42 Judgments on AI's Evolution: Agents, Organizations, and the Physical World

I. AI Industry Just Entering the Agent Application Stage

Judgment 1: AI may bring a productivity revolution exceeding the Industrial Revolution. The interview frames AI not as a mere product upgrade after mobile internet, but as a shift in who can complete tasks and how tasks are completed. If AI capabilities continue on their current trajectory, the impact on industrial-era institutions could surpass the internet, rendering today's products, organizations, and experience less valuable. This is a forward-looking industry projection, not an accomplished fact.

Judgment 2: The AI industry will pass through three phases: infrastructure, application explosion, and native applications. Phase one: technology becomes widely callable infrastructure. Phase two: entrepreneurs use these capabilities to find new tasks and products. Phase three: after long exploration, native applications suited to the new world become mainstream. The second phase cannot be skipped; base technology provides capability but does not automatically tell entrepreneurs which problems to solve. The interview cites the evolution of Toutiao, recommendation algorithms, and short video to illustrate that new technology requires massive product trial-and-error to find its fitting content form and user demand.

Judgment 3: 2026 marks the basic completion of phase one and the opening of phase two. This milestone rests on intelligence becoming standardized and token emerging as a unit of intelligence measurement, allowing calculation of resource consumption per call. Products like OpenClaw show AI moving from answering questions to independently completing tasks, though they remain immature and not yet mass-market. Historically, a true browser-level product that lowers barriers for developers and users has not yet appeared; Yahoo- or Google-style platforms are still further out.

Judgment 4: Large models will ultimately become AI cloud infrastructure. In this framework, large-model companies' long-term position resembles AI cloud. As enterprises depend more on base intelligence, supply must be stable, costs controllable, and vendor lock-in avoided. Massive investment means the market cannot long sustain many homogeneous players, leading to an oligopoly with stronger regulation. Once models become infrastructure, customer evaluation shifts from capability breakthroughs to stability, price, and invocation experience.

Judgment 5: The window for general large-model entrepreneurship has closed. This refers specifically to creating another public base-intelligence company along existing lines. Leaders have accumulated model, compute, and organizational capabilities and are already using AI in model R&D itself. Improved training efficiency turns leads into compounding intelligence advantages. At the time of the interview, Anthropic and OpenAI were seen as leaders. However, bottom-layer innovation continues; New Lab's exploration of new technical routes and application-driven questions may still spawn new companies.

Judgment 6: Technical value does not automatically become corporate profit. A technology can create huge social value while the providing firm captures little profit. Once similar supply appears, customers compare costs and competition squeezes prices. Technical leadership, early adopter willingness, and short-term revenue growth do not alone prove long-term value. Entrepreneurs must judge substitutability and market expandability. Whether value stays in infrastructure or flows to applications affects profit capture.

Judgment 7: Today's model leaders may not be the ultimate winners of the AI-native era. The infrastructure phase relies on original research and breakthroughs; the application phase demands user understanding and new task/interaction modes. Capabilities differ per phase, and first-phase leadership rarely transfers naturally. Historical analogies (Yahoo, AOL, Google, BlackBerry) show first-wave firms can gain huge value yet lose central position when the industry advances. History cannot name specific winners but warns against treating current rankings as final.

Judgment 8: Infrastructure oversupply may lower application entrepreneurship costs. As long as intelligence is insufficiently strong, stable, or cheap, infrastructure investment continues. Cyclical building can create temporary supply exceeding demand, pressuring some investors but giving application builders cheaper, stronger, more stable base intelligence. Whether infrastructure investment pays off and whether the application ecosystem benefits are separate questions.

Judgment 9: New Lab explores both new technology and new organization. Some researchers see a ceiling in current technical routes; others want to apply mature large-model capabilities to fields like AI for Science. These teams explore model architectures while testing how researchers should collaborate. The interview views early-stage bubbles positively: before technical routes converge, no one can prejudge the correct path. Market willingness to fund potential returns gives more teams trial opportunities, and today's fringe directions may become tomorrow's consensus. As large models become infrastructure, the industry's core question shifts: what tasks should callable intelligence actually complete for users?

II. Product + AI Is a Necessary Condition; Complete Tasks Are the Watershed

Judgment 10: The essence of an Agent is independently completing tasks. The standard lies in action. The interview distinguishes chatbots from Agents by whether the system can finish a task around a goal without step-by-step human command. The internet let people share information; Agents and Skills let people share capabilities. Research and design methods can be encapsulated and invoked; programming and operations capabilities can be composed by intelligent systems. User concerns shift from "where to find information" to "who can help me get this done."

Judgment 11: The real opportunity lies in complete, complex, high-value tasks. When model capability is lacking, many apps plug gaps with patches and workflows; these are easily covered when models upgrade. Tools solving only a simple step hit market ceilings quickly. The interview uses advertising: generating a creative or executing a placement are partial steps; helping an enterprise achieve growth is a complete task. Only important tasks give enterprises room to accumulate context and feedback, gradually forming proprietary technology. The watershed is whether the product can complete a whole high-value task for the user. Evaluating completeness looks at what the user ultimately receives, who bears responsibility for failure, and whether the system grows stronger from real feedback. A single impressive generation is not yet a task the user trusts.

Judgment 12: Applications must enter scenarios that can land today and scale tomorrow. The interview explicitly mentions education, health, enterprise growth, and organizational brains as worthy exploration directions—not validated track lists. Scenario selection requires two conditions: real demand and payment reason today, plus expansion space when capability improves. Entrepreneurs can start from a small entry point, but it must lead to more complex tasks and not lose its raison d'être after the next model upgrade.

Judgment 13: Application doldrums do not mean the opportunity has ended. The previous batch of simple apps was quickly covered by models; new complex scenarios are not yet fully proven, leaving entrepreneurs and investors temporarily without success samples. This period is seen as a counter-consensus exploration phase. OpenClaw showed geeks and some managers Agent potential, but deployment and usage barriers kept it from the mass market—it is more a directional drill than a mature product. Entrepreneurs cannot wait on the sidelines for models to strengthen; task boundaries, user trust, and real feedback are only earned in practice. Without such accumulation, they cannot catch the opportunity when base capability leaps again.

Judgment 14: Short-term ARR cannot prove long-term room. New tech always attracts early adopters willing to try and pay. Revenue is a starting point, not proof of a lasting position. Annualizing one month's revenue and extrapolating current growth ignores market ceilings and model substitution risk. The framework values whether users perceive a clear experience change, whether the product spreads organically, and where it can grow in two years. ARR is an operating result, not a substitute for product-space judgment.

Judgment 15: Agent companies still need their own technical accumulation. Application firms need not retrain general base models for society, but must not become mere wrappers around model APIs. Context management, long-term memory, interaction modes, and task-specific learning mechanisms offer innovation space. Whether to train models, and which parts, depends on the task. If product value comes only from temporarily patching base-model weaknesses, it will not survive model upgrades.

Judgment 16: The data flywheel upgrades to intelligence compounding. Agents invoke base intelligence to complete tasks, gain feedback from real work, use feedback to improve algorithms and task flows, and accumulate new data—forming intelligence compounding. This loop requires the Agent to participate in execution; if the system only generates suggestions and humans do all the work, complete outcome feedback is hard to capture. Extending capability from one scenario to others, taking on more complex tasks, expands the moat. This extension is summarized as a "black hole effect": scenario diversity and new connections between capabilities deepen competitive barriers; same-type data volume is only one part.

Judgment 17: The next-generation entry point must solve standards and trust. As Agent numbers grow, users need to judge which system can complete a task, whether it can access data and funds, and what the charging and failure-handling rules are. Choosing an Agent costs far more than opening a webpage. The new entry point will connect developers and users, first establishing standards and evaluation mechanisms, then offering recommendations. Platform builders typically need to first create a successful Agent, understand how tasks run, before gradually forming an ecosystem and de facto standards. The interview predicts a browser-like key opportunity may appear within two years, and a clearer Agent entry point within three to five years—this is a temporal projection from the interview time, not a fixed product roadmap.

Judgment 18: Agent OS belongs to the more distant native-application stage. Agent entry points help users pick capabilities; Agent OS further takes over task scheduling. Users only express desired results; the system decomposes tasks and composes Agents, deciding at critical nodes whether human involvement is needed. This form requires sufficiently mature Agent supply, reliable invocation, and more complex collaboration needs. Building a grand OS before the ecosystem exists struggles to solve real problems. The ultimate direction helps entrepreneurs judge what to accumulate today, but cannot replace immediate product practice. Phase two first needs large numbers of task-completing Agents, repeatedly used and verified, before unified scheduling has a realistic foundation.

III. AI Transformation Ultimately Enters Organizations, Not Just Tools

Judgment 19: Future organizations will shift from role-centric to task-centric. Traditional companies set roles first, then assign responsibilities and reporting lines; tasks are allocated to fixed positions. AI-native organizations first clarify what needs solving, then call suitable humans and Agents to complete it together. Power relationships change accordingly: seniority no longer guarantees the best fit for the current problem; judgment flows to those holding key information and relevant capabilities. Division of labor persists, but collaboration basis shifts from administrative rank to task needs.

Judgment 20: Adding AI to old processes is only a transitional solution. Existing enterprise workflows are designed around department boundaries and human capability/management constraints. Inserting generation and summarization functions and connecting automated execution steps improves local efficiency but does not automatically change how tasks are decomposed. AI transformation cannot start with tool procurement; it must first re-decompose tasks and organizational collaboration. Enterprises must decide which links require human judgment and which can be continuously executed by AI. Responsibility boundaries must change accordingly. Without this restructuring, more tools may only make old processes more complex.

Judgment 21: The one-person company is only a transitional form. A one-person company is not the organizational endgame. AI lets one person do work that previously required multiple people, but individual time and capability still have limits; complex tasks still need division and cooperation. A more common future form may be a small team of more well-rounded individuals with complementary skills, each leveraging Agents to shoulder a broader task range. The author's five-person project exemplifies this: it proves small-team capability boundaries are expanding, not that everything should be done by one person.

Judgment 22: Contribution and delivery will matter more than titles. As role hierarchies weaken, promotion is no longer the sole incentive. The interview emphasizes subjective agency: proactively finding problems, claiming tasks, and owning results. Task allocation and completion make individual contribution more transparent. The market calls for capability; teams need delivery; administrative titles can replace neither. Expanded responsibility brings freedom, but many problems can no longer wait for superiors to break down and assign.

Judgment 23: Shared context will replace part of process control. The premise for autonomous decision-making is knowing shared goals, existing constraints, and why the team made current choices. This is summarized as "context replacing control." AI also needs access to the same information: it must understand not just immediate instructions but the task's relation to long-term goals. Further extrapolation suggests this capability could let OKRs land in a new way: systems continuously observe whether tasks deviate from goals and alert teams. This remains future projection. AI can help preserve and sync context, but whether the goal itself is worth pursuing still requires organizational judgment.

Judgment 24: CEOs must shift from controllers to servants. When organizations rely on a few high-agency talents, managers can no longer sustain cooperation through authority alone. CEOs must attract people suited to current tasks and provide them with ample information and creative space. Repetitive, tedious work and sustained high pressure consume creativity. As AI takes on part of execution, managerial value shifts to removing obstacles and maintaining goals; process approvals should decrease. This does not cancel management requirements: the more important the task, the clearer goals, authority, and result standards must be set in advance—management just no longer relies on watching every action.

Judgment 25: Organization and strategy must change together. The old method was to set strategy first, then adjust organization to fit. In the AI era, decision frequency is higher; new strategy-generation modes cannot run in old organizations. Organizations need to let humans and AI continuously process problems, forming new judgments in shared information. Strategy then decides what capabilities are needed next and how the organization should further adjust. The two become a mutually reinforcing loop.

Judgment 26: The application era will re-amplify the product manager's value. The infrastructure phase relied on scientists for technical breakthroughs. Entering the application phase, enterprises must judge which users to serve, which problems to solve, and which technologies deserve investment. AI product managers are shifting from requirements-document producers to problem definers, solution judges, and delivery acceptors. PRDs and prototypes can be generated faster; the saved time must not be spent producing more documents. PMs must judge whether a task is worth doing, then verify whether delivery actually solves the user problem. When models are unstable, they must also establish matching acceptance criteria.

Judgment 27: Excellence requires mission and complementary teams. The interview distinguishes excellent from exceptional. Excellence can advance on continuous positive feedback; exceptional often persists in a counter-consensus direction when the outside world disagrees, eventually proven by results. Such persistence cannot rely solely on personal desire to prove oneself. Founders must put the mission before self and be willing to hear different judgments. One person watches direction, another excels at execution; complementarity is more realistic than forging an all-round leader. Mission cannot stay a slogan written for employees. When financing and growth change, it must still help the team make trade-offs, not be rewritten by external evaluations at any time.

IV. The Stronger the Uncertainty, the More Strategy Cannot Rely Solely on Annual Planning

Judgment 28: The stronger the uncertainty, the more important strategic judgment becomes. In relatively stable industries, firms can predict the future from the past and break long-term goals into annual plans. AI is different: model capabilities may leap suddenly, new products rapidly shift user expectations, and previously valuable businesses can be impacted within months. Studying business history aims to extract reusable laws of industrial evolution, helping entrepreneurs locate their stage. Mechanically mapping today's company to past winners misses the point. The interview explicitly avoids short-term predictions. Laws provide coordinates but cannot calculate a definite route. Enterprises need continuous judgment and readiness to correct course. A choice must be made, but no single choice can be a fixed answer for years.

Judgment 29: Strategy must use a mid-term path to connect vision and action. The interview summarizes the strategic method as "look ten years, think three years, do one year," noting the middle "think three years" is most easily ignored. Long-term vision answers where the enterprise wants to go; current action answers what to do today. A bridge is missing in between. Without a mid-term path, vision becomes a slogan and annual plans degrade into linear extensions of old business. Mid-term thinking requires the team to see at least the next two or three steps: what to verify first, how subsequent milestones unfold, and which step this year's resources serve. Conflicts arise among short-, mid-, and long-term; strategy's value lies in making trade-offs within these conflicts. Ten, three, and one year are not fixed cycles; different industries can use different time scales. The rhythm can change, but the middle cannot be empty.

Judgment 30: In the exploration phase, raise trial-and-error efficiency first; pursue scale only after direction converges. When judging what strategic method a firm should adopt, the interview weighs its development stage more than its category (model, application, or robotics company). In the strategic exploration phase, direction is undetermined. The number-one leader must take the team to try different possibilities, lower trial costs, and gradually form consensus from feedback. The most dangerous move at this stage is using strong execution to scale an unverified direction. After direction gradually converges, the focus shifts to building a sustainable business model, improving organizational efficiency, and driving scale. Stabilize the prototype first; only then can repeated investment bring growth. When the external environment fundamentally changes again, mature firms must return to exploration mode. Traditional enterprises cannot manage continuously changing AI capabilities with a single project approval and annual budget. Roadmaps should record the current most-worth-verifying hypotheses and update with evidence, not become unchangeable commitments.

Judgment 31: Counter-consensus does not guarantee success; the scarce skill is discerning signals and continuous correction. Successful firms often make counter-consensus choices, but opposing consensus does not automatically bring success. Many seemingly bold directions fail; studying only survivors easily reverse-engineers outcomes into methods. Discerning signals requires expanding perception of the real world and thinking problems more deeply. Too narrow contact leaves judgment stuck in local optima. By the time a trend heats up, latecomers often see only the leader's step-three results, missing the cognitive accumulation of steps one and two. Past training builds capability but also creates path dependence. Founders must believe their direction while reserving reality's right to overturn it. When new external evidence conflicts with original judgment, facts come first. The author's view: AI-era strategy cannot stay a static annual document written once and awaited for execution; it must become a continuous verification and correction mechanism.

V. Robotics Will Open the Physical World; Old Giants Have No Safe Zone

Judgment 32: Robotics is still in early exploration; two routes may both succeed. The robotics industry is analogized to the large-model stage before ChatGPT. The industry is still searching for landing modes; routes are far from converged. A real constraint is the lack of massive ready-made data like internet text. One route pursues a general robot brain, seeking cross-scenario generalization. The other enters specific scenarios first, letting robots gain data in actual work to improve models. Each has difficulties: the general route must prove generalized capability is good enough in large scenarios; the scenario route must prove it can escape narrow demand. The interview uses the metaphor of climbing a mountain's south slope vs. north slope: at departure, no one can tell who will reach the summit first.

Judgment 33: Industrial breakthrough depends on affordable mass production. Demos are just the starting point. Robots have a long way to become mass-market products. The interview analogizes to Ford's Model T and assembly line: the industry's key node comes from stable delivery and cost decline, enabling large numbers of users to afford and use them. A phased milestone: whoever can actually sell 10,000 robots has achieved a breakthrough worth noting. This describes a future threshold, not a claim that any firm had reached it at interview time. After market demand diversifies, single standard products will face new competition.

Judgment 34: Robots are the interface for AI entering the physical world. The significance of robots can be analogized to electrical appliances. After electricity entered homes and factories, it didn't produce just one device type; intelligence entering the physical world may spawn multiple major categories. Home-service robots and companion robots may take different forms; industrial scenarios will yield another batch. Many smaller-scale hardware products will also exist. This poses entrepreneurs a question more practical than technology: usefulness proves demand exists, but market capacity and technical barriers must still support a company's long-term development.

Judgment 35: "All in AI" is only a necessary condition for giants to stay at the table. The interview judges that no giant is absolutely safe before AI. Large companies investing in models, talent, and compute only earn the right to participate in the next round of competition, not a guarantee of remaining the next generation's leader. Old organizations also create resistance. Transforming a 100,000-person company into an AI-native organization is vastly harder than letting a 10-person team grow gradually in the new way. Mature companies' resources are advantages, but old processes and vested interests limit the speed of directional change. Adding AI or Agent features to old products is seen as a transitional solution. Old products can persist, but new entry points require the enterprise to re-understand how users complete tasks. Existing user scale cannot automatically substitute for this cognitive layer.

Judgment 36: Infrastructure advantage does not automatically become the next-generation consumer entry point. The interview treats model/cloud capability and consumer-grade entry as two distinct competitions. Microsoft's cloud accumulation and early OpenAI investment, Google's chip and model capabilities help them compete in AI infrastructure, but cannot directly answer who will own the next user entry point. The same logic applies to ByteDance: its model capability and organizational resources give it a strong chance to become a major AI cloud company, but past product success cannot guarantee it naturally becomes the next era's consumer entry. At interview time, Doubao was still classified as a first-generation chat-style AI product because it had not yet met the requirement of Agents independently completing tasks. This is a time-conditioned product judgment; subsequent capability changes need separate observation.

Judgment 37: Agents will change the relationship structure of social and transaction. Tencent owns the human relationship network; Alibaba owns the human-goods connection. Both advantages may be challenged by Agents. When Agents start finding services and executing tasks on users' behalf, the acting entities connected by platforms change. Old platforms' accumulated relationships retain value, but how entry points form and where value flows may be redefined. The interview does not give a concrete picture of next-gen social or shopping products; it raises structural risk: neither social nor transaction lies outside AI's transformation.

Judgment 38: Chinese AI competition more easily trends toward homogenization. This framework believes China and the US will form not entirely identical AI ecosystems. US early movers had longer independent exploration periods, more easily forming their own technical routes and organizational characteristics. Many Chinese teams see similar overseas prototypes and follow simultaneously, competing in similar talent markets, resulting in large numbers of companies crowding into the same direction, so the industrial landscape takes longer to converge. A half-year or one-year lead is not just a time advantage; earlier entry into real problems gives teams a chance to form different product cognition and practical accumulation. Catching up to public capabilities only gets an entry ticket; understanding user tasks may open the next gap.

VI. When Intelligence Is Callable, Human Value Returns to Creation

Judgment 39: Human value will shift more toward original capability and complex problem definition. The AI era is summarized as the creativity era. When knowledge and base intelligence can be called externally, structurally processable work will increasingly go to AI; humans must establish new value beyond ready-made answers. In this definition, writing and drawing are only part of creativity. Originally defining complex problems and seeing unmet needs also count as creation. Emotional companionship and new human-to-human interactions may generate previously unsatisfied value. The role of experience also changes: once experience that can be structured into knowledge is absorbed by models, senior practitioners can no longer rely on years of service alone for advantage. Younger people can enter unfamiliar fields faster; experienced people must keep updating cognition. This connects directly to the AI product manager's evolution: document generation and information organization can be handed to AI; deciding which problems are worth solving and which results are acceptable still requires human ownership.

Judgment 40: Education must shift from knowledge instillation to interest and creativity. Per the interview, the current education system is built mainly around knowledge transmission, with interest and potential development long secondary. When knowledge acquisition changes, original cultivation goals are impacted. The contradiction has appeared: youth feel old learning-to-employment paths shifting, but parents struggle to let children leave familiar evaluation systems. Answers have not yet emerged. Even knowing education must change, people lack verified new methods. Education is seen as one of the systems where conflict will concentrate next; society needs to invest more resources to seriously find feasible methods. The interview does not provide a mature answer; this boundary is important.

Judgment 41: Society digesting the AI transition will take at least a generation. The interview estimates social adaptation to the AI revolution needs a generation; twenty years is just the start. This number describes a generational-replacement scale, not a precisely executable timetable. Those truly adapted to the new world may grow up treating AI as part of the environment. Office and virtual space relationships will change accordingly; people will maintain more continuous collaboration with intelligence via wearables. These projections stay within the currently discussable AGI scope. For how superintelligence will ultimately develop, the interview explicitly states there is no consensus, so no further conclusions are given.

Judgment 42: AI can expand freedom, but humans must still rearrange their own rhythm. The interview's overall attitude toward AI leans optimistic. It believes technology will continuously increase what people can do, and more intelligent solutions may let people invest energy in more interesting, growth-bringing activities. It does not assert the new civilization will make everyone happier. General living conditions are measurable, but philosophical happiness is hard to prove directly via technological progress. Another contradiction has emerged: AI processes tasks so fast that humans, trying to keep up, become more exhausted. Speed is not always better. The interview therefore proposes that humans retain appropriate isolation, find a rhythm their biological individuality can bear, and establish a new balance with AI.