After the Model Race: Zeng Ming on AI Applications, Organizational Change and Strategic Judgment

1. Guest Background

Zeng Ming approaches artificial intelligence through enterprise strategy and industrial change. A professor and strategy expert, he formerly served as Alibaba Group’s chief strategy officer and participated in the founding of Cheung Kong Graduate School of Business and Hupan. In this episode of Zhang Xiaojun’s business interview podcast, host Zhang Xiaojun asks him to connect that background with the choices facing AI companies and their leaders.

Zeng describes his method as a combination of academic training, participation in Alibaba’s rapid development and continuing research into emerging industries. Practice gave him an opportunity to test whether ideas he had taught actually worked. It also supplied material from which to develop new frameworks. That combination explains the interview’s movement between broad historical patterns and the practical demands of running an organization.

His recollections of Alibaba establish a useful starting point. He identifies the creation of Taobao, Alipay and Alibaba Cloud as consequential choices, then stresses that a major strategic commitment leaves many decisions unresolved. Alibaba Cloud still had to confront questions about service scenarios, architecture and technical direction. The relevant experience is therefore more than selecting a promising market: it concerns the contested work of translating direction into an operating business. Those recollections provide context for his analysis, without independently validating his predictions about AI.

2. What the Episode Covers

The interview’s organizing framework divides the industrialization of a general-purpose technology into three stages. First, the technology becomes widely usable infrastructure. Second, a proliferation of applications explores what can be done with it. Third, native applications emerge alongside deeper changes in how the new environment operates. Zeng uses this sequence to distinguish the capabilities companies need to accumulate, rather than treating every AI business as a participant in the same race.

He places 2026 near the completion of the first stage and the opening of the second. Token-based measurement and billing are central to his interpretation: intelligence becomes something that can be purchased and called at scale. Agents showing the potential to execute tasks independently supply another signal. This is his assessment of an industrial transition. It does not establish that every relevant capability is reliable or that mature consumer products have already arrived.

His explanation of the application stage draws on the progression from Toutiao to Douyin. In his account, early products helped develop recommendation algorithms; improvements in bandwidth and computing then supported short video, whose combination with recommendation produced a different kind of application. Applications consequently help direct technical development. They are not simply customers waiting for technology to finish. His suggestion that the second stage could offer eight to ten years of opportunity is a forecast within this framework, not an independently established timetable.

This leads to his characterization of model companies as potential AI cloud providers. He expects collisions and integration between them and existing cloud businesses. Large investment requirements, social dependence and the need for alternative suppliers underpin his prediction of oligopolistic competition with strong government oversight. The discussion presents a proposed industry structure, rather than evidence that this endpoint has already been reached.

For applications, he sees an opportunity resembling the browser: common standards could make agents easier to build and easier to invoke. The resemblance concerns the removal of barriers on both sides of a market. He explicitly says that an abstract historical pattern cannot reveal the actual interface. An eventual Agentic OS would go further, interpreting an intention and directly orchestrating agents. He assigns that possibility to the third stage, dependent on substantial exploration of agents and native hardware.

The product referred to in the conversation as “Lobster” illustrates the distance between a promising direction and broad usability. Zeng describes an early experiment attracting enthusiasts and adventurous CEOs, with immature technology making deployment difficult. He also qualifies his assessment: within his limited observation, he has not seen a sufficiently good mass-market consumer agent. His optimism about an application era therefore coexists with an acknowledgment that its compelling consumer products remain to be built.

3. Core Views: Reasoning, Examples, and Limits

Zeng’s skepticism about today’s model leaders becoming tomorrow’s native-application winners rests on a change in required capabilities. Infrastructure development initially asks whether a technology can work. Application development asks which human problem deserves attention, how people should experience the solution and what supporting technical work is necessary. Product originality and empathy for ordinary users do not automatically follow from model performance. This explains his position on OpenAI and Anthropic, but does not prove that either company must fail to make the transition. It remains a historical inference about organizational difficulty.

He also recognizes a mechanism through which model leaders may extend their advantage. His term “intelligence compounding” describes AI entering the model-training process while organizational focus and technical operations reinforce improvement. The qualification matters: he explicitly distinguishes this from fully autonomous self-learning. Assistance within a development process can be consequential without establishing that a model independently controls its own improvement. Removing that distinction would turn his account into a stronger technical claim than the evidence supports.

For application companies, the clearest example is advertising. Zeng contrasts generating creative material or conducting a single placement with undertaking the more complex task of growth. A narrow function may have limited room to develop and may be absorbed as foundation models improve. A broader task could require interacting decisions, execution and learning. Complexity matters in his argument because of the work it creates for a learning system; a more ambitious product description alone does not establish a defensible business.

The proposed application loop has several demanding links. An agent executes a task independently, receives feedback from the real world and improves through technical innovation, including context management, memory and interaction. Its accumulated capability may then extend into other scenarios. Without meaningful feedback, activity need not become learning; without transfer, improvement can remain confined to one task. Zeng says he has not yet seen a mature scenario demonstrating the larger opportunity he describes. “Compounding” is consequently a hypothesis to investigate, not a credential that the interview grants to an existing product.

His interpretation of network effects follows the same logic. More examples of the same data type may eventually yield diminishing returns, whereas diverse scenarios and richer connections could support new insights. He calls the expanding attraction of scenarios a “black hole effect.” The phrase is a conceptual description of a possible advantage. The episode supplies no quantified relationship between the number of connected scenarios and intelligence, so a company cannot demonstrate the effect merely by listing integrations.

Platforms, meanwhile, must solve a judgment problem. Choosing an agent can expose users to consequences involving important tasks, money and personal information. Zeng imagines an intermediary helping establish confidence in capability, charges and authorization. His proposed sequence begins with a successful agent that attracts users and developers, enabling a de facto standard. This helps explain why he doubts that inherited traffic alone guarantees a future gateway. A directory or a declaration of standards cannot substitute for understanding execution.

Inside organizations, he makes an equally consequential distinction: the institution he expects to decline is the hierarchical company organized around positions and command lines. Cooperation and division of labor continue. His alternative takes the task as its basic unit, with people and AI participating in definition, decomposition, acceptance and delivery. He emphasizes agency, visible contributions and complementary abilities. The one-person company serves as an illustration of amplified individual capacity, while its limits reaffirm the need for collaboration. This is a proposed direction of organizational development, rather than a universally tested replacement structure.

“Strategy generation” connects that structure to decision-making. Shared context and feedback from task execution could allow people and AI to produce insights that were not specified in advance. The implication is that organizational design shapes what strategy can emerge. However, the interview does not establish how reliably this mechanism improves decisions or resolves competing interpretations. It identifies a design challenge and a possible mechanism, leaving implementation and performance open.

Zeng also resists turning unconventional founders into a formula. Contrarian ideas can fail; excellence of the exceptional kind is recognized after sustained achievement. A mission must be sincerely held. His broader proposition is that human value increasingly involves creativity: originally defining complex problems and discovering needs involving emotion, growth and happiness. That widens the field beyond technical invention. His call for education to support exploration follows from this outlook, but remains a direction for inquiry rather than a validated educational program.

Entry, capability matching and automatic orchestration also need to be separated. An entry layer first reduces selection costs through standards and trust, helping users judge capability, price, privacy and authorization. An Agentic OS belongs to a later stage in which the system orchestrates agents from the user’s intent. Treating these as the same product would mistake the conditions for building an ecosystem for a system capability that already exists.

The idea of an organizational brain also changes the CEO’s role. Zeng treats organization-facing agents, an organizational operating system or a CEO agent as possible application directions. If collaboration relies less on command and control, the founder cannot retain a conventional boss mentality; the role moves toward a service-oriented CEO who supplies context, resources and conditions for co-creation. This is his prediction about organizational change, not a management model already validated across companies.

4. Learning and Application

A practical use of the interview is to change the opening question of a strategy review: is the team still discovering its direction, or does it already have an operating prototype worth reproducing? Zeng assigns different priorities to those conditions. Exploration calls for leadership involvement, inexpensive experiments and willingness to revise beliefs. Development calls for a robust, closed operating loop and efficient execution. An application of this distinction would document what makes delivery repeatable, where improvisation remains necessary and what would invalidate the prototype before expanding it. This is an editorial translation of his framework, not a reported implementation result.

His robotics discussion offers a useful discipline for comparing uncertain approaches. A team starting with a generalizable brain must test the breadth of its capability and its effectiveness in important settings. A team starting with one scenario must test whether that setting is large enough and whether its capability can expand. Zeng considers both routes logically plausible. The useful output is therefore a set of conditions under which each route could work, rather than an early declaration of a winner. Preserving options also has to be balanced against obtaining concrete evidence.

Agent companies can similarly separate purchasable foundation capabilities from technical work specific to their tasks. Zeng’s argument against duplicating foundation-model development does not remove the need for substantial application engineering. One practical assessment would ask what performance is lost when a model supplier changes, and which execution, memory or context mechanisms retain value. Because he acknowledges an evolving boundary between models and application technology, this cannot become an unconditional rule against training any model.

Revenue needs its own examination. Zeng warns that annualizing current growth can conceal a nearby ceiling, and that competition may move pricing toward marginal cost. A team could use these concerns to examine scenarios in which models absorb its feature, competing supply increases or prices weaken. Payment demonstrates present value; it does not establish durable bargaining power. His economic argument is useful as a stress test, without being a ready-made profit forecast for every company.

To connect direction with current spending, define the next two or three milestones and specify which one each near-term commitment serves. Zeng’s emphasis is the tension between short, medium and long horizons, rather than mandatory adherence to ten, three and one years. Timetables can change. Milestones should also identify capabilities required for the next stage; a sequence consisting entirely of revenue targets may leave the underlying strategic question unanswered.

Organizational tools deserve bounded experiments. Zeng imagines AI tracking natural-language goals and flagging departures from OKRs. He also recounts Li Zhifei describing the construction of an internal operating-system framework in three days. Neither establishes improved organizational performance. A limited trial could examine whether shared goals reduce misunderstanding, whether alerts assist decisions and who confirms a change of direction. Rapid construction is evidence about making a framework, not about its subsequent effectiveness.

Finally, evaluate the human pace alongside system throughput. Zeng warns that efficient AI tools can leave their users exhausted and recommends limiting inputs to recover a sustainable biological rhythm. Teams can reserve time for interpretation and recovery, then ask whether faster processing actually produced better judgment. This preserves a condition for the creativity his analysis values: people need enough attention and energy to decide what the machinery should be doing.

Source

• Original episode: 153. 和曾鸣聊产业史观：残酷的真相、会消亡的公司、优秀≠卓越、“OAI、Anth大概率不是原生时代大赢家”