Conversation | Zhang Peng

When everyone rushes to develop "digital employees" and "agent tools," endlessly competing within niche scenarios, where is the true moat for AI entrepreneurship?

Recently, Zhang Peng, founder and president of Geek Park, and Liu Ye, founder of VisionFlow, engaged in a forward-looking discussion following the emergence of OpenClaw. As one of China’s first-generation programmers, born in 1979, Liu Ye has experienced the full cycle from底层 hardware to software, and from enterprise-level integration (B2B) to online education (industrial internet). After months of seclusion and extensive conversations with researchers from top global AI companies and leading domestic entrepreneurs, he reached a sobering conclusion: treating AI as a “digital employee” to replace individual tasks is an over-simplification of real business operations driven by an engineer’s mindset.

In this conversation, Liu Ye introduced a series of highly insightful concepts and frameworks, such as “progressive exposure” and the “high-low dimensional matrix of tasks.” Through the discussion, a future possibility became increasingly clear: the next step for AI is not an oversaturation of tool-like agents, but the construction of a “digital organization” equipped with collaboration, reporting, and reflection mechanisms. When corporate culture becomes unnecessary and low-dimensional work is fully eliminated, the future CEO may no longer be a “Chief Executive Officer,” but rather a “producer” with极致 aesthetic sensibility.

This is an exploration and simulation of organizational structures, business barriers, and the ecological niches of the next generation of entrepreneurs in the AI era, aiming to spark deeper discussions among future entrepreneurs.

The following is a curated transcript of the conversation compiled by GeekPark:

The battle of 0.1 million A has already begun, and there's so much you can do.

But what matters most is what to do.

Zhang Peng: From Homework Box to today, your deep interest in exploring the changes brought by OpenClaw—what changes have you personally experienced?

Liu Ye: I am part of China’s first generation of programmers, having started learning programming from a young age. I’ve witnessed the evolution from BASIC to DOS, then Windows, and now the Mac era, as well as the rise of the three major internet portals. I worked in enterprise informatization, aiming to build China’s IBM; later, I transitioned to Zuoyebang, deeply engaging in online education. Online education is a profoundly significant industry—the highest form of industrial internet and the “last train.” This experience taught me that the core of industrial internet is not technology, but the industry itself—its business. The pattern of industrial internet follows this sequence: first, information matching; then standardized products; followed by supply chains; and finally, complex non-standardized services. The later the stage, the higher the gross margin—and the harder it is to execute.

So when the AI wave hit, the first thing I did was spend nearly six months doing nothing but talking—having HR arrange conversations with everyone I could. From chief scientists at top startup companies to core algorithms engineers and researchers at major foundational model firms, and emerging AI entrepreneurs—I talked to them all, accumulating nearly a thousand hours of deep discussions. How deep? So deep that when someone said the first half of a sentence, I could predict the second half. Everyone’s consensus had become remarkably aligned.

After chatting with everyone, the conclusion was surprisingly consistent: everyone is doing the same thing—digital employees. This reminded me of a strategic misjudgment by a prominent figure regarding cloud computing, who said that Alibaba’s cloud was essentially just a cloud storage drive. When you try to understand something new through an old framework, you can only ever see the surface layer.

Today, everyone thinks it’s easy to become a digital employee and use Claude to build a “digital sales agent” or “digital customer service.” Where’s the technical barrier? Where’s the moat? When burning hundreds of millions of tokens per day becomes commonplace, it feels more like manufacturing—it simply can’t scale. So I ask every entrepreneur the same question: Why you? What makes you different? Are you younger? Smarter? Better at pulling all-nighters? Competing on a single dimension is just like the difference between “69 seconds in 10” and “70 seconds in 10.”

Zhang Peng: Hmm, there’s so much we can do today, but what we should do is what matters most. Do you have any thoughts on this?

The ten-year journey of industrial internet will be replayed today.

Liu Ye: AI is very different, but I believe it still shares some underlying patterns with industrial internet. In the early stage, focus on tools; in the middle stage, focus on business; finally, move into consulting. When technology is immature, the first wave to enter will always be engineers, who are skilled at over-abstracting the world—for example, Baidu’s “box computing,” which treats everything as a box. But the latter half of mobile internet was about content and services, not boxes.

People with engineering backgrounds often oversimplify business when imagining organizations. Look at the three major first-generation internet portals—those that ultimately succeeded best were Tencent and Alibaba, which were slightly farther from technology but much closer to industry. Today is no different: technology is becoming increasingly less important.

Zhang Peng: This wave of humanities students seems quite happy—being unable to code doesn’t seem to matter anymore. But in the long term, what exactly does the AI era require of people? What has changed?

Liu Ye: In China’s talent structure, I’ve noticed something. China’s first-generation programmers were also product managers, because the role of product manager didn’t exist back then. The position of product manager only became widely recognized around 2010, after Steve Jobs launched the iPhone 4 and Zhang Xiaolong articulated his product philosophy, leading to the phrase “everyone is a product manager.” Before that, programmers took on the responsibilities of product managers—programmers came first, then product managers—so the first-generation programmers were all product managers. These early programmers didn’t learn coding for work; they did it out of passion, driven by genuine interest. It was precisely these unconventional, boundary-defying individuals who stood out the most.

But over the past decade, the industrial internet has turned programmers into "code farmers," while product managers became architects—turning code farmers into individuals who no longer think about business. Now that AI has arrived, the "coding" part has been eliminated; without evolving, they are left with nothing but the "farmer." These young people are highly talented, but their understanding of industry is nonexistent. Thus, today’s "Ten Thousand A War" is fundamentally still an overabundance at the tool layer.

In the later stages of industrial internet, companies like Alibaba and Meituan routinely hire professionals with backgrounds from top consulting firms (MBB) for business analysis, and rely on these consultants to guide product managers through business processes, because internet product managers inherently lack systemic thinking. Feishu was built this way. Although ByteDance is a pure internet company, it extensively uses consulting firms to establish internal processes. In the AI era, this trend will only intensify, not diminish.

03 Problems in a company are never about employees—they are about the organization.

Zhang Peng: So, you think focusing on "digital employees" isn't very meaningful.

Liu Ye: This is my most fundamental insight: Digital employees are not the end goal—digital organizations are. If digital employees become widespread and even recruitment roles disappear, and everyone can have high-quality digital employees, then what? Will all companies automatically become profitable and successful? In reality, every company’s challenges are fundamentally strategic and organizational—not issues of employees.

So today, agents are still working for people, not making decisions for them. Internally, we’ve transformed OpenClaw into something called MetaOrg—a core engine capable of generating agent teams. When we tackle any task, we don’t assign a single employee; instead, we build an “organization” to solve it. This organization has collaboration structures, reporting lines, a mission, goals, and defined ways of operating.

Zhang Peng: But in the future, is it possible for one person to be an entire department—or even an entire company?

Liu Ye: That’s an excellent question. Let’s get down to the specifics—when using AI to create a short video or write a document, multiple rounds of dialogue are needed. You say something, it responds, then you give feedback. This is tool-like usage—it’s just very intelligent.

The concept of individuals and departments is not about quantity—more or fewer. When describing a senior role’s job description, we typically say: first, the ability to get things done across various tasks and use multiple tools. A senior role, however, is one who understands intent, proactively plans pathways, takes initiative in execution, delivers results, provides regular updates, reflects on and summarizes outcomes, and dynamically adjusts strategies based on deviations in results. This is what constitutes advanced capability.

Zhang Peng: A qualified department must be like "Level 4 autonomous driving."

Liu Ye: Yes. When given a skill, it can accomplish complex tasks; when given a skill system, it can handle complex integrated tasks; and when multiple agents are orchestrated together, it can achieve even more sophisticated outcomes—like producing a short film. I often tell my employees during meetings: when using MetaOrg, don’t think of yourselves as managers—think of yourselves as chairpersons. Push its boundaries and explore them.

In the future, young entrepreneurs may no longer receive 500,000 yuan from their families to start a business—instead, they might be given a TOKEN budget to experiment. How many TOKENs you’re willing to spend determines the level of sophistication of the role you can pursue. The more advanced the role, the longer the reasoning chain, and the more iterative testing, refinement, and learning it requires.

Zhang Peng: Returning to the previous question, if there is a group of agents that can be broken down into finer units—similar to role and capability specialization—when they form a team to tackle core tasks, the quality of each individual determines success or failure. This brings us back to the business organization competition logic of the previous era: talent density—where higher talent quality makes it easier for an organization to achieve and outperform its core objectives.

The core of this issue is that if, in the future, all AI systems are omnipotent and we can all access the best AI, then beyond commercial organizations creating value by delivering specialized services more efficiently, another dimension must return to the concept of “talent density”—namely, the higher the atomic-level capabilities of your agents and bots within this system, the higher the talent density, leading to better outcomes, efficiency, and even innovation in complex tasks. I’m not sure if this reasoning is correct.

Liu Ye: I agree with this perspective. Within companies, there’s typically a department—large enterprises often call it OD, or Organizational Development. To assess whether an organization can win, the standard approach is to benchmark all of the opposing side’s talent, evaluating how well individuals match their roles and how strong their capabilities are relative to those roles, in order to predict the outcome of the competition. Therefore, companies generally rely on organizational capability rather than business strategy to win. The most classic example is Alibaba. Alibaba places great emphasis on organizational development, which is why it is now experiencing a “second spring.” Founding teams age over time, but organizations can be endlessly renewed. Fundamentally, if one day we were competitors both using AI, and I had built a powerful AI organization with strong AI organizational development capabilities, how would I construct it? I would systematically analyze each competitor’s agent skill system, examining their skill code. Then, within my own system, I would develop superior skills and even fill in the functional gaps they lack. For instance, if I have a strategy department, I would begin with observation and analysis.

Huawei has a methodology called "Five Looks and Three Decisions." I jokingly told my friends that if we entrepreneurs simply applied this framework, we could outcompete 99% of our rivals. The "Five Looks" involve examining industry trends, market and customers, competitors, our own capabilities, and strategic opportunities; the "Three Decisions" are defining control points, setting goals, and formulating strategies. This methodology is sufficient to eliminate the vast majority of competitors, because most people play chess randomly—they rely on fast thinking—while experts default to deep thinking and analytical reasoning. My first instinct is to think like a commander: how should I approach this?

Zhang Peng: What is meant by "five observations and three determinations" is essentially about avoiding knee-jerk reactions and instead establishing a structured, long-term reasoning process.

Liu Ye: Experts rely on deep research combined with critical thinking—they first examine global best practices and information, then synthesize and analyze it with deep reasoning before delivering an answer that strikes directly at the core.

So I believe the core of future competition will come down to one thing: modeling traditional industry operations, abstracting them into systems capable of intelligent agent orchestration. This is the next-generation organizational development (OD) capability, which will evolve into AIOD—the sole core competitiveness of the future.

Alibaba’s core strength lies in building organizations; once the organizational structure is solid, it can compete effectively against any opponent and in any business area. Moreover, Jack Ma once said that the purpose of a battle is not necessarily to capture a specific market, but to achieve organizational growth through the battle. Alibaba judges whether a battle is worth fighting based on whether it contributes to organizational development—a highly sophisticated mindset. Jack Ma himself functions like a super information hub, traveling 200 times a year to gather diverse insights, which he then uses to refine the organization. He is truly a chairman in the fullest sense, not merely a CEO.

This is the highest form of organization we have seen—capable of spanning generations, covering diverse industries, consistently achieving success, and rebounding after periods of decline. Typically, a company that appoints the wrong CEO within a decade is likely to decline. Therefore, learning from history and viewing current developments from a higher-dimensional perspective allows us to achieve greater efficiency by refining and optimizing existing models than by building from scratch.

Anyone can now easily build an agent, with extremely low barriers for employees to get started, and with the support of open-source communities, there are few secrets left in the industry. Competition at the tool level can never match the power of open-source communities. So, what is the core competitiveness that open-source communities lack and cannot replicate?

The Physics of AI Organizations: Why Is "Progressive Exposure" Key?

Zhang Peng: In the previous era, discussions about organizations emphasized a range of elements such as organizational culture, values, and KPIs. As we transition from traditional organizational management in the previous era to the new era of AI agent organizations, which elements can be completely discarded, and which can be retained but require transformation?

Liu Ye: The primary reason Anthropic introduced skills is the concept of "progressive exposure" in AI coding—when AI is overwhelmed with excessive, disorganized information, it suffers from context degradation and attention deficits, leading to confusion; only progressive exposure enables the AI to maintain focus and produce high-quality outputs. Relying on humans to achieve progressive exposure is essentially equivalent to fully manual conversations, which is inefficient. Therefore, the core value of skills lies in decomposing complex tasks into layered steps to enable progressive exposure for the AI.

This aligns with the company’s management logic: the board focuses on strategic issues, the CEO handles tactical matters and manages senior leadership, while employees deal with routine tasks. If 300 people were to participate in the same meeting, it would be impossible to conduct. The fundamental purpose of organizational structure is to enable layered information processing—just as the third normal form in databases improves efficiency through hierarchical data compression. Complex problems must be broken down into layers and revealed incrementally, rather than overwhelming participants with excessive context all at once. This is precisely the core logic of traditional organizational structures, given that computational capacity is inherently limited at any given time.

Zhang Peng: The model requires enormous computational power to generate from scratch each time, which is too inefficient.

Liu Ye: It’s impossible to achieve; the core still relies on layered, progressive exposure—resources that must be called need to be called, as determined by the capability boundaries of AI models. Additionally, Anthropic introduced skills for another reason: complex tasks have surpassed basic physical laws, and skills enable the decomposition of complex tasks into a series of low-dimensional, simple tasks. The key dimension for distinguishing tasks is not difficulty, but complexity—there are different types, such as low-dimensional difficulty and high-dimensional difficulty. For example, programming and solving math problems are low-dimensional, high-difficulty tasks.

Yu Kai from Horizon proposed a classic model: all jobs can be categorized into four quadrants based on “level of competition” and “dimensionality”—high dimension/high competition, low dimension/low competition, low dimension/high competition, and high dimension/low competition. Sales and engineers fall into the low-dimension/high-competition quadrant; product managers and CEOs belong to the high-dimension/high-competition quadrant; scientists are in the high-dimension/low-competition quadrant—these fields may have only one researcher worldwide, with minimal competition but extremely high dimensionality. Tasks such as producing high-quality short dramas or great novels, which are high-dimension/high-competition, are currently beyond AI’s capability; whereas low-dimension/high-competition tasks like code optimization are already well within AI’s capabilities. The higher the dimensionality of a task, the fewer the available data sources, yet the greater the volume of data required to train models—this explains why text models emerged first, followed by image and video models, and why short-video models remain difficult to deploy. This supply-demand imbalance between high-dimensional tasks and high-dimensional data can only be addressed by decomposing tasks into skills—just as companies, when unable to find talent for senior roles, break them down into three junior positions—except for roles like CEO, which remain irreplaceable.

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

Liu Ye: It will be completely replaced, and this replacement has already occurred.

Zhang Peng: Indeed, so all low-dimensional, highly competitive tasks should be resolved by AI as soon as possible. These can be broken down into skills and implemented through agents, without necessarily requiring human involvement.

Liu Ye: I have an initial concept: IBM and Accenture, as the world’s two largest consulting firms, fundamentally focus on distilling industry best practices and aligning them with digital transformation—they sell processes, not tools. When enterprises procure risk management processes or intellectual property, they always engage consulting firms to implement them. Our current core task is to build skill clusters by identifying top experts across various domains, distilling their capabilities, aligning them, and forming standardized skill sets. This is similar to the model of Zuoyebang: Zuoyebang collaborated with Beijing No. 4 High School, Renmin University Affiliated High School, the national college entrance exam question-setting team, and Xueersi teachers to distill core methods such as question design, instruction, and grading, then partnered with Baidu’s algorithm engineers to build a system—essentially aligning best practices. The core of organizational capability lies in assembling high-quality cross-disciplinary teams that understand both industry and engineering, can coordinate with top experts in various verticals, and possess business development, talent acquisition, and management skills—this is the essential composition of a new-generation AI SaaS company.

Zhang Peng: Further extrapolating, in the future, we should reverse-engineer the required organizational structure from a business perspective. An organization is essentially an orchestration system—similar to a business operating system—where placing people as units of productivity into appropriately aligned structures maximizes their value; otherwise, the system cannot operate efficiently. Today, the key productivity factors have shifted: human labor has been replaced by AI, which is infinitely scalable and can continuously expand as long as a positive feedback loop is established. Past organizational cultures may now be transformed into goals and context, no longer requiring slogans, three-step meetings, icebreakers, or similar formalities.

Liu Ye: Culture is about managerial intent, not business intent. In the previous era, strategy began with vision; vision determined value, the organization aligned with strategy, business operations validated everything, and culture was merely a means of governing the organization—not directly serving strategy, and sometimes merely the founder’s personal preference.

Zhang Peng: In the past, there were many gaps in the process of serving customers— is AI eliminating these gaps?

Liu Ye: Yes, culture is no longer important in the AI era. Culture is the belief system that organizes human societies, but AI does not require it. AI has no physical form and does not need cultural guidance. The core requirement of AI is computing power.

Zhang Peng: You mean AI needs goals and principles. A single document is enough to clearly define those goals and principles, allowing all productivity units to immediately synchronize and faithfully execute them without deviation. A significant portion of friction in human organizations disappears.

Liu Ye: Yes. The original organizational structure was: Strategy → Culture → Talent → Execution. The current AI organization is: Goal → Principles → Skills → Orchestration. The entire management chain has been halved.

05 The Final Barrier: Aesthetics and Arrangement

Zhang Peng: What is the new barrier for enterprises? Talent quality has been replaced by Skill Set—if I have good taste, I can access the best skills from around the world. Then, one level above that is “orchestration,” correct? What changes will this bring about?

Liu Ye: Just as you can buy all electronic components in Huaqiangbei, why isn’t everyone able to create an Apple? The biography of Jobs defines aesthetics very clearly: having seen enough good things in the world to distinguish quality is aesthetics. If you’ve never encountered good products, good processes, or good organizations, you cannot produce outstanding results.

Zhang Peng: Experience is the foundation of aesthetic judgment.

Liu Ye: It’s just experience plus talent.

Zhang Peng: Aesthetics is expressed in two ways: one is active design and arrangement; the other is recognizing and selecting high-quality emergent elements within chaos. These two approaches are not mutually exclusive.

Liu Ye: Indeed, there is no conflict. Some of Apple’s achievements are developed in-house, while others come from acquiring third-party companies; the core is having good taste—you don’t need to reinvent the wheel, and you can develop independently when necessary.

Zhang Peng: The core issue is whether to let the agent confirm the path after running within the configured module, achieving emergent orchestration; or to predefine all paths for designed orchestration?

Liu Ye: Emergence is non-manipulative; it requires first establishing seed rules and principles, which reflects a person’s aesthetic. Just as an excellent engineer can create a functional Openclaw with 500 or 5,000 lines of code, while an incompetent one cannot achieve the same result even with 50,000 lines—the underlying seed rules still must be set by humans.

Zhang Peng: So, you can't wait for emergence to arise from chaos—that would take an extremely long time. Coordination remains crucial. Does this coordination ultimately have to come only from the founder, or is it more like that of a “producer”?

Liu Ye: I think this definition of a producer is excellent. Indeed, even with emergence and economies of scale, data labeling, data cleaning, and continuous algorithm alignment are still required to prevent uncontrolled growth.

The number of people required depends on business complexity—complex tasks cannot be handled by one person alone, such as filming short dramas or writing prompts, which present numerous practical challenges. The concept of a "one-person company" has been overused; the world cannot be infinitely simplified. Although a computer can be operated by one person, it is extremely difficult for an individual to master all high-level skills. Exceptional talents like Elon Musk or Fei-Fei Li, who excel across multiple domains and can take on any role, are exceedingly rare.

Zhang Peng: If we could access the world’s top agent and skill systems—such as an outstanding screenwriter—could we theoretically produce a globally renowned and profitable film? Although the screenwriter brings a core strength (a great script), they cannot handle all aspects of production. Is this “core strength + global resources”闭环 feasible?

Liu Ye: This is fundamentally a data issue—whether there exists data that stores high-dimensional information. For example, training a model on CEO skills is currently unsupported by sufficient data: Ren Zhengfei’s ten-thousand-word essay and Jack Ma’s oral statements cannot fully capture their high-dimensional cognition; even if we collected all global corporate financial reports and every statement made by CEOs, we still couldn’t train a model capable of performing as a CEO, because the core competencies of a CEO are tacit knowledge that cannot be fully revealed through text.

Zhang Peng: In other words, the CEO’s core capabilities cannot yet be vectorized. This constrains the ideal vision of a “one-person company”—even if everyone can leverage their strengths in a single dimension and access top global resources, there is still a lack of a central orchestrator; fundamentally, it’s an issue of orchestration capability. Ultimately, having the best “components” still requires strong orchestration ability.

Liu Ye: The same applies to product managers—their tacit knowledge cannot be fully documented. This is also the fundamental reason why current AI companions and AI-generated content lack vitality—they lack data support for high-dimensional tacit knowledge. When data volume is limited, focus on skills; when data volume is abundant, then build models. Robots currently cannot be effectively deployed, primarily due to insufficient data.

Zhang Peng: This suggests that the future competitive edge of companies will no longer be about access to top-tier models—initial AI resources appear uniform, and computing power is tied to financial resources and business closure capabilities. Ultimately, the key differentiator will still lie with the “producer” themselves: their ability to orchestrate and the innovation and significance of their goals, which together form a company’s core competitiveness.

Liu Ye: A former partner at McKinsey once told me that McKinsey’s core business is to extract best practices, develop models, and then assist companies in implementing them step by step. For example, when consulting for Chinese automobile manufacturers, we would consult with our Japanese colleagues to understand Toyota’s methods—essentially replicating and implementing best practices.

Mi Meng’s case in producing short videos is highly instructive. A graduate in Chinese literature, she assembled a core team composed of talent from the mathematics and computer science departments of Tsinghua and Peking Universities, specifically to deconstruct the logic behind viral short videos, ultimately achieving an exceptionally high rate of viral success. This approach essentially involves modeling a social engineering framework for the industry; even if overfitting is possible, the direction of the modeling is correct.

IBM, Accenture, and McKinsey all do this—McKinsey's first generation modeled best practices into their partners, while IBM transformed them into digital processes; at their core, both are essentially "selling management and processes."

Zhang Peng: The core is to distill best practices and repeatedly validate their implementation—this will be the key to success for future business organizations. Only by thoroughly breaking things down can you achieve efficient orchestration. So, your next core direction is to move forward with this approach?

Liu Ye: Over the past three years, we’ve primarily focused on AI-to-consumer (AI-to-C) services, rebuilding our entire teaching and curriculum development system using the MetaOrg approach. This is not simply a story about “using AI to improve efficiency.” We’ve built an entire agentic curriculum development organization, powered by virtual teams: a language learning research team tracking the latest theories in second language acquisition, a vertical corpus collection team gathering authentic expressions from real-world contexts, a dialogue evaluation team establishing multidimensional standards for assessing spoken proficiency, a dialogue design team translating pedagogical methods into natural human-AI interactions, a question container design team solving problems related to matching exercise formats with content, and a data analysis team uncovering genuine signals of learning outcomes from user behavior. Each team has its own skills, workflows, and evaluation criteria. Currently, around 80% of tasks—including textbook data labeling, monitoring and evaluation, user insights, and product iteration—are handled by AI.

Our development path is evolving from "AI as a feature" to "AI as an organizational capability." The English teacher role, which has moderate complexity, has already been abstracted and generated into other roles via MetaOrg; when combined with the latest skill architecture, it holds the potential to create even more advanced roles.

We have now completed the full development of the AI tutor, including the abstraction and engineering implementation of orchestration capabilities. In the future, it is highly likely that Meta tutor will evolve into a Meta organization, where the fundamental unit is the role rather than the individual, with the core focus on collaboration and management between roles. Our current priority is to connect with the top CEOs across industries, as they are the central “producers.”

Zhang Peng: So what you've launched is more like a scalable department?

Liu Ye: The goal is to move toward a corporate structure—large companies are essentially composed of multiple smaller units, with the smallest unit being the position. While it’s important to focus on industry-wide strategic choices, product iteration must also begin at the position level; if positions are not effectively managed, even the most capable leaders cannot build an efficient organization.

Zhang Peng: To build a strong department, first break down the relevant capabilities and roles, then further break down the skills required for each role, and strive to achieve SOTA levels in these skills.

Liu Ye: There is only one core approach: co-create with the most elite service recipients. The skills developed must be evaluated by top-tier enterprises to ensure they meet real needs—just as a subordinate’s proposal requires approval from a superior; you can’t just be self-satisfied. For example, when building models for short dramas, recognition from industry-leading institutions is essential; otherwise, it doesn’t qualify as truly top-tier. Everything must be assessed and measured.

Midjourney produces high-quality images because its team consists of photographers and engineers with exceptional visual aesthetics; LV’s image model, trained using Stable Diffusion, far surpasses ordinary models because LV possesses the world’s most advanced visual aesthetics and data. Clearly, evaluation capability is the core. To build an AI company, you must emulate IBM and Huawei—IBM mastered best practices in car manufacturing after serving top automakers and then exported those insights; Huawei spent 4 billion purchasing the IPD process, using it both for internal management and external licensing—this is the true competitive advantage.

Zhang Peng: Essentially, it’s about breaking down skills according to best practices, achieving SOTA for each skill, then elevating them to SOTA for roles and departments, and ultimately orchestrating them into SOTA for the business—this is a clear path to becoming top-tier in business. Another critical question: How do we keep skills up to date? Just as biological evolution drives variation in Earth’s biosphere, today’s SOTA may be rendered obsolete tomorrow—how do we respond to such change?

Liu Ye: The core logic aligns with human and biological evolution—perception, planning, action, and reflection. Maintain high talent density and cross-disciplinary attributes within the organization: one end connects with technological frontiers (researchers), the other explores business models, while co-creating with top industry clients to continuously evaluate and optimize in real-world scenarios. This is the only approach.

Zhang Peng: Conversely, systems built on best practices from top-tier companies can help mid-tier companies achieve leapfrog growth—but such systems are likely only accessible to companies with substantial resources and financial capacity, making them unaffordable for SMEs and young entrepreneurs. The consulting industry has evolved from traditional services into tool-based products. Are the opportunities for the new generation limited only to the skill level? How can disruptive innovation be achieved at the skill level to prevent the industry from falling into a “privilege cycle”?

Liu Ye: In the previous generation of SaaS, companies like Salesforce, Palantir, Notion, and Slack—some building general-purpose tools, others offering integrated services—demonstrate that young entrepreneurs still have opportunities: avoid areas where you lack competitive advantages, focus on universal skills, and find the right ecological niche. Notion is a classic example—it doesn’t engage in specific business processes but abstracts the function of text note-taking into a general-purpose tool. Ultimately, the world will be shaped by the division of labor among countless agents. Young people must first identify their ecological niche, then leverage their unique strengths, align with future trends, and avoid becoming adversaries of time. Over the past decade, the first generation of internet entrepreneurs were mostly returnees (leveraging cognitive advantages), the second generation were mostly programmers (leveraging tool-driven breakthroughs), and the third generation of industrial internet entrepreneurs were largely serial founders—the pattern is clear. Young people must understand the mid-game and their own strengths.

Zhang Peng: So you believe that localized innovation and optimization at the skill level have limited impact; the greatest opportunity for the new generation may lie in goal innovation—identifying newly emerging goals of the era, combining them with high-quality skills, and continuously evolving to build new systems and achieve breakthroughs on these new goals.

Liu Ye: The competition around skills is extremely subtle. Although skills are currently popular, if someone aligns with top human experts and creates superior skills, existing ones will be replaced. This brings us back to the issue of moats: early movers may not necessarily win in the end—they could very well become “nutrients” for a higher-dimensional competitor.

Zhang Peng: The fear is becoming merely a “loader,” only laying the groundwork for a more advanced opponent. If we merely optimize efficiency within existing goals, it’s meaningless—any efficiency advantage will eventually be erased. Therefore, for the new generation to achieve breakthroughs, they must create fundamental differences in their goals.

Liu Ye: That's right—failing to grow into a core force only nurtures higher-dimensional competitors. The essence of business is simple: clearly identify who your customers are, how to serve them, and how to make them indispensable to you. Any young person who cannot clearly define their customers cannot achieve optimization.

Zhang Peng: You also need to focus on incremental markets, as competition in existing markets is extremely difficult. If your business succeeds, it will elevate companies in the same field to the same advanced level—these companies already possess both wealth and expertise, making it very hard for young players to compete against them in established markets.

Liu Ye: In the previous generation of SaaS industries, the success of companies like Notion and Slack stemmed primarily from targeted differentiation.

In the early stages of the previous generation of SaaS, Chinese venture funds tended to invest in scientists. Later, they realized that scientists are better suited for collaboration and communication rather than entrepreneurship—the high-dimensional, low-competition environments scientists operate in differ fundamentally from the high-dimensional, high-competition logic of the business world. The higher the dimension of a field, the more difficult it is to transition into a new one, as the core thought patterns are entirely different. In the early stages of any field, competition is technological (low-dimensional, high-competition, with immature technology); once the technology matures, competition shifts to business (high-dimensional, high-competition, dominated by industry professionals, product managers, and business practitioners). For example, when the iPhone was first launched, most top-ranked apps were developed by programmers; a few years later, with the rise of industrial internet, all programmer-dominated apps on the rankings were replaced.

If the AI era follows the logic of the mobile internet era, Silicon Valley’s core strength will still lie with experienced professionals, much like how China’s industrial internet is largely driven by serial entrepreneurs. Young people’s opportunities still lie in identifying differentiated goals.