kimi, and DeepSeek AI cloud companies. Their relationship with traditional cloud providers is becoming complicated: model firms may build their own cloud services and chips, while cloud providers question how much value they are getting from their investments.

Zeng is unsure whether a foundation-model company can become a successful application company. Coding is becoming part of foundational intelligence, he says, rather than simply an application built on top. He also cautions against assuming “this time is different” because Anthropic has grown so quickly.

Yahoo reached a market capitalization of $120 billion by 2000, only five years after its founding. Zeng estimates that $100 billion then would be roughly comparable to a trillion dollars today; AOL later reached an even higher valuation, and Google crossed $100 billion in 2005.

BlackBerry also reached a $100 billion valuation unusually quickly. Zeng says early winners can become enormously valuable, even if they do not lead the native-application era. He compares refining with chemicals: applications built on processed resources can eventually be more valuable than the basic suppliers.

An oil company is essential to cars, Zeng says, but that does not make an oil company a carmaker. He discusses Qi Lu's prediction that OpenAI could become many times larger than Google, then clarifies that he doubts today's leaders will necessarily define the next era.

Zeng describes himself as a researcher of business history, studying how companies move through technological stages. First-stage firms often struggle to survive the second stage, because making a technology usable relies on different work than building original applications that understand ordinary users.

The application stage requires new technical capabilities, including context management. Zeng says these advances may be substantial enough to count as the large models of a new era. New “neo labs” may develop the technology applications need, and progress may shift to a new paradigm.

Zeng expects new organizational forms to combine with technical advances, and says it is too early to know how the model-company landscape will settle. For infrastructure essential to society, he expects a limited number of suppliers, heavy investment, oligopoly, and strong regulation to ensure reliable, low-cost service.

Zeng says the eventual market structure does not make today's competition unimportant. Zhang asks whether OpenAI, Anthropic, and Google can coexist and differ in quality; Zeng says competition matters even if infrastructure later becomes a regulated oligopoly. Anthropic, OpenAI, and Google have recently taken turns leading.

Zeng connects Anthropic's lead to “intelligence compounding”: using AI tools in the core process of training AI. Large models have begun helping train other large models, improving development speed, though this is not yet fully autonomous self-learning. Even partial use has brought substantial gains.

Zeng says organizational focus combined with new technical methods can create a powerful lead. He expects Chinese model companies to keep catching up, though the U.S. and China remain different ecosystems. He predicts an elimination race around training models at the three-trillion-parameter scale.

Zeng describes the Chinese model race as moving from eight competitors toward four and then two. He names Zhipu, Kimi, MiniMax, and Qwen among the companies to watch, while saying it is too early to know whether the top two or three will separate decisively.

Zeng says Anthropic and OpenAI are developing an intelligence flywheel that could widen their lead—a black-hole effect that also makes U.S. cloud providers uneasy. For foundation-model infrastructure, he thinks the startup window is closing because leading companies already have accumulated resources and compounding learning.

The eventual model providers may be differentiated enough to earn money while remaining interchangeable. Zeng sees models becoming a good but mature service: customers may notice them mainly when they fail. He expects the next generation of companies to create Agent standards and user entry points.

At the 2026 WAIC conference, Zeng sees application builders in a small area while established companies fill the main floors. He predicts an Agent entry point could emerge within three years, but users will need help judging whether an Agent is capable, safe, and trustworthy with money or private information.

Users may trust an intermediary's standards and recommendations before trusting individual Agents. Zeng says the ecosystem is crowded with experiments; Silicon Valley demo days last summer featured many Agent startups, but most built simple applications that have struggled to show lasting value.

Many early demos generated advertising materials or placed ads. Zeng says such simple tasks have little room to grow, while complex work such as growth can produce a feedback loop and valuable technical accumulation. Founders should build for sufficient intelligence rather than scaffolding that models can absorb.

Zeng says an Agent needs a complex enough job to build on; simple tasks are easily absorbed into broader products. Manus had good timing when coding capabilities spilled over, but a general-purpose Agent now needs an entry point many times larger than a narrow task.

A presentation or report tool alone may not be a large enough entry point. Zeng says founders should look for substantial verticals such as education or healthcare. Business products may aim to become an organizational brain or operating system; consumer products may focus on problem-solving or companionship.

Asked whether Agents can build a data flywheel, Zeng says the idea has evolved into an intelligence flywheel. On both business and consumer sides, the product must do meaningful work. Agent applications also need algorithms and scenarios that let intelligence improve through use.

An Agent can develop a data flywheel by completing a specific task independently and receiving real-world feedback. Improvements to context management, memory, and interaction can compound into greater intelligence, which may then extend into other scenarios and more complex work.

Zeng thinks the most scalable and generalizable scenario could create a “black-hole effect,” but says he has not yet seen such a strong native application. A breakthrough product should spread by word of mouth, as OpenClaw and Yahoo did; he recalls the browser craze at UIUC in 1993.

Zeng remembers the campus quickly asking why there was so little to see online, which led people to build websites and discover business opportunities. Asked whether OpenClaw feels like that moment, he says network effects will persist, but their AI-era form will be broader.

More diverse data can generate new insights, while more application scenarios can expand a system's intelligence. Zeng calls this an upgraded network effect, combining data intelligence with network coordination into a “black hole” that can absorb more scenarios.

Zeng expects platforms to remain important because judgment is scarce even when information is abundant. A two-sided market may match Agent developers with users, but users need more assurance than when deciding whether to open a webpage because a poor Agent can cause serious consequences.

Users may delegate judgments about Agent capability to a trusted independent third party. Unlike old platforms that match information, new platforms must match capabilities to needs. Zeng expects Agent-native companies to emerge, though a full Agentic OS may be three to five years away.

Zeng places an Agentic OS in the third stage. It would take a user's intention and coordinate mature Agents without exposing each one individually. That requires a sufficiently rich and reliable Agent ecosystem, so he thinks this interface is further away than many people assume.

The eventual Agent platform could come from a model company, an application business, an internet company, a compute provider, or a hardware-software company such as Apple. Zeng says genuine innovation often comes from a company whose future form is not obvious at the beginning.

Zeng says discussing the third stage can guide work in the second. Agent startups have struggled because investors ask whether their products are merely model features. He calls this a contrarian period for founders: no breakthrough application has yet demonstrated the field's full potential.

Video generation offers one example: many companies want to build the AI-era equivalent of Douyin, but today's products are far from that goal. AI short dramas and animation had a low barrier to entry, so teams quickly converged on similar offerings as model capabilities improved.

Zeng says founders underestimated how hard it is to build Agents. Earlier products could show ARR from simple tasks, but those features were absorbed as models advanced. What remains difficult is finding complex, valuable work where an Agent uses intelligence fully and develops a distinctive technical base.

The real sign of the application stage, Zeng says, is when people trust AI to do important work. He rejects the idea that models will devour every application: once intelligence becomes sufficient, new technical requirements will emerge for different scenarios. He illustrates this with the long development of electrical applications.

Electric grids and household applications came after the light bulb. Washing machines and refrigerators appeared around 1910, radios also emerged, and air conditioners and televisions arrived later. Zeng says applications follow their own timelines and may eventually be compressed into newer devices such as phones.

Edison focused on power generation and founded General Electric, but an assistant left to build a power-transmission company. GE later prospered through products such as refrigerators and washing machines. Zeng says consumer breakthroughs helped make the grid essential infrastructure, while the grid brand itself became less visible.

IBM grew alongside the need to process information in modern factories, then expanded after computers appeared. Zeng says AI-era leaders will be those who use intelligence to build valuable hardware or software products. He frames Agents as the main actors of the new era, distinct from model companies.

Zeng says the boundary between Agents and models will become clearer. Agent companies may develop important technology, but large models will increasingly be commoditized infrastructure; an Agent startup generally does not need to train a giant model. Cursor first probed the limits of coding models, then those capabilities were absorbed into the models.

Zeng says these are different stages that were hard to distinguish while the field was changing dynamically. Agent companies can call foundation models without training their own, provided there are reliable substitutes. Manus still has to explore where that boundary lies and what model investments its Agent needs.

Zeng says OpenClaw was a useful rehearsal, even though it mainly attracted technically adventurous users and was still immature. He has not yet seen a strong consumer Agent, but remains optimistic that one will emerge; current investor pessimism follows disappointment with the previous wave.

Zeng thinks a few model companies could build major applications, but success will depend on their organizational and innovation capabilities. He advises struggling application founders to find their own direction and execute well. He says fundraising is hard when the work is genuinely new.

Zeng says new labs are appearing in both China and the U.S. because some researchers see a ceiling in the current paradigm, including diminishing returns to scaling, and want to test ideas they have long held. Others want to explore foundational directions such as AI for science.

Zeng adds that large models have reached a level where teams can try fundamental alternatives to the current technical path. He sees these new labs as early forms of a different organization that will continue to evolve, not just a temporary phenomenon.

Zeng says companies arose with the Industrial Revolution, replacing craft workshops with an organization suited to modern management and large-scale expansion. If AI challenges the industrial era's economic laws, it may also challenge the company as an organizational form.

A “neo lab” is closer to a group of researchers collaborating than to a conventional company, Zeng says. OpenAI and Anthropic were early versions, but their successors are exploring different work. The future organization may not be called a lab or a company; its form is still undefined.

Zeng expects organizations to continue, because people will always need ways to work together, but their forms change historically. AI may bring a transformation as deep as the Industrial Revolution, overturning industrial-era institutions, while also challenging humanity's long-held identity as the only intelligent, language-using species.

Zeng says AI can understand human knowledge and display intelligence, making the future difficult to imagine; many familiar things could disappear within twenty years. He identifies a foundational change in organizations: the basic unit may shift from a job position to a task.

Today's companies define positions, job descriptions, reporting lines, hierarchy, and compensation around roles. In an AI-native organization, Zeng says, people and AI would instead organize around tasks. Adding AI to existing human-designed workflows is only an interim step; a native organization would start by defining and decomposing tasks.

In a task-based organization, people could claim work, deliver results, and have their contributions become visible. Zeng expects a more network-like structure with fewer layers. Hierarchies exist partly to decide who has the final say, but future work can call on whoever is most capable.

If command chains disappear, the traditional middle layer and promotion ladder may disappear too. Zeng says motivation would come from self-direction, which Silicon Valley calls “high agency”: people choose what capabilities to build, assess their own growth, and move when they no longer find meaningful work.

Zeng says people need to identify what they want to learn and where those capabilities can be used. Task work also makes contributions and rewards more transparent. He notes that researchers are increasingly recognized by their field-level contribution rather than by titles such as VP.

Zeng thinks the neo lab is an early form of a future organization. Labs appeal to researchers, but eventually all creative people may work in new kinds of organizations that are neither universities nor companies. He rejects the idea that the one-person company will be the final form.

A one-person company is not the final form, Zeng says: AI can let an individual do work that once required many people, but individual capacity remains limited. Organizations and collaboration will continue; people may become more versatile contributors within them.

Zeng says traditional companies need to understand the chasm to AI-native organizations; simply layering AI onto existing practices may leave them behind. Today's neo labs are technology-oriented exploratory forms, but future organizations will also explore difficult questions beyond technology alone.

Zeng says people may eventually describe the ideal organization as a team, like a professional sports club, rather than a company. He thinks a bubble around neo labs can be useful: high potential returns encourage experimentation, and market feedback helps allocate resources.

High valuations reflect the possibility that one new lab could become the next OpenAI or Anthropic. Investors in today's leaders also need continued rapid growth to justify their bets. Zhang asks why Zeng puts organizational questions before strategy in his book.

Zeng says AI is a paradigm shift for the industrial era, so new organizational forms and strategy must be developed together. The desired strategy may not fit the existing company; a new form may be needed to generate it. Decision-making is also happening more frequently and under higher quality demands.

Zeng says an AI-native company needs leaders who understand AI deeply enough to redesign its internal operating system. He cites Li Zhifei's account of spending three days at home writing a new framework with AI's help: leaders who understand organizations can rapidly turn a working prototype into reality.

Other CEOs may settle for adding AI to Feishu because migration is costly and they cannot judge alternatives. Zeng sees this as a dividing line: the earliest AI-native leaders may personally rewrite how their organizations operate, while others keep the existing system.

Zeng compares old ERP systems to fixed roads: they impose rigid workflows, unlike a neural network that adapts to tasks. For CEOs, he says, competing at the technological frontier now requires a greater commitment to AI-native work; organization and strategy are inseparable.

Zeng says strategy and organization grow from the same new DNA. In AI-native companies, teams may initially be small and highly proactive, with strong learning ability; once they discover methods for collaborating with AI, those methods could be opened to more people.

Zeng urges people to imagine a change larger than the internet, perhaps larger than the Industrial Revolution. The first stage of a technology shift is led by researchers and technical experts; once technology becomes infrastructure, product operations, user connection, and empathy become much more important.

Researchers currently create much of the value because they interact with AI most and understand its limits. Zeng says product managers will also become extremely important: they decide which customers to serve, what to build, what value to create, and where limited time and technology should go.

Zeng says many model companies are researcher-led because they are technology-first. But technical founders must understand the economics of business: impressive technology and substantial value creation do not automatically produce high profits when supply is homogeneous.

Zeng says today's model-generation founders need technical backgrounds, but the next application generation may not. Embodied-AI companies already pair model expertise with application and business experience. AI-native organizations also need contributors whose work is indispensable, rather than people whose role an Agent can replace.

Zeng expects future enterprises to look increasingly like partnerships. Law firms, consulting firms, and investment banks have long used this form, where every partner's contribution matters. He thinks more employees may become partners, with early members called co-founders and later ones partners.

Zeng says the right team size depends on the task; people must have complementary strengths rather than duplicate one another. He expects scientist-founders to be common in the current model-building stage, but their share to fall when the industry moves into the next stage.

Zeng distinguishes excellence, which grows through positive feedback, from greatness, which emerges after people endure repeated negative feedback and prove an early judgment right. Many contrarian ideas fail; an extraordinary founder must combine an unusual idea with exceptional execution.

Zeng says people who achieve something extraordinary often credit others and attribute their success to luck. He sees this humility as evidence that the cause is larger than the individual and that greatness requires a smaller ego, empathy, and a mission beyond personal recognition.

A founder focused on proving personal greatness may seek short-term validation and filter out information that challenges their assumptions. Zeng says leaders on a different path ask how to do the work well, listen to varied views, and build cohesion around a shared dream.

Zeng distinguishes opportunity-, strategy-, vision-, and mission-driven companies. A real mission must be sincerely believed, not written on a wall to motivate others. When he asks founders what kind of person they hope to be ten years later, their answer can reveal what is driving them.

A desire to build a $10 billion company may reflect a search for external recognition and leave a founder vulnerable to every valuation round. Zeng says a better answer centers on genuine curiosity or a cause bigger than oneself. He cites Jack Ma's goal of making it easy to do business anywhere and Elon Musk's mission of taking humanity to Mars.

To distinguish signal from noise, Zeng recommends experiencing changes in the outside world broadly and thinking deeply about them. Following whatever is currently fashionable means arriving after others have already moved ahead; narrow exposure can also produce a locally optimal judgment that misses the wider picture.

Zeng prefers internalizing a way of thinking over memorizing theoretical rules, which people can apply mechanically. Zhang asks whether strategy still matters when it must change in real time. Zeng says it matters more than ever: one major mistake can eliminate a company within months.

Zeng summarizes AI-era strategy as “look ten years out, think three years, act on one year.” Industrial-era companies could extrapolate stable growth and outsource strategy, but the AI era is too uncertain for a linear plan. The three-year view creates milestones and connects near-term action to a longer-term ambition.

Thinking three years ahead forces leaders to imagine several milestones and decide what the current year's work should make possible. Zeng says this creates tension between short-, medium-, and long-term goals, enabling better tradeoffs than simply optimizing the next step.

The exact strategy rhythm depends on the industry: a company might look five years ahead, think two years out, and act in half-year cycles. But Zeng says leaders need a medium-term perspective and should see several steps ahead, not just one.

Zeng recalls asking a fast-growing startup what its business would look like after two more years and whether its ARR had a ceiling. The founders calculated their addressable market, saw that the ceiling was not high, and returned to explore new directions.

ARR can create an illusion because it annualizes the current moment and projects it forward. A service may seem to be growing rapidly until it hits a ceiling; once a model's capabilities spread or competition increases, customers may stop paying a premium.

The discussion turns to a basic economic distinction: when competition is limited, prices reflect customer value; with many competitors, prices tend toward marginal cost. The hosts connect this to how companies should measure progress when goals cannot be reduced to simple KPIs.

They explain OKRs as a way to coordinate around qualitative goals, though implementation is demanding and often degenerates into another KPI system. In future organizations, alignment means sharing goals, forecasts, and the cognitive models behind decisions.

The guest argues that AI could make OKRs more practical: an AI-enabled organizational system could monitor natural-language goals and flag drift or emerging trends. Rather than prescribe strategy by industry, he says companies should be understood by whether they are exploring, forming, maturing, or executing a strategy.

Foundation-model firms have largely moved into strategy formation, while embodied-robotics companies are still exploring. The guest calls the shift from strategic planning to “strategic generation”: building an environment where strategic insight can emerge continuously.

Strategy, he says, is a timely decision made by the right people. An organization should therefore work like a neural network and task flow, with humans and AI contributing to an evolving collective model that can surface ideas no individual had previously considered.

Such an organization needs a culture suited to broad participation. The guest returns to “context, not control”: give people and AI enough shared context, align their goals and understanding, and let decisions emerge from that common picture.

Humans cannot process information at AI scale, but they can make creative connections that are not yet explicit in the data. The hosts ask whether hostility toward “old-timer” leaders reflects the mismatch between legacy hierarchies and the new organizational model.

The guest traces respect for elders to the value of accumulated experience. In a period of intense pressure and changing work, younger employees may see hierarchical norms as offering little value and as asking them to sacrifice their own time and potential.

He suggests that experience which can be codified has increasingly been absorbed into large models, letting individuals use AI to acquire expertise quickly. In the new culture, creation remains central: like a sports team, an organization needs distinct roles while recognizing every contributor.

The guest describes that culture as open, transparent, and based on co-creation. He contrasts this with much knowledge work of the previous era, which was repetitive, offered little growth, and was managed through close control.

He expects AI to give people more freedom to pursue meaningful work, making happiness and emotional value necessary features of future organizations. Founders must adapt too: employees are drawn by inspiration and purpose, not by a boss's authority or manipulative tactics.

Returning to strategy, the guest says firms in a development phase should improve organizational and execution efficiency and turn a directional strategy into a repeatable operating model. Exploration-stage firms instead need leaders who encourage varied experiments and reduce the cost of learning.

Exploratory companies must remain willing to change direction when evidence contradicts their convictions. They should absorb others' strongest ideas and turn them into the basis for another round of innovation; here, innovation efficiency matters more than execution efficiency.

The hosts discuss how consensus around agentic coding can create homogeneous competition, while new labs often pursue ideas outside the consensus. Today's contrarian approach may become tomorrow's accepted direction, as the market continually renews itself.

Homogeneous competition has weakened expectations that model companies can sustain monopoly-level profits. That raises questions about whether they can continue large capital expenditures and whether the infrastructure built to serve them will earn an adequate return.

The guest still sees substantial room for large-model progress and expects infrastructure investment to continue, though cyclical overbuilding may eventually exceed near-term demand. That could benefit application companies by making existing infrastructure available at lower cost.

As intelligence becomes cheaper, more capable, and more reliable, application companies may gain room to build. They should not simply wait: participating now helps them develop experience before adoption accelerates, and the guest urges founders to retain confidence in their future-oriented mission.

Robotics remains at an early exploration stage, constrained in part by data. One path begins with a general robot-brain model; another starts with a particular scenario, closes the loop through embodied data, and uses that experience to improve the model.

The guest says neither robotics path can be declared the winner in advance. General-purpose efforts must prove their breadth, while scenario-led firms must find markets large enough to support strong performance and potential generalization; both are like climbing the same mountain from different sides.

The guest compares robotics to the period before ChatGPT for foundation models, and also to the early automobile industry, when many companies assembled cars from bought-in parts. Ford's 1913 powered assembly line made mass production possible and brought affordable cars to ordinary buyers.

The hosts discuss Google's uncertain chances of becoming a defining consumer application provider, while the guest thinks it can remain a strong AI cloud provider thanks to its infrastructure. They see corporate rise and decline as a normal cycle that gives new firms room to grow.

ByteDance has strong models and products such as Seedance, but the hosts distinguish that capability from creating a defining consumer application. They debate whether Doubao remains a chatbot rather than an agent, which must independently complete tasks.

Turning a chatbot into an agent is part of the race to build the AI-era operating system and its coordinating entry point. The guest says older services may be displaced and social relationships will be reshaped, with agents playing a larger role.

The guest calls many current products transitional, comparing QQ's limited role in mobile internet with WeChat's central role. The hosts ask whether incumbents can create the next product internally, and the guest says mobile internet combined disruption with continuity from the internet's architecture and network effects.

That continuity let established firms move into mobile, while disruption enabled native companies such as Douyin and Pinduoduo. The guest sees AI as more disruptive: it may overturn the Industrial Revolution's productive forces, whereas the internet mainly changed relations of production.

Because AI may have less continuity with existing technology, legacy capabilities may not transfer linearly and incumbents may struggle to build next-generation products internally. Organizational culture is one barrier; another is understanding agents, platforms, and ecosystems.

A large incumbent faces a different transformation challenge from a small AI-native startup. The guest says a company must first understand agents and build a successful one, then provide infrastructure that can become a standard and attract users and developers; the hosts raise xAI's strategy and organization.

The guest says he has seen xAI engineers exhausted and argues that sustained extreme pressure can suppress the open conditions innovation needs. He questions whether a tightly controlled, top-down approach leaves room for answers to emerge from the organization.

In his view, xAI's strategy is driven by Musk's will rather than generated through the organization. Meta also faces leadership and resource-allocation challenges; by contrast, the guest praises the Altman–Sutskever pairing at OpenAI and Anthropic's strong organization and mission-led focus.

Asked about China, the guest describes Kimi as relatively native, Qwen as benefiting from an early start and major-company resources, and ByteDance as accelerating after Zhang Yiming committed to the field. The hosts ask why Chinese competition is slower to converge and often becomes a crowded race.

The guest says Chinese firms often start from similar prototypes and recruit from overlapping talent pools, encouraging homogeneous competition. Still, he credits companies associated with Yang Zhilin and Yan Junjie, as well as Zhipu, with earlier or deeper insight that has kept them ahead.

The hosts debate model-building at application companies such as Xiaohongshu and miHoYo. The guest says their focus is applications, while in-house models can supply foundational capabilities; building a strong model can be a reserve capability, though the right degree of investment remains unsettled.

The guest compares the market to the opening phase of a chess game and avoids short-term forecasts. He expects an agent entry point may emerge within about five years, while a major consumer application comparable to Douyin or Pinduoduo could take eight to ten years; society may need at least a generation to absorb AI as a new means of production.

As intelligence becomes available externally, creativity may become a key source of difference between people. Drawing on Peter Drucker's framework, the guest describes productivity, management, and knowledge revolutions, and sees AI as ushering in an age of creativity.

AI will take over structured knowledge work, the guest argues, freeing people to develop new potential. Creativity is not limited to art: it can mean defining complex problems and imagining roles earlier generations could not foresee.

People may create new forms of companionship and support for growth and happiness. The conversation turns to the possibility of new human needs and forms of happiness that do not necessarily advance the material world.

The guest expects companies, offices, and business schools to change as virtual life expands; the hosts suggest the metaverse may have been a sound direction pursued too early. He says younger entrepreneurs imagine a world unlike today's and remains optimistic about the scale of change.

He expects education to move beyond cramming facts toward developing human capabilities, but young people feel caught between their own ideas and conventional schooling and career expectations. Society must invest in exploring new educational approaches.

Within organizations, humans and AI will collaborate closely; across society, AI will become pervasive and may be accompanied by wearable cameras, voice devices, and other tools that augment people. Consumer hardware is already a lively field.

The guest compares hardware's future to early electrical appliances: a few major product categories may emerge alongside many small ones, but the winners are not yet known. He advises entrepreneurs to embrace uncertainty as an opportunity rather than fear it.

The guest urges entrepreneurs to embrace uncertainty as an opportunity. Asked about his mission, he calls himself a happy researcher and hopes AI will make people happier, while acknowledging that it will bring new problems.

He frames technological progress as repeatedly creating new possibilities that people learn to master. By ordinary measures, this has improved life: industry scaled subsistence, consumer society eased scarcity, and many physical goods are now abundant, though philosophical judgments about happiness are harder.

The internet reduced information asymmetry, and the guest expects AI to enable further possibilities. He says organizations should be judged partly by whether their people are happy, recalling Google's influence in Silicon Valley around 2007–08.

Google's culture emphasized difference, creativity, and time for personal projects, and helped the company sense that new organizational forms were needed. Asked about his daily routine, the guest says he talks with entrepreneurs and uses interviews and other information as input for his research.

He warns that AI tools can drive people into exhaustion because their processing speed exceeds human pace, so people need boundaries and a healthier rhythm. In rapid-fire questions, he names chili-fried pork and Shanghai as favorites, identifies a good consumer-facing agent as an important bet, and discusses his studio's name and books.

The guest recommends management classics including Built to Last and Peter Drucker's books. He also recalls a short book by Milton Friedman on how markets work, which shaped his understanding of their basic operating mechanisms.

He says his interest in management began with a preference for history and the humanities over science competitions; after finding economics too mathematical, he chose strategy in business school. He describes strategy as an organization's brain and says it belongs with the CEO, then distinguishes excellent strategy-driven founders from exceptional mission-driven ones.

The guest declines to judge Zhang Yiming's mission from limited personal interaction, but calls him a remarkable, vision-driven founder with a developed management philosophy. Zhang's idea that “the company is the product” anticipated a culture resembling Netflix's, and ByteDance became highly efficient once its strategy was clear.

In the closing exchange, the guest says clearer strategy eventually shifts a company toward efficient execution, but changing times can make an old playbook obsolete. Organizations must transform or risk being left behind.