Conversation with Liu Ye: OpenClaw is Just "Limbs," We Need to Progress from "Digital Employees" to "Digital Organizations," from "Building Soldiers" to "Deploying Formations"

Conversation | Zhang Peng

As everyone rushes to develop “digital employees” and “Agent tools,” and endlessly compete in niche scenarios, where is the true moat for AI startups?

Recently, GeekPark founder & CEO Zhang Peng and VisionFlow founder Liu Ye had a forward-looking discussion after the OpenClaw explosion. As a first-generation Chinese programmer born in 1979, Liu Ye has experienced the full cycle from low-level hardware to software, from enterprise-level integration (ToB) to online education (industrial internet). After months of closed-door research and conversations with top global AI researchers and leading domestic entrepreneurs—“talking about everything”—he reached a cold conclusion: treating AI as a “digital employee” to replace individual tasks is an over-simplification of real business from an engineer’s perspective.

In this dialogue, Liu Ye introduced a series of insightful concepts and frameworks, such as “gradual exposure” and “task high- and low-dimensional matrices.” A possible future gradually became clearer: AI’s next step is not a flood of tool-like workers, but the construction of a “digital organization” with collaboration, reporting, and reflection mechanisms. When corporate culture becomes unnecessary and low-dimensional work is thoroughly flattened, the future CEO may no longer be a “Chief Executive Officer,” but a “producer” with exquisite taste.

This is a speculative discussion on organizational forms, business barriers, and the ecological niche of new entrepreneurs in the AI era. It aims to spark deeper future discussions among entrepreneurs.

Below is a curated excerpt of the dialogue from GeekPark:

01 The battle for 10,000 A has begun, there are too many things to do,

but what is most important to do?

Zhang Peng: From Zuoyeg盒子 to today’s intense exploration of OpenClaw, what changes have you experienced yourself?

Liu Ye: I am part of China’s first generation of programmers, started coding from a young age. I’ve seen the rise of BASIC, DOS, Windows, and today’s Mac era, and witnessed the emergence of the three major portals. I’ve worked in enterprise informatization, aiming to be China’s IBM; later I shifted to Zuoyeg盒子, deeply involved in online education. Online education is a profound industry, the highest form of industrial internet, and the “last train.” This experience made me realize that the core of industrial internet isn’t technology, but industry itself—business. The pattern of industrial internet is: first, information matching; second, standardized products; third, supply chains; and finally, complex non-standard services. The higher the margin, the harder it gets.

So when the AI wave arrived, my first move was to spend nearly six months doing nothing but talking to all the capable people in HR. I talked to chief scientists from star startups, core algorithms, engineers, and researchers from top tech giants, as well as emerging AI entrepreneurs—almost a thousand hours of intense conversations. How deep? I could predict the second half of their sentences from the first half. The consensus was remarkably similar.

After a full circle of conversations, the conclusion was startlingly consistent: everyone is working on the same thing—digital employees. It reminded me of a strategic misjudgment by a big shot about cloud computing, who said Alibaba’s cloud is basically just a cloud disk. Using old frameworks to understand new things always leads to superficial insights.

Today, everyone thinks creating a digital employee with Claude to generate “digital sales” or “digital customer service” is easy. What are the technical barriers? The moat? When it becomes normal for someone to burn billions of tokens a day, it’s more like manufacturing—impossible to take off. So I ask every entrepreneur the same question: Why are you? What makes you capable? Are you younger? Smarter? Able to stay up late? Competing on one dimension— isn’t that just the difference between “10 seconds 69” and “10 seconds 70”?

Zhang Peng: Yes, today there are too many things to do, but what should be prioritized? Do you have thoughts on this?

02 A decade of industrial internet, today will repeat itself

Liu Ye: AI is very different, but I believe some patterns from industrial internet still apply. Early on, tools were built; mid-term, business was developed; finally, consulting emerged. When technology is immature, the first wave is engineers—they excel at overly abstracting the world, like Baidu’s “frame computing,” which sees everything as frames. But in the later stages of mobile internet, content and services matter more than frames.

Engineers tend to oversimplify organizational imagination. Look at the first-generation internet portals—Tencent and Alibaba thrived because they were closer to industry, even if slightly less technical. Today, the trend is the same: technology is becoming less important.

Zhang Peng: The current wave of liberal arts students is happy—they don’t need to code, and that seems okay. But in the long run, what are the evolving demands on people in the AI era? What changes?

Liu Ye: In China’s talent structure, I see a problem. The first generation of Chinese programmers were also product managers, because there was no formal product manager role back then. The concept of “everyone is a product manager” only emerged around 2010, after Jobs launched iPhone 4 and Zhang Xiaolong proposed the product philosophy. Before that, programmers handled product work as well—they learned to code not just for work, but out of passion. These unconstrained, passionate people are often the most outstanding.

But the second generation of programmers, over the past decade of industrial internet, have become “code farmers,” while product managers became architects. Programmers have been domesticated to not think about business. Now, with AI, the “coding” part is being eliminated; they are only “farmers” now. These young people are excellent, but they lack understanding of industry. So, the current “battle of 10,000 A” is fundamentally a flood of tools.

In the later stages of industrial internet, companies like Alibaba and Meituan all use top consulting firms (MBB) for business analysis, with consultants leading product managers to design workflows—because internet product managers are inherently lacking in systematic thinking. Feishu was built this way. ByteDance, though purely internet, also heavily employs consulting firms to build internal processes. In the AI era, this pattern will only strengthen.

03 The problem with enterprises is never employees, but organization

Zhang Peng: So, you think focusing on “digital employees” as a single point isn’t very meaningful.

Liu Ye: That’s my core judgment: digital employees are not the end goal; digital organizations are. If digital employees flood the market and even recruitment positions disappear, and everyone has good digital employees, then what? Will all companies be profitable and successful? Actually, all company problems are strategic and organizational, never just employee issues.

Today’s agents are still doing work for people, not making decisions. We internally restructured OpenClaw and created MetaOrg. It’s essentially a core that can generate agent teams. For any task, we don’t assign a single employee but build an “organization” to solve it. This organization has collaboration, reporting, mission, goals, and action modes.

Zhang Peng: But in the future, could one person be a department? Or even a company?

Liu Ye: That’s a very good question. We still focus on micro-tasks—for example, using AI to make a short video or write a document requires multiple rounds of dialogue. You say a sentence, it responds, and you give feedback. That’s tool-like use; it’s just very smart.

So, the concept of people and departments isn’t about quantity. When describing a senior role’s JD, we usually say: first, capable of doing various tasks; second, able to use all kinds of tools. A high-level role can understand intent, proactively plan, execute, deliver, report regularly, reflect, and adjust strategies based on deviations. That’s advanced capability.

Zhang Peng: A competent department is like “L4-level autonomous driving.”

Liu Ye: Exactly. When given a skill, it can complete complex tasks; with a skill system, it can handle integrated complex tasks; when multiple intelligent agents are orchestrated, it can do even more complex things, like producing a short drama. I often tell employees: when using MetaOrg, don’t see yourself as a manager, but as a chairman. You need to push its boundaries.

In the future, young entrepreneurs might get a TOKEN budget to experiment, rather than a fixed 500,000 yuan. How many tokens you’re willing to spend determines how advanced the role can be. The higher the role’s reasoning chain, the more back-and-forth trial-and-error, iteration, and reflection are needed.

Zhang Peng: Returning to the earlier question, if an agent group can be broken down into finer units or roles, then when it forms a team, the talent quality of each individual determines success or failure. This echoes the old organizational competition logic: higher talent density—more capable talent—makes core tasks easier to achieve and outpace competitors.

The core is: if in the future AI becomes omnipotent and we can call upon the best AI, then besides organizations providing different niche services efficiently, we also need to look at “talent density”—meaning your agents or bots are broken down into more atomic capabilities, increasing talent density. In complex tasks, results, efficiency, and innovation will be better. Is this a correct deduction?

Liu Ye: I agree. Inside a company, there’s usually an OD department—organizational development. To see if an organization can win, the common approach is to benchmark all competitors’ talent, assess their ability to match roles, and predict outcomes. So, in warfare, organizations rely on their organizational capacity, not just strategy. A prime example is Alibaba. They value organizational building so much that they’re experiencing a “second spring.” The founding team ages, but the organization can regenerate. Essentially, if one day we’re competitors and both use AI, I’d build a strong AI organization with excellent development capacity. I’d analyze all competitors’ agent skill systems, study their capabilities, and develop better skills or fill gaps. For example, I’d have a strategic department to observe and analyze.

Huawei’s “Five Looks and Three Fixes” methodology is a good example. I joke that if we use this approach for startups, we can beat 99% of competitors. The five looks are: industry trend, market and customers, competitors, internal capabilities, and strategic opportunities; the three fixes are: control points, goals, and strategies. This methodology can filter out most competitors because most people play chess randomly, relying on quick thinking, while experts engage in deep reasoning and strategic thinking. The first instinct of a master is to think as a commander—how to lead the battle.

Zhang Peng: The “Five Looks and Three Fixes” essentially means avoiding “reactive responses” and solidifying a long-term reasoning process.

Liu Ye: Experts rely on deep research and thinking models—first, look at global best practices and information; then analyze and reason deeply. When they give answers, they do so decisively.

I believe the core of future competition is modeling traditional industry operations, abstracting them into systems capable of orchestrating intelligent agents. This is the new organizational development (OD) capability, evolving into AIOD— the ultimate core competitiveness.

Alibaba’s strength lies in building organizations. Once the organization is in place, regardless of competitors or business, it can be competitive. Jack Ma once said that the purpose of war isn’t necessarily to seize a domain but to foster organizational growth. Alibaba measures whether a battle is worth fighting by organizational growth—a very high-level mindset. Ma himself is like a super-information hub, flying 200 times a year to gather intelligence, then using it to improve organizational building. He is truly a chairman, not just a CEO.

This is the highest form of organizational structure we see—able to span generations, cover different industries, continuously succeed, and rebound after decline. Usually, if a company appoints the wrong CEO within ten years, it’s likely to decline. So, looking at history and viewing current development from a higher dimension—even making some cuts and optimizations—are far more efficient than building from scratch at the bottom.

Today, anyone can easily build an agent, with very low employee onboarding barriers, plus open-source community support. The industry’s secrets are no longer many. In tools, the competition can never surpass open-source communities. So, what core advantage does open source lack and cannot replicate?

04 The physics of AI organizations: Why is “gradual exposure” key?

Zhang Peng: In the “old era,” organizational discussions emphasized culture, values, KPIs, etc. When transitioning from the old organizational management to the new AI agent organization era, which elements can be completely discarded, and which should be retained but transformed?

Liu Ye: Anthropic’s launch of skills is fundamentally based on the “gradual exposure” concept in AI coding—if AI receives a large amount of messy information, it risks context corruption and attention deficit disorder. Gradual exposure helps AI maintain good focus and output high-quality results. Relying on manual implementation of gradual exposure is essentially full manual dialogue, which is inefficient. Therefore, the core value of skills is to layer and decompose complex tasks, enabling AI to be gradually exposed.

This aligns with company management logic: the board focuses on strategy; the CEO on tactics and senior management; employees handle simple tasks. If 300 people participate in the same meeting, it becomes unmanageable. The core purpose of organization is to enable layered information processing—like database normalization, which improves efficiency through layered data compression. Complex problems must be decomposed and exposed gradually, not input all context at once. This is the core logic of traditional organizational forms, given the limited computing power at any given time.

Zhang Peng: Models consume enormous computing resources to generate from scratch, which is inefficient.

Liu Ye: Impossible to do otherwise. Relying on layered, gradual exposure is essential; resources must be called upon as needed, determined by the capabilities of AI models. Another reason Anthropic launched skills is that complex tasks have surpassed basic physical laws; skills can decompose complex tasks into low-dimensional, simple tasks. The key dimension isn’t difficulty but complexity—low-dimensional, high-difficulty tasks like coding or solving math problems are examples.

Horizon’s Yu Kai proposed a classic model: all job types can be divided into four quadrants based on “competition level” and “dimensionality”: high-dimensional high-competition, low-dimensional low-competition, low-dimensional high-competition, high-dimensional low-competition. For example, sales and engineers are low-dimensional, high-competition; product managers and CEOs are high-dimensional, high-competition; scientists are high-dimensional, low-competition—possibly the only person worldwide researching such topics, with low competition but high dimension. Tasks like high-quality short dramas or novels are high-dimensional, high-competition, currently beyond AI’s reach; while code optimization, a low-dimensional, high-competition task, AI can handle well. The higher the dimension, the fewer data sources, but the more data needed to train models. That’s why text models appeared first, followed by image and video models, and why short-video models are hard to implement. The supply-demand contradiction in high-dimensional tasks and data can only be addressed by decomposing tasks into skills, similar to splitting high-level roles into basic ones when talent is scarce. Only high-level roles like CEOs are irreplaceable.

Zhang Peng: Low-dimensional, high-competition tasks are most likely to be fully replaced by AI.

Liu Ye: Absolutely, and that’s already happening.

Zhang Peng: That’s why all low-dimensional, high-competition tasks should be solved by AI as soon as possible—decomposed into skills and implemented through agent organizations, often without human involvement.

Liu Ye: I have a preliminary idea: IBM and Accenture, as the world’s largest consulting firms, fundamentally extract best practices from industries, align with digitalization, and sell processes rather than tools. When companies buy risk management processes or IP, they hire consulting firms to implement. Our current core work is to build skill clusters, find top experts in each field, extract and align their capabilities, and form standardized skill sets. This is similar to Zuoyeg盒子—collaborating with Beijing No. 4 Middle School, Renmin University Affiliated High School, Gaokao question setters, and TAL’s teachers to extract core methods like question setting, teaching, and grading, then working with Baidu engineers to build systems—essentially aligning best practices. The core organizational ability is to assemble high-quality cross-disciplinary teams that understand industry, engineering, and can coordinate top industry experts, with capabilities in business, recruitment, and management. This is the core of the new AI SaaS enterprise.

Zhang Peng: Further extrapolating, future organizations should be designed backward from business needs. An organization is essentially an orchestration structure—like a business operating system. Placing people as productive units within suitable organizations maximizes value; otherwise, it can’t operate efficiently. The productivity elements have shifted from reliance on human labor to unlimited AI supply, creating a positive cycle of continuous expansion. Past organizational culture might now transform into goals and context, no longer needing slogans, meetings, or icebreakers.

Liu Ye: Culture is a management intent, not a business intent. In the old era, strategy started with vision, which determined values; organizations served strategy, and business validated everything. Culture was just a governance tool, not directly serving strategy, and could even be a founder’s personal preference.

Zhang Peng: In the past, there were gaps between people serving strategy. Is AI eliminating these gaps?

Liu Ye: Yes, in the AI era, culture is less relevant. Culture is part of human belief systems, but AI doesn’t need it. AI’s core requirement is computing power.

Zhang Peng: You mean AI needs goals and principles. A document can clarify goals and principles, and all productivity units can immediately synchronize and faithfully execute, eliminating friction in human organizations.

Liu Ye: Exactly. The old organizational chain: strategy → culture → talent → execution. The new AI organization: goals → principles → skills → orchestration. The entire management chain is compressed by half.

05 The last barrier: Aesthetics and orchestration

Zhang Peng: What are the new barriers for enterprises? Talent quality is replaced by Skill Set; as long as I have aesthetic judgment, I can acquire the best skills worldwide. Then, the next layer is “orchestration,” right? What changes will that bring?

Liu Ye: Just like Huaqiangbei can buy all electronic components, but not everyone can make an Apple. Steve Jobs’ definition of aesthetics was very clear: seeing enough good things in the world and being able to distinguish quality is aesthetic judgment. If you’ve never seen good products, good processes, or good organizations, you can’t produce high-quality results.

Zhang Peng: Experience is the prerequisite for aesthetics.

Liu Ye: Experience plus talent—nothing more.

Zhang Peng: Aesthetics manifest in two ways: one is proactive design and orchestration; the other is recognizing and selecting emerging high-quality things in chaos. These two are not mutually exclusive.

Liu Ye: Not mutually exclusive. Some of Apple’s achievements come from independent R&D, some from acquisitions. The core is having aesthetic judgment—no need to reinvent the wheel; when necessary, develop independently.

Zhang Peng: The key is whether the agent operates within a set module to confirm paths, enabling emergent orchestration; or whether all paths are preset, leading to a designed orchestration?

Liu Ye: Emergence is non-manipulative; you need to set seed rules and principles first, which reflects one’s aesthetic judgment. Like a good engineer who can produce a useful OpenClaw with 500 lines of code, while a poor engineer might write 50,000 lines but still not achieve the same effect. The underlying seed rules still need human setting.

Zhang Peng: So, waiting for emergence in chaos is too long; orchestration remains crucial. Ultimately, does this orchestration have to come from the founder or be more like a “producer”?

Liu Ye: I think the “producer” analogy is very fitting. Even with emergence and scale effects, data labeling, data cleaning, and continuous alignment of algorithms are needed to prevent disorderly expansion.

The orchestrator depends on business complexity—complex tasks like shooting a short film or writing prompts can’t be done by one person alone. The “one-person company” concept is often misused; the world can’t be infinitely simplified. Computers can be operated by one person, but mastering all high-dimensional capabilities is impossible for a single individual. Super talents like Elon Musk or Fei-Fei Li, who excel across multiple fields and can take on any role, are extremely rare.

Zhang Peng: If we can call upon the world’s top agents and skill systems—say, a top screenwriter—could we, in theory, leverage these resources to produce a globally renowned, profitable film? The screenwriter has a core strength (a good script), but can’t handle all aspects. Is this “core strength + global resources” closed loop feasible?

Liu Ye: It’s fundamentally a data issue—whether there exists data that stores the highest-dimensional information. For training CEO skills, there’s currently insufficient data: speeches by Ren Zhengfei, oral accounts by Jack Ma, can’t fully capture their high-dimensional cognition; even collecting all global company reports and CEO speeches can’t produce a CEO-level model because the core capabilities are implicit knowledge, not fully exposed in text.

Zhang Peng: That means the core ability of a CEO can’t yet be fully vectorized. This constrains the “one-person company” ideal— even if each person excels in a single dimension and leverages top global resources, the lack of a core orchestrator remains. Ultimately, having the best “components” still requires strong orchestration.

Liu Ye: Same applies to product managers. Their implicit knowledge can’t be fully textualized. This is also why current AI companions and content generation lack “vitality”—they lack high-dimensional implicit knowledge data. When data is scarce, focus on skills; when abundant, develop models. Robots can’t be deployed effectively now mainly because of insufficient data.

Zhang Peng: From this, the future competitive advantage of companies will no longer be about access to top models—initial AI resources seem uniform, and computing power correlates with financial strength and business loop capabilities. Ultimately, the difference will come down to the “producer” itself—its orchestration ability and the innovativeness and significance of its goals, which form the core competitiveness.

Liu Ye: A McKinsey partner once told me that McKinsey’s core business is extracting best practices, building models, and helping companies implement them step by step. For example, when consulting for Chinese automakers, they learn from Japanese counterparts like Toyota—essentially copying and applying best practices.

The case of Mimi Meng making short dramas is very instructive. She’s a literature major, but her core team includes top Tsinghua and Peking University math and CS graduates, who analyze the logic of viral short videos, achieving very high hit rates. This approach is essentially social engineering modeling for the industry—though overfitting is possible, the modeling direction is correct.

IBM, Accenture, McKinsey do similar things—first-generation McKinsey modeled best practices into partner knowledge; IBM transformed them into digital processes; both are fundamentally “selling management and workflows.”

Zhang Peng: The core is extracting best practices, repeatedly validating and implementing—that’s the key to future business success. Only with thorough decomposition can high-efficiency orchestration be achieved. So, your next focus is to advance along this path?

Liu Ye: Over the past three years, we’ve mainly done AI ToC (Theory of Constraints) business, reconstructing the entire teaching and research system with MetaOrg. This isn’t just about “using AI to improve efficiency.” We built a complete agentic research organization, with virtual research teams: language acquisition research team tracking latest theories, corpus collection team sourcing authentic expressions, dialogue evaluation team establishing multi-dimensional speaking standards, dialogue design team translating teaching methods into natural human-machine interactions, question design team solving content adaptation, data analysis team mining real signals from user behavior. Each team has its own skills, workflows, and evaluation standards. About 80% of tasks—lesson data labeling, monitoring, assessment, user insights, product iteration—are now handled by AI.

Our development path is from “AI as a function” to “AI as an organizational capability.” The role of an English teacher, a medium-complexity position, has been abstracted and generated into other roles via MetaOrg; with the latest skill architecture, more advanced roles can be built.

We’ve completed the full process of AI tutor, including orchestration abstraction and engineering implementation. In the future, we’ll likely upgrade from MetaTutor to MetaOrganization—its smallest unit is a role, not a person, focusing on collaboration and management among roles. Our current priority is connecting with top CEOs across industries, because they are the true “producers.”

Zhang Peng: So, you’re launching a more scalable department?

Liu Ye: The goal is to move toward a “company.” Large companies are essentially composed of smaller companies, with the smallest unit being a role. We need to focus on industry-wide strategic choices and also push product iteration from the role level—if roles aren’t well designed, even strong managers can’t build an efficient organization.

Zhang Peng: To build a good department, you must first decompose related capabilities and roles, then decompose roles into skills, aiming for these skills to reach SOTA levels.

Liu Ye: The only core method is co-creating with top-tier client companies. The skills developed must be evaluated by leading enterprises—like subordinate proposals needing supervisor approval—nothing self-congratulatory. For example, in short drama modeling, industry top institutions’ recognition is essential; otherwise, it’s not truly top-tier. Everything must be evaluated and measured.

Midjourney produces high-quality images because its team includes top photographers and engineers with excellent aesthetic judgment; LV’s training of image models with Stable Diffusion surpasses ordinary models because LV has the world’s top image aesthetics and data. Clearly, evaluation ability is the core. To build an AI company, emulate IBM or Huawei—after serving top automakers, they master best practices and output; Huawei spent 4 billion on IPD processes, used internally and externally, which is the real core advantage.

Zhang Peng: Essentially, the path is to decompose skills along best practices, achieve SOTA in skills, then upgrade to SOTA in roles and departments, and finally orchestrate to top-tier business. How to keep skills up-to-date? Just like biological evolution, each era’s SOTA may be replaced in the next. How to respond to this change?

Liu Ye: The core logic aligns with human and biological evolution—perception, planning, action, reflection. Maintaining high talent density and cross-disciplinary attributes, connecting to cutting-edge research, understanding business models, and co-creating with top industry clients—this is the only way.

Zhang Peng: From this, the best practice systems of top companies can help mid-tier firms leapfrog, but such systems are likely only accessible to resource-rich companies. Small and medium enterprises and startups face high barriers. The consulting industry has shifted from traditional services to tool-based products. Is the new generation’s opportunity only at the skill level? How to achieve disruptive innovation at the skill layer and avoid the “noble cycle” of high costs?

Liu Ye: In the last SaaS wave, companies like Salesforce, Palantir, Notion, Slack proved that young entrepreneurs still have opportunities—by avoiding areas where they lack advantages, focusing on universal skills, and finding niche ecosystems. Notion is a prime example: it abstracts text note-taking, not involving specific business processes, becoming a universal tool. Ultimately, the world will be a division of labor among countless agents. Young people need to find their niche, leverage their strengths, and align with future trends to avoid becoming victims of time. Over the past decade, the first internet generation was mostly returnees (relying on cognitive advantages), the second was programmers (relying on tools), and the third is second-time entrepreneurs in industrial internet—patterns are clear. Young people must understand the mid-game and their own advantages.

Zhang Peng: So, you believe that the limited role of local innovation at the skill level means the greatest opportunity for the new generation is goal innovation—identifying emerging new goals, combining high-quality skills, and continuously evolving to build new systems and achieve breakthroughs.

Liu Ye: The competition at the skill level is very subtle. Currently, skills are hot, but if someone aligns with top human experts and develops better skills, existing skills will be replaced. This circles back to the moat—early movers may not last; they could become nutrients for higher-dimensional opponents.

Zhang Peng: I worry about becoming just a “loading program,” helping higher-dimensional opponents lay the foundation. If I only optimize efficiency on existing goals, it’s meaningless—the efficiency advantage will eventually be leveled. To make breakthroughs, the new generation must make fundamental differences in goals.

Liu Ye: Exactly. If you can’t grow into a core force yourself, you’re just nourishing higher-dimensional opponents. The essence of business is simple: knowing who your customers are, how to serve them, and making them indispensable. If young people don’t understand who their customers are, they can’t optimize.

Zhang Peng: Also, focus on incremental markets. In saturated markets, competition is extremely fierce. If your business succeeds, it will elevate competitors to the same advanced level—these companies have wealth and cognition, making it hard for young firms to compete in the existing market.

Liu Ye: The success of SaaS companies like Notion and Slack in the last wave was driven by goal differentiation.

In early SaaS development, Chinese funds favored investing in scientists, but later found that scientists are better suited for collaboration than entrepreneurship—high-dimensional, low-competition fields are very different from high-dimensional, high-competition business logic. The higher the domain dimension, the harder it is to switch to new fields. The core thinking shifts from technical competition (low-dimensional, high-competition, immature tech) to business competition (high-dimensional, high-competition, industry practitioners and product managers). For example, when Apple launched the iPhone, top-ranked apps were mostly developed by programmers; years later, with the rise of industrial internet, those programmer-led apps were replaced.

In the AI era, if the logic continues from mobile internet, Silicon Valley’s core strength remains experienced practitioners, just like China’s second-time entrepreneurs in industrial internet. The opportunity for young people is still to find differentiated goals.