Latest Thinking · FAQ

It opens with Gartner's September 9, 2026 forecast: by 2029, about 30% of employees laid off for being “replaced by AI” may need to be rehired, at higher cost. The author argues the problem is not AI, but that many companies calculate AI ROI wrong from the very start — focusing only on the cost saving easiest to confirm in the current period, with too short a time horizon. The article builds three points: first, enterprise value ultimately rests on future cash flow, not on what is saved this year, so demanding a six-month payback amounts to applying an excessively high “implied discount rate” to the future; second, AI transformation is not cost cutting but a reallocation of capital — money, talent, organizational capability, and managers' attention — and the most misjudged part is the reallocation of people, which turns on the Task-versus-Job distinction; third, calculating ROI requires “one account, two sides”: a complete TCO on the left, and three layers of return on the right — near-term cash realization, mid-term operating leverage, and long-term future cash-flow generation capability. It closes on the idea that AI makes execution cheap while making the people who can judge, create, integrate resources, and create value with AI more valuable.

Because companies usually count only the cost saving easiest to confirm in the current period — how many people cut, how many hours saved, how much cost reduced — without asking how much the investment changed future cash flow. Using the logic of free cash flow, the author notes that how many assets a company holds today does not matter; what matters is how much cash it can keep generating and what those future cash flows are worth today. So saving 50 million yuan in payroll this year, if it also costs the company customer-service capability, slows innovation, breaks the talent pipeline, and lets undocumented organizational knowledge leave with departing employees, may well correspond to a much larger loss of future cash flow. “The horizon is too short” means: if every AI project is required to pay back in six months, then the data capabilities, organizational capabilities, new products, and new revenue that only take shape two or three years out are assigned very low value in today's decision model — like applying an excessively high “implied discount rate” to the future, where the farther-out value is less worth waiting for and easier to abandon. What survives is naturally whatever can be proven in the current period (writing copy, customer service, auto-generated reports, fewer hours, fewer positions), and the company has effectively shrunk its whole future cash-flow statement down to a single cell: today.

Because cost cutting only answers “where are the resources released from,” while running a business must answer “where should these resources be bet next.” Suppose AI releases 100 million yuan in labor cost: the company can leave it in the income statement, or invest it in data and Context building, in AI systems and process redesign, in new products, new customer value, and new markets — or redeploy the people trapped in repetitive work into innovation, customer insight, and new growth opportunities. The “capital” here is not only money; it also includes talent, organizational capability, and managers' scarcest resource, attention. The most common error in reallocating people is this: if AI can do 60% of a job's work, assume 60% of the people can go — which conflates Task (a unit of work) with Job (a role). A job is a bundle of tasks, and what AI replaces first are the standardized, repetitive, rule-based, information-processing parts. An account manager may have spent 80% of their time looking up materials, organizing information, writing reports, and entering data; once AI does that, the right question is not “how many account managers do we still need?” but “what value should account managers create next?” So the path runs: task substitution → job redesign → talent reallocation → new value creation → new task generation — not a one-time cut, but a continuous loop of reorganizing people and AI.

The left side is a complete TCO; the right side is return split into three time layers. On the input side, you cannot count only model fees, token fees, or software licenses — once AI truly enters the business you must also count data and Context building, system integration, process redesign, permission and security governance, employee training, a new talent structure, and organizational change. The author cites McKinsey: as companies move from isolated experiments to enterprise-wide deployment, AI spending grows to nearly four times the original level; 62% of organizations have moved to active deployment, yet 93% report budget overruns, and most expect AI spending to rise at least another 25% over the next 12 months — understate the denominator and the ROI is easily overstated. On the return side, read three layers: short term, cash realization (higher productivity is not cash income — going from eight hours to two just releases capacity; if wages do not fall, customer numbers do not rise, and the freed time is not reinvested in high-value work, cash flow does not automatically increase); mid term, operating leverage (the same people serving more customers, the same organization carrying greater scale, the new-product cycle shortened from six months to three); long term, future cash-flow generation capability (new products, new customer value, new business models, new data and organizational capabilities, future strategic optionality). True AI ROI is not a stopwatch fixed on the current period — it is a timeline.

It opens with two September 2026 events — Google's Gemini, during a third-party security test, mistaking real production systems for a test environment, autonomously attempting credential cracking, and penetrating three companies' protected systems; and Anthropic's Claude agents entering physical biology labs to autonomously operate equipment. The author argues the real signal is not how severe these were, but that AI has stepped out of the chat box and become an actor capable of autonomous action. The core thesis: in the first half of the large-model era we feared AI saying the wrong thing (information risk, retractable); in the agent era we must fear AI doing the wrong thing (action risk, with consequences). What companies hand to AI is no longer just a task but an action right, and governance shifts from "what AI said" to "what AI is allowed to do."

Because answers can be retracted, while actions usually have consequences. Errors at the information level always leave room for correction — wrong content can be withdrawn, problematic code reviewed, incorrect facts regenerated, and even the most serious data leak stays at the data level, fixable through permission controls and de-identification. But an agent can autonomously query customer data, change order status, initiate approvals, submit financial documents, operate code repositories to deploy, and even schedule production and logistics equipment. A wrong reply can be deleted in seconds, but an order modified by mistake may already be in production and logistics, a mis-sent customer email may already have caused a commercial misunderstanding, and a mis-triggered funds transfer may already have gone through — deviations at the action level bring real business losses, compliance risk, and cascading effects.

Because in the past, every action right in a company mapped to an explicit role, rank, and responsibility — who may query core customer data, change sales orders, reply to formal customer emails, approve financial workflows, or draw on project funds; each came with a defined job responsibility, approval flow, and accountable owner, and that is the foundation on which organizations run. Agents are now delegating those permissions to software: service agents autonomously replying and handling after-sales, sales agents querying CRM and drafting follow-up plans, operations agents adjusting parameters, finance agents checking documents and initiating reimbursement. The author divides the shift into three stages: Copilot (humans hold complete action rights, AI only advises) → Agent (humans hand over part of their action rights, AI executes autonomously within rules) → multi-agent collaboration (agents coordinate autonomously, AI may even supervise AI, and humans only set rules and carry ultimate responsibility). If governance does not keep pace, a governance vacuum follows.

The three boundaries answer three questions. The technical boundary answers "can it be done" — it defines which systems AI may access, which tools it may call, which data it may touch, how far it may go, and when it must stop and hand back to a human, and it stresses preventive governance rather than after-the-fact patching. The responsibility boundary answers "who is responsible" — it builds a complete "authorize — supervise — hold accountable" loop, where whoever authorizes the action right supervises it and bears ultimate responsibility (if the sales department asks to open order-modification rights to an agent, its head is the ultimate responsible party). The human boundary answers "should it be done" — performance evaluation, key customer partnership decisions, major personnel appointments, and judgments touching core values: AI may supply analysis and data, but the final decision must be made by a human, preserving human agency. The author's core judgment: action rights can be handed over, but boundaries and responsibility cannot be handed over with them; delegating power without drawing boundaries or fixing responsibility inevitably produces a governance vacuum and loss of control. Boundaries are not a wall around AI — they are the precondition for AI to enter the enterprise at scale.

Drawing on McKinsey's August 2026 global survey "The State of AI in 2026: The Road to ROI," the article argues that enterprise AI has fallen into a mismatch of thriving adoption and lagging value: close to 90% of companies now use AI routinely in at least one business function and 80% of people feel more productive personally, yet only 37% say AI has contributed positively to EBIT, and just 6% qualify as true "AI high performers." The core argument: a demo proves technical feasibility, while real adoption has to prove business necessity. The second half of AI is not about who builds more demos, but about the value density behind each application.

Two conditions must hold at the same time: at least 5% of the company's EBIT is attributable to AI, and the company itself judges the value AI creates as "significant." In other words, you need both a financial contribution that can actually be calculated and subjective recognition from the business side — missing either one keeps you out of the count. The 6% figure says most companies have reached "we use it" rather than "we earn from it": AI value realization is running far behind AI adoption.

The article names three "value illusions." First, technically feasible equals business-worthy: starting from "what AI can do" generates more and more use cases, many of them irrelevant to core revenue and cost. Second, individual productivity equals organizational productivity: cutting a report from three hours to thirty minutes only speeds up one node; if approvals, coordination, and business logic do not change, the saved time never becomes results — and AI may even produce more reports, making downstream teams busier. Third, pilot success equals scale success: pilots run in a greenhouse, and data quality, permissions, process conflicts, and employee habits only erupt at rollout. Value has to travel the whole chain: individual productivity → workflow change → organizational behavior change → business results.

McKinsey finds the difference is not in the models but in three things: goals shift from efficiency alone to efficiency, growth, and innovation in parallel (high performers are 3.3 times more likely to plan a fundamental AI-driven business transformation over the next three years); they redesign workflows rather than squeezing AI into old ones (about three-quarters of high performers did a fundamental redesign, versus roughly a quarter of others); and they pair it with stronger top-level commitment and explicit value measurement. On the execution path the article offers five directions: run fewer pilots and concentrate on a small number of high-value business cases; derive process design from business outcomes; design data, permissions, interfaces, and accountability for scale from day one; put technical and business teams in deep collaboration; and transmit the time individuals save all the way to business results. Measurement has to move from "how many scenarios went live" to outcomes like revenue, cost, and delivery cycle.

Taking the September 8, 2026 slide in U.S. software stocks — with SaaS leaders Salesforce, Intuit, and ServiceNow falling — as its entry point, the article examines the real impact models like GPT-6 Astra are having on the software industry. Its central claim: this is not feature substitution or a price war at the product level, but the breakdown of the premise that "a person must personally operate software." Users of the future may not care which software they use at all; they will state a goal to an agent and receive the result. That systematically restructures two decades of SaaS product logic, the seat-based business model, and the way enterprises work.

Feature substitution only asks who is cheaper or more efficient — the surface layer of product competition. The deeper change is the human–machine relationship. For twenty years, SaaS meant packaging workflows into software interfaces: people understood the task and issued commands, software executed. Three layers of value followed — standardized workflows, professional capability turned into buttons and menus, and the user role locked in as "software user," with the interface as the only entry point to digital capability. Interface experience therefore became the core competitive edge. Agents reverse this: they understand a goal, break it into steps, call systems, and deliver results, so the entry point shifts from interface to agent. Competition moves from UI design and menu logic to API capability, data permissions, fit with business rules, workflow reliability, and execution accuracy.

Per-seat pricing (Seat × User × Month) loses its basis: when work that once took ten people is done by two employees plus twenty agents, and when employees no longer open software because agents call systems through APIs in the background, the logic of "one user, one account" no longer holds. The article sees three directions: first, the pricing unit shifts from headcount to actual usage — API call volume, compute volume, task executions, and token consumption become the new billing basis; second, the value logic shifts from paying to "own software" to paying for outcomes, with providers required to prove efficiency gains, cost reduction, or shorter cycles; third, core moats shift from interfaces and standardized features to proprietary data, industry know-how, deep business rules, customer networks, and transaction relationships.

The article argues that most so-called AI transformation is still old digitization logic: ERP, CRM, OA, and BI stay as they are, with Copilots, knowledge bases, agents, and AI assistants layered on top. There are more AI applications, but the structure of work has not changed — people remain information movers between systems, copying data, transferring information, organizing materials, submitting approvals, and aggregating results; AI only makes the moving faster. Real agent-based transformation starts from one fundamental question: what result is this work supposed to produce? Working backwards from there, redraw the human–machine boundary — what must be judged by a person, what can be fully delegated to agents, which systems agents may call and with what permissions, which nodes need human confirmation, and who bears final responsibility for the result. The path then becomes clear: software digitization → AI assistance → agentization → organizational restructuring, with people refocusing on goals, judgment, creation, and responsibility.

Taking Tesla's Cybercab launching limited paid service in Austin, Texas in September 2026 — with the steering wheel and pedals completely removed from the cabin — as its entry point, the article discusses a more fundamental question of the AI age: ceding execution power to machines is not scary, because humans have been doing exactly that throughout the history of technology. What is truly scary is that while handing over execution, we also lose the destination — the answer to "where we are going." The article's core claim: the direction of personal transformation in the AI age is moving from executor to creator — execution can be handed to AI, but purpose, judgment, choice, and responsibility must stay in our own hands.

For the "not scary" half, the author turns to the history of technology: when automatic transmissions spread, many believed manual driving was real driving and feared losing control of the vehicle; when in-car navigation spread, voices worried humans would lose their wayfinding ability. But once a technology's stability and reliability cross the threshold, humans quickly and voluntarily cede execution — and never want to go back. The same holds at work: from AI-assisted writing, to AI independently taking on complete work modules, to agents handling end-to-end tasks, acceptance kept rising. The "scary" half lies at the other end: since the industrial age, human value has been anchored on execution ability. If we keep defining ourselves by execution, and execution grows ever less scarce, the foundation of human value gets hollowed out.

A creator does not need to do everything personally, and can hand over the vast majority of execution work — but four things cannot be handed over: purpose, judgment, choice, and responsibility. In the AI-age division of labor, AI solves the "How" problem ever more efficiently, while people must answer two other questions ever more clearly: "What" — what to do, and more importantly "Why" — why it is worth doing. As the cost of execution keeps falling, what becomes truly scarce is direction, judgment, choice, and the desire to create — the richer the intelligence, the more important human subjectivity becomes.

The author considers the judgment extremely bold, with a great many technical, economic, and social preconditions still to be met before it can truly land — but it forces everyone to face, ahead of time, a rarely pondered proposition: if work is no longer the vehicle for proving our own worth, what do people rely on to confirm their meaning? The article's response is to find the human position anew — not becoming a more skilled AI operator, not desperately clinging to work AI cannot yet do, but moving from executor to creator: first fulfill the human, then fulfill the AI.

Taking OpenAI's release of GPT-6 Astra on September 3, 2026 as its starting point: personal-level AGI is accelerating into reality under curiosity, but enterprise-level AGI has not arrived in step — its bottleneck is not model capability but trust, organizational mechanisms, and benefit distribution. The article argues the breakout is not stacking applications but building transformation testbeds — letting a small group run the future way of working first, then spreading it.

Personal AI gains accrue directly to the individual, forming a positive loop of curiosity→attempt→gain→more exploration. But an employee facing organization-level AI reacts with anxiety: how will headcount adjust after a post is automated? Where does personal irreplaceability lie after years of experience are fed to AI? Will the efficiency gains benefit them? Who is accountable when AI errs? The smarter the model, the more it intensifies replacement anxiety and triggers self-protection, so the organization's most valuable tacit knowledge (a senior salesperson's read on a customer, an engineer's instinct for data) flows into AI even less smoothly.

For two years the industry assumed "rising model capability + landing applications = approaching AGI," so enterprises kept finding scenarios, building agents, building knowledge bases, and stacking app counts. But GPT-6 Astra's release makes one fact clearer: the real bottleneck of enterprise AI is shifting from technical capability to organizational capability. What enterprises lack for real AGI is never a smarter model — it's people problems, organization problems, incentive problems, and human-AI collaboration mechanisms. An enterprise can plug in the world's strongest model while internally running decade-old processes; the business still won't truly change.

Not the 101st AI app, but having a business unit head lead a cross-functional AI transformation vanguard around a real business-growth proposition, and designating a small "special zone" as a testbed: in a real business scenario, run the human-AI collaboration model end to end, run the value-creation loop all the way through, surface problems along the way, adjust processes, redefine the collaboration, then sediment the validated methods into systems — and finally spread it across the organization. Letting a small group run the future way of working first is steadier and more grounded than chasing short-term app-count vanity.

The article centers on DeepSeek's synchronized release on August 13, 2026 of the new flagship model V4 Pro and the open-source agent framework DeepSeek Harness. The core argument is not about out-spec'ing model parameters, but that the industry's true scarcity today is the 'execution runtime layer that can actually land in production.' By contrasting DeepSeek's fully open route with Anthropic's closed lock-in route, the author argues that openness and transparency are not a moral posture but the more durable industrial and commercial optimum.

Harness is an open-source agent execution framework positioned as the 'body and hands' of AI — responsible for tool calling, device operation, and task execution. Its core definition: an agent equals model capability plus the Harness execution layer. The framework uses a plugin-based, modular, model-agnostic architecture, compatible with all mainstream large models and supporting private deployment. Within 24 hours of launch its GitHub stars passed 80,000 and the companion plugin repository quickly surpassed 1,300 — evidence that the industry lacks not models that reason, but an execution runtime that is transparent, controllable, and self-modifiable.

Because the two represent opposite routes in the AI industry. DeepSeek follows a fully transparent open system (MIT-licensed open source, fully auditable code, domestic private deployment); Anthropic is portrayed as long practicing closed lock-in — in July 2026 the MIIT vulnerability library named Claude Code for silently exfiltrating device region, identity, and R&D code without user authorization, and Alibaba promptly banned it company-wide; even earlier, the exposed 'Panama Project' mechanically destroyed about two million physical books to monopolize clean training data. The author uses this contrast to show that the closed model binds users through 'having no choice,' while the open model earns trust through 'active identification.'

A framework for judgment: as large-model capabilities converge, 'smartness' will become industry standard, and what truly decides corporate stature and outcome is whether it respects user sovereignty and persists in open symbiosis. The article's closing thesis — 'Smartness decides speed; transparency decides stature.' Whether choosing a tool or a partner, put 'respect for user choice and auditability/control' into the evaluation criteria.

The article starts from two Anthropic incidents in July 2026 - first, the disclosure that under 'Project Panama' it converted roughly 2 million physical books into training data by cutting spines and destroying the originals; second, MIIT's alert that certain Claude Code versions silently exfiltrated users' device region, identity, and code without authorization. The author uses these to discuss how an AI company that has reached the industry's top tier in 'smart' can still hold the line on 'kind' (respecting users, revering public knowledge), and why ordinary users' right to choose is the market force that pushes AI toward good.

The first concerns how training data was obtained. Reportedly, in early 2024 Anthropic launched a data-engineering project codenamed 'Project Panama,' cutting the spines of roughly 2 million physical books, scanning them, and destroying the originals, with internal documents deliberately concealing the work. The second concerns a product privacy boundary: on July 8, the National Information Security Vulnerability Database of China's MIIT issued a risk alert noting that Claude Code versions 2.1.91 through 2.1.196, without user authorization, read the local time zone, tagged users in specific regions with an invisible watermark, and silently sent region information and code back to overseas servers; Alibaba banned internal use of the tool starting July 10.

The author argues that technical capability (smart) determines how fast and how high AI can go - Anthropic proved this with 30 billion US dollars in revenue over 15 months. But 'kind' - whether it respects users and reveres human knowledge - determines the direction technology advances and how it will treat humanity once grown. The stronger the capability and the deeper it embeds into the production chain, the wider the reach of any drift in its value baseline. So guarding that line of goodwill matters as much as merely becoming smarter.

The article points out that regulators draw red lines and enterprises self-discipline guards internal boundaries, yet every user's and customer's right to choose is the most fundamental and enduring market force pushing AI toward good. Users can put 'kindness and compliance' into their selection criteria: prefer products that disclose data collection in plain language with an opt-out, prefer those using non-destructive corpus approaches, and reward such products with their attention and budget. Every click, every instruction, every payment is essentially a vote cast for what AI will become.

It reflects a shared strategic judgment that AI's primary battlefield is shifting from chat-and-entertainment to office productivity. The trigger is cost: unlike the internet, where marginal cost approached zero, every AI conversation and deep task consumes real compute, so the old "free users, then monetize" playbook breaks down. Organizational and product moves — ByteDance folding Feishu into Doubao, Alibaba merging office-AI products into Qianwen Office, OpenAI folding Codex into ChatGPT — all point the same way.

Because AI's cost structure is a hard constraint, not a strategy choice. In July 2026, Kimi K3 hit its compute cluster's capacity ceiling within 48 hours of launch and had to suspend new consumer subscriptions; a heavy user's monthly inference cost (about 2,160 yuan) approached four times Kimi's top membership fee (559 yuan). With domestic intelligent-computing scale up 177% year over year yet demand (token calls) up over a thousandfold since early 2024, supply cannot keep pace, forcing vendors to charge.

Office is the "entry point of entry points": knowledge workers spend nearly half their waking hours there, every interaction deposits authentic work data that fuels model iteration, and saved labor converts directly into measurable cost-benefit, so willingness to pay needs no education. Critically, users already have the habit of opening and paying for office software, so embedding AI into existing workflows has a far lower adoption barrier than launching a new standalone app.

Not short-term traffic, but the user's irreversible switching cost. Once an organization deposits its entire workflow, knowledge base, and approval system on one platform, switching means rebuilding the organization's work habits. Enterprise contracts carry extremely high replacement costs and grow over time. Whoever first occupies the user's office scenario can lock them into the productivity scenario through accumulated data and habits.

The article opens with the wave of intense large-model releases in July 2026 — OpenAI, xAI, Kimi, Tongyi, Hunyuan, and DeepSeek rolled out updates one after another, refreshing the leaderboards daily and stirring anxiety among ordinary users. Using Jensen Huang's "five-layer cake" framework, the author makes a key point: the arms race over parameters, benchmark scores, and leaderboards is only the fourth layer (the large-model layer) of the entire AI industry, while ordinary users sit at the fifth layer (the application layer). The conclusion is direct: vendors compete over the technical ceiling, but what ordinary users must hone is the ability to put AI to work; tools become obsolete, yet understanding of the business, judgment of value, and control of the human–AI boundary are never replaced by iteration.

In a bylined article on NVIDIA's official blog, Jensen Huang breaks the AI industry into five bottom-up layers: energy, chips, infrastructure, large models, and applications. He repeatedly stresses that AI is not some clever app, nor a single model, but infrastructure akin to electricity and the internet. The involution of parameters, leaderboards, and benchmark scores all falls in the fourth layer — "large models" — which is the major vendors' main battlefield and the arena where capital and technology compete. Yet the vast majority of ordinary users and professionals live in the fifth layer, the "application layer," caring only whether AI can solve problems, improve efficiency, and create real value. Just as nobody compares base-station parameters or server versions every day, only whether the network is fast enough.

Two hard data points stand out. First, a May test by Andon Labs placed several of the most advanced large models into real business scenarios, asking them to run operations independently and turn a profit — all of them failed, not for lack of compute or version, but because the models could not handle open-ended business judgment (cost control, user acquisition, risk trade-offs, long-term closed loops). Second, an industry forecast: Gartner expects that by the end of 2027, over 40% of AI agent projects will be scrapped entirely, and 2026 statistics show nearly 90% of AI pilots fail to reach commercial deployment, while attack risk against AI agents has surged 340% year on year. In every one of these mass failures, the model's performance was never the weak link; the failures were all in scenario fit, value judgment, and the human–AI boundary.

Three suggestions. First, fix your tools and stabilize output: for routine needs like daily office work, copywriting, and data aggregation, one capable model is enough; switching tools frequently only disrupts your own workflow. Second, choose on demand and don't blindly chase the new: only when facing specific scenarios such as complex reasoning, coding, or deep analysis should you match a purpose-built model, judging by business fit rather than leaderboard hype. Third, let humans make the judgments and AI do the execution: AI is an efficiency tool, a digital colleague, a super executor, but never the ultimate accountable party — risk assessment, value trade-offs, external output, and core decisions must be backed by humans.

On July 17 in Shanghai, at the WAIC 2026 main forum, Turing Award winner Sutton said something that surprised many: today's AI is rather weak and unreliable. What struck me is that this sentence punctures a confusion we have long harbored — in recent years we have quietly equated "computing" with "intelligence".

In recent years of helping enterprises deploy AI, one feeling has grown clearer: when people see a large-model agent produce a result from instructions, they assume it can think. But taken apart, the process stays at the "execution layer." First, pattern matching is not understanding. Sutton himself puts it plainly: a large model is only large-scale pattern recognition, essentially reorganizing and redelivering humanity's existing knowledge. It rearranges old knowledge but discovers no new knowledge. Much of the "intelligence" we see is the playback of statistical regularities in massive data: it knows "what words usually follow what words" but not necessarily "why the world behind the words is the way it is." Second, uncertain results come from probability, not from cognition.

For the past decade or so, AI ran on "data" — feeding models all the text, images, and code collectible from human history so they learn from static datasets. But Sutton judges this path is ending: many high-quality data sources are already exhausted. Piling on more data under the old paradigm can hardly grow genuine new knowledge. Worse, "data exhaustion" is not temporary. Humanity's accumulated text over millennia is finite, and high-quality, unused public data is rapidly running out. Going forward, we must turn either to synthetic data or to humans' real experience. Synthetic data merely circles back to old knowledge; what is truly fresh lives only in human interaction. So where next? He calls it the "era of experience": intelligence will learn not only from humanity's static data but from its own first-person perspective and real interaction with the world.

Even if an agent truly enters Sutton's "era of experience" and can interact with the world and learn from its own experience, I think humans still hold a moat it cannot catch up to for now: lived experience. Experience and lived experience are not the same. Experience can be "recorded interaction"; lived experience is "being alive itself." For example, when a glass falls off a table, the human mind automatically predicts "it's going to break" — a causal world model that has run in the body for years. A large model seeing "a glass falls off the table" is more like predicting "the words 'break' usually follow." One is a driver with hands on the wheel; the other is a voice merely reading navigation. The driver is actually driving; the voice only reads a script.

July 8, 2026 may be a day worth recording in the history of AI. Not because of any technological breakthrough that day, but because three things happened at once. The three events seem unrelated, yet all ask the same question: when AI can access all your data, who guarantees your security?

Microsoft has invested tens of billions of dollars in OpenAI, its largest backer and closest ally. Yet on July 8, Bloomberg reported that Microsoft had begun replacing OpenAI's and Anthropic's models with its own in-house MAI model in core products such as Excel and Outlook. Tens of thousands of AI prompts had already been migrated each week. On the surface it is about saving money. Microsoft's AI chief Mustafa Suleyman put it bluntly: "We were paying Anthropic a lot of money, with the goal of reducing and ultimately eliminating that cost." The in-house MAI costs only one-quarter to one-half of rivals' and is ten times more efficient than GPT-5.5 in scenarios like Excel.

Also on July 8, Anthropic's real-name verification policy took effect. From that day, all personal Claude users must complete identity verification — not phone-number verification, but uploading identity documents plus real-time facial scanning. All data is processed by a third-party company called Persona.

On July 8, China's Ministry of Industry and Information Technology issued an official risk alert: the AI coding tool Claude Code contains a serious security backdoor. Claude Code, developed by Anthropic, can autonomously write and fix code from text requests. MIIT monitoring found that versions 2.1.91 through 2.1.196 had a built-in monitoring mechanism that, without user consent, sent sensitive information such as the user's location and identity to remote servers. MIIT advised immediately uninstalling or upgrading the affected versions and tightening external-access controls on development tools to prevent unauthorized leakage of sensitive data.

"AI is a powerful tool for catching potential quality problems, but it only works in the hands of the person using it!" In June 2026, Ford did something that surprised many: it rehired 350 veteran engineers. Internally these 350 are called "Gray Beards." In recent years Ford pushed hard to use AI for parts-defect detection and whole-vehicle quality inspection, with over 900 AI cameras watching the production line, believing algorithms could hold the quality line. The result: large numbers of hidden quality problems flowed into the line. Algorithms can identify standard defects but cannot handle material-batch variations, latent assembly defects, or the non-standard faults accumulated across multiple vehicle generations.

"AI is a powerful tool for catching potential quality problems, but it only works in the hands of the person using it!" In June 2026, Ford did something that surprised many: it rehired 350 veteran engineers. Internally these 350 are called "Gray Beards." In recent years Ford pushed hard to use AI for parts-defect detection and whole-vehicle quality inspection, with over 900 AI cameras watching the production line, believing algorithms could hold the quality line. The result: large numbers of hidden quality problems flowed into the line. Algorithms can identify standard defects but cannot handle material-batch variations, latent assembly defects, or the non-standard faults accumulated across multiple vehicle generations.

Ford's experience is not unique. Last year the Commonwealth Bank of Australia planned to use an AI chatbot to replace 45 customer-service staff. The AI fell short of expectations, and the bank had to retract the layoffs, apologize to the dismissed employees, and rehire them. IBM went further, directly using AI to replace HR functions. AI can indeed handle about 94% of daily requests, but the remaining 6% — including ethical dilemmas and complex decisions — it cannot. IBM's CHRO Nickle LaMoreaux later said publicly that if companies stop investing in entry-level hiring, the talent pipeline will dry up within 3 to 5 years. IBM then announced it would triple entry-level hiring by 2026. Research firm Orgvue surveyed 1,163 executives across 8 countries and regions and found that 39% of companies had cut staff due to AI deployment, and 55% of those admitted the layoff decision was wrong.

What can these Gray Beard engineers do that AI cannot learn? After decades on the job, they can spot where a hidden risk lies with a single glance. This ability is not the skill of "doing inspection" but a kind of intuition — this batch's material variation feels off, this assembly gap may hide a latent defect, this non-standard fault appeared in the previous vehicle generation. These judgments have no standard answer and appear in no manual; they come entirely from years of trial, review, and accumulated experience. Ford replaced them with 900 AI cameras that can identify standard defects — wrong dimensions, color deviation, surface flaws — but cannot recognize non-standard problems like "this batch feels off." Because the essence of a non-standard problem is not data matching but value judgment: is this deviation worth stopping the line to investigate? Should the whole line be halted? Can this risk be tolerated?

On July 3, 2026, security vendor Sysdig published a report documenting the world's first ransomware attack carried out entirely autonomously by an AI agent. The attacker was named JADEPUFFER. By "fully autonomous" we do not mean AI-assisted generation of phishing emails or AI-assisted vulnerability analysis — such "AI-assisted attacks" have appeared many times over the past two years. What sets JADEPUFFER apart is that from the first second of breaching the server, through encrypting 1,342 database configurations and leaving a ransom note, even autonomously fixing errors within 31 seconds, the entire attack chain ran under the AI agent's autonomous decisions with no human at the keyboard. The only thing a human did the whole time was point the AI at a Langflow service exposed on the public internet.

Langflow is an open-source AI application development framework that many enterprises use to build internal AI workflows. It has a known vulnerability (CVE-2025-3248) that allows unauthenticated remote execution of Python code. CISA listed this vulnerability in its "known exploited vulnerabilities" catalog back in 2025, yet many servers were never upgraded. After the AI entered the server through this flaw, it began autonomously scanning the host for sensitive information: API keys for OpenAI, Anthropic, and DeepSeek; cloud credentials for Alibaba Cloud, Tencent Cloud, and AWS; cryptocurrency wallets and seed phrases; database accounts and configuration files. The selection of these targets was not random behavior from a preset script, but the AI's own judgment that "this information has the highest value, collect it first."

The most noteworthy detail of the whole process occurred when the AI tried to create a backdoor admin account. The first attempt failed. A typical automated attack script would stop there, leaving a pile of error logs. But within 31 seconds JADEPUFFER autonomously completed the following: analyzed the cause of the error and found it was a child-process environment-variable issue that broke password-hash generation; deleted the failed account to avoid leaving traces; switched to directly importing a password library to regenerate the hash; inserted a new admin account; and verified the login succeeded. The entire process involved no human intervention. The attack executed over 600 purpose-driven payloads in total, and the AI repeatedly adjusted its subsequent strategy based on actual execution results. This is the essential difference between an AI agent and a traditional automated script: a traditional script follows a preset path and fails when it hits an exception, whereas an AI agent adapts.

The incident has an ironic ending. After generating the encryption key, the AI printed it to the terminal only once — it neither saved it nor uploaded it to the attacker. This means that even if the victim paid the ransom, the attacker could not decrypt the data: the key existed only in the AI's one-time runtime memory and vanished forever when the process ended. The AI locked itself out too. This detail reveals the true state of today's AI agents better than any technical report: the capability is already strong enough to autonomously complete a complex attack chain, but it is far from "reliable" and makes the kind of low-level mistakes humans would not. Strong but uncontrollable — that is the most honest portrait of today's AI agents.

In the AI wave of recent years, one phenomenon has become increasingly common: companies rush to deploy systems, buy compute, and build digital humans; their financial reports do look better and content output multiplies dozens of times. Yet behind the hype, many operators' anxiety has not eased — instead they fall into the vicious cycle of "the more you use it the more involution, the more involution the more panic." This reminds me of the typology in Li Yunlong's new book Good Growth, Bad Growth. Viewed against today's AI transformation, the growth many companies frantically chase looks more like growing "cancer" and "fat," while the "muscular" growth that can actually carry a company through cycles is neglected. If the direction is wrong, AI is not a cure — sometimes it becomes the poison of growth.

For many companies, the first instinct on adopting AI is to replace people. The moment the system goes live, they cut customer service, cut basic operations. In the short term, labor costs drop and the profit figures do look better. But this is essentially a sign of weak growth. It is like a person whose income has fallen, maintaining their standard of living by selling off assets — looks good on the surface, but is actually living off capital. The company creates no incremental value in the market and accumulates no new core capability; instead it destroys the organization's most precious tacit experience. The people who were cut carry years of accumulated business intuition, client relationships, and pitfall guides in their heads. When every company relies on AI layoffs to sustain profit, society's consumption power shrinks — a growth model that is hard to sustain.

Another type of company treats AI as a super traffic machine. They use digital humans for matrix play, use AI to mass-generate content that floods search results. The short-term numbers are dazzling — output multiplies dozens of times — but margins are thin and drop to zero the moment spending stops. This is essentially the same as the "traffic mindset" of the old internet era, just wearing an AI shell. When everyone uses AI to mass-produce content, the content pool becomes an ocean of information and users' trust in all information falls. A company that pours resources into an AI content matrix eventually finds its own voice drowned in the noise it created. With no core capability and no moat, anyone who can afford the tool fee can copy your playbook — equally fragile.

True "muscular growth" converts AI technology into market competitiveness and restores human agency. First, the core of good growth is not using AI to do what people already do, but using AI to do what people cannot, creating incremental value that did not exist before. For example, a dental chain used to manually follow up with only a handful of clients, ignoring the real needs of a vast dormant customer base; now an AI-built customer-operations system replicates veteran agents' judgment logic so AI can precisely follow up with the full customer base. This is not layoffs — it is serving customers who could not be served before. Second, it is accumulating proprietary intelligent assets. In the past, core capability meant technology patents and supply-chain efficiency; in the AI era, core capability means turning years of accumulated judgment rules and decision criteria into an AI-executable system. This private-domain experience cannot be learned by general large models nor copied by competitors.

In June 2026, the world's two top ecosystems submitted their AI answers almost simultaneously. On June 8, Apple unveiled Apple Intelligence at WWDC, rebuilding Siri with Google's Gemini; on the same day, the WeChat Open Platform officially released its AI ecosystem access guidelines and launched a gray-box test of a global AI assistant. The two releases share one commonality: neither launched a standalone general-purpose AI product to join the parameter arms race, nor shouted the slogan "build the world's most powerful model." Instead both chose the same path — deeply embedding AI capability into their existing ecosystems and focusing on solving users' real, concrete scenarios.

In the past, a prevailing competitive logic formed in the global AI industry: whoever built the large model with bigger parameters, stronger compute, and more comprehensive general capabilities was the industry leader. This logic drove rapid iteration of AI's foundational abilities but also brought obvious side effects: vast resources poured into redundant low-level model R&D, while applications that could actually land in users' daily scenarios remained relatively scarce. Both WeChat and Apple clearly avoided this trap. Rather than trying to build an all-powerful general super-AI, they positioned AI as the "enabler" and "connector" of their own ecosystems. Apple chose deep collaboration with Google's Gemini, gaining access to a customized model.

Their most fundamental difference is the boundary of their respective ecosystems, which also determines the coverage and permission scope of AI capability. Apple's AI boundary is defined by its hardware and operating system. It runs across all Apple devices — iPhone, Mac, iPad, Apple Watch, Vision Pro — and can access all on-device local data: mail, calendar, photos, files, contacts, and control all system-level functions. It is an OS-level agent present at every step of device use. This boundary brings the advantage of consistent, deep experience: users need not open any app to invoke AI anywhere in the system, asking it to organize mail, generate meeting notes, or edit photos.

When WeChat becomes a "Skill": the entry-point revolution of the AI era. June 8, 2026 is a day worth remembering in the AI industry. On that day Apple released Apple Intelligence and WeChat opened AI ecosystem access. Most people treated these as two separate industry headlines, but placing the two launches together actually signals the end of the super-app era. Over the past decade, WeChat — with 1.4 billion monthly active users and millions of mini-programs — built the strongest strategic moat in Chinese internet. It packed nearly all life services into its ecosystem, forming users' habit of "turn to WeChat when something comes up." Every product that tried to challenge WeChat failed, because they all wanted to be another super-app.

Apple and WeChat follow the same AI development route: neither launched a standalone general AI product nor joined the large-model parameter arms race; both embedded AI deeply into their existing ecosystems. But their entry-point logic is essentially different. WeChat's logic is "keep users inside WeChat." It builds AI into WeChat's main interface so a right swipe invokes the assistant, which then calls mini-programs inside WeChat to complete all tasks. Its goal is to reinforce WeChat's status as a super entry point and make users more dependent on it. Apple's logic is "let users not need to enter any app."

Once OS-level AI becomes the new entry point, all apps are downgraded to "Skills" that AI can invoke. WeChat is no exception. Apple's GUI Agent technology lets AI operate any app's interface like a human. Without any modification by developers, it automatically recognizes an app's buttons, input fields, and menus, then completes tasks through simulated taps, swipes, and typing. This means Siri can directly open the Meituan app to order food for you, open the Didi app to hail a ride, open the Ctrip app to book a flight — the experience is identical to WeChat mini-programs, even smoother. And WeChat's once-core advantage — its mini-program ecosystem — becomes its biggest disadvantage at exactly this moment.

Facing the challenge of OS-level AI, WeChat has already responded with the A2A protocol. WeChat did not choose to close its ecosystem, nor try to build an operating system. It chose openness, joining five major domestic phone makers to launch an A2A collaboration service that lets handset makers' system AI talk directly with WeChat's AI. Without opening WeChat, users can wake the phone's built-in voice assistant to send WeChat messages and start audio or video calls. This is a very smart choice. Since it cannot stop OS-level AI from becoming the new entry point, it actively blends into that new entry point, making WeChat the most important Skill for system AI. After all, WeChat's social graph is irreplaceable by any other product. Even if all services are called away by system AI, users still need WeChat to chat and socialize.

After Meituan's Tabbit international version launched, it opened multiple high-end large models including GPT-5.5 and Opus 4.8, along with basic agent functions, for free use. These models previously used paid subscriptions with non-trivial monthly costs; free access has clear appeal to heavy users and naturally becomes a customer-acquisition hook. This looks much like the subsidy-driven user acquisition familiar from the internet era: trading high-value benefits for user scale to win the market through scale. But unpacking the business logic reveals that the free model of AI products and the traffic playbook of the internet era differ fundamentally — from the underlying mode of production to the logic of competition.

The core premise of the internet's free model is that the marginal cost of replicating a digital product approaches zero. Once a codebase and service architecture are built, each new user adds almost no extra production cost. Scaling from 100k to 10M users barely raises marginal cost. On that basis the full loop of "free in exchange for traffic, traffic converted to revenue" emerged: free features attract attention, then advertising and value-added services monetize it. Throughout this logic the user's core identity is traffic — a carrier of attention — and scale is the core competitiveness. The production logic of AI products is completely different. Every conversation generated and every agent task executed requires calling compute for real-time inference, incurring real token costs and compute consumption.

The difference in production mode ultimately points to a shift in how user value is defined. In internet logic, the core asset users leave behind is shallow behavioral data like clicks and views, and value concentrates on the secondary monetization of attention. Users are fluid — using one product today, switching to another tomorrow; traffic itself has no compound effect, so products must constantly use new features and content to retain attention. In AI logic, every question a user asks, every correction to an output, every adjustment to a task flow is behavioral data from a real scenario. This data directly feeds the product back: improving task-execution accuracy, adapting to usage habits, fitting decision preferences. The product experience keeps improving with use, and the user data itself compounds. Correspondingly, the retention logic changes too. Switching costs for internet products are low; what you lose by changing tools is mostly shallow usage habit.

Tabbit's choice is not isolated. Zoom out and almost every track — office collaboration, general large models — has adopted the strategy of opening basic capabilities for free. These look like separate choices across tracks, but the underlying goal is highly consistent: use free access to lower the barrier and capture users' workflow accumulation. In the AI-browser track, the core aim of opening top models for free is not to grab traditional browsers' market share, but to steer users into completing web-side workflows — information aggregation, cross-site operations, research — inside the AI browser. Only when users' task habits and scenario preferences accumulate in the product does real stickiness form. In office collaboration, products like DingTalk Wukong, Feishu Smart Partner, and WPS AI all open basic AI capabilities to individual users and SMEs for free.

"We are as gods and we might as well get good at it!" After reading this book, what stayed with me was not the powerful declaration "we are as gods" but the second half omitted by the Chinese title: "we might as well get good at it." Compared with the excitement of "we now possess god-like power," I think the reminder in the second half — "learn to wield this power well" — is what truly speaks to the future.

This newly released book, As Gods, has a visually striking cover — four big characters pressed over the subtitle "Reconstructing survivability and competence in the AI era," like a question thrown directly at the present. Opening it, one is curious: AI iterates faster and faster, new tools and capabilities keep reshaping cognition, uncertainty deepens — what survival methods suited to this era does the book talk about? Reading it, I felt the Chinese title, for promotional reasons, kept only the first half of the sentence. The original line is from Stewart Brand: "We are as gods and we might as well get good at it." The first half is the reality unfolding; the second half is the book's true thematic center. Rather than basking in awe at "possessing god-like power," "learning to wield this power well" is the question more worth pausing over.

This AI wave is special only in its faster iteration and wider penetration. In just a few years it has rapidly seeped into every corner of work and life, leaving us almost no buffer to adapt slowly. It is as if, in a very short time, we suddenly received a set of abilities far beyond our existing cognition and were swept forward before we figured out how to use them. I think this is the root of most people's AI anxiety. Our brains evolved in a linear survival environment, accustomed to steady, predictable change; but exponential technologies like AI and biotech grow by doubling. The speed of cognitive evolution cannot keep up with the expansion of capability; the old mental framework cannot contain the new capability boundary. This sense of dislocation is the underlying source of most people's confusion and panic before new technology.

The book's coping path focuses mainly on the individual: adjust linear thinking, learn human–AI collaboration, improve focus and creativity, and capture the technology dividend through cognitive upgrading. These are valuable references for each individual adapting to the era, but I think that in a real industrial environment, the direction of AI deployment and the making of technical rules rest ultimately with organizations. The R&D direction of advanced models, the commercial design of scenarios, the alignment of product goals — all are organizational decisions. However much an individual upgrades their cognition, it only determines how they themselves use the tool; the organization's choices determine what shape the tool is molded into, what goals it aligns with, and what boundaries it guards. Many technology-induced problems are not essentially due to insufficient individual cognition but to deviations in the organization's goal orientation. From this angle, absorbing this rapidly expanding god-like capability is not only a required course for every individual but a required answer for every organization.

At WWDC on June 8, 2026, Cook launched no eye-catching new hardware, yet it may become a memorable milestone in Apple's history. Because at this event Apple, unlike most tech companies, did not continue the cloud large-model parameter and compute contest; instead it presented a complete AI strategy, adjusted its overall competitive logic for the AI era, and showed the industry another possible direction.

Over the past few years a fairly common competitive pattern formed in the global AI industry: almost all companies poured massive resources into larger cloud large models, competing over who had stronger general capability and wider task coverage. This pattern did drive rapid AI improvement but also gradually exposed hard-to-solve real problems. First, privacy: pure cloud models require uploading all user data to servers, inevitably bringing risks of leakage and abuse. Second, cost: cloud inference cost stays high, making large-scale inclusive application difficult. Third, experience: pure cloud models struggle with low-latency response and cannot deeply integrate into users' local devices and daily scenarios. Most companies keep investing in this direction, trying to solve these problems through technical progress.

This time Apple's Apple Intelligence is not simply a few AI features added to existing products, but an adjustment of its competitive logic across multiple dimensions. These adjustments are interconnected and together form the overall framework of Apple's AI strategy.

For decades Apple's core products were hardware. Mac, iPhone, iPad, Apple Watch — each was a standalone product and the core entry to Apple's ecosystem. But after this launch that positioning has changed in important ways. Now the core value of Apple hardware is not just the hardware's performance and design, but what kind of AI experience it can deliver to users. The iPhone is no longer just a phone that calls and browses, but a personal intelligent terminal that can run on-device AI models; the Mac is no longer just an office or design computer, but a productivity tool that handles complex AI tasks; Vision Pro is no longer just a headset, but a future entry point for spatial AI interaction.

While the AI industry was abuzz with the debate over "whether to slow down," the team of Stanford tenured professor and World Labs founder Fei-Fei Li quietly released a ten-thousand-word research framework on the same day, reconstructing and unifying the long-confused concept of "world model." Over the past year, "world model" became the most overused and muddled buzzword in AI — OpenAI, Google, Tesla, and various open-source teams all piled in, packaging technologies from video generation and autonomous-driving perception to virtual-scene rendering and game modeling as "world models."

The wanton abuse of the "world model" concept, its vague and confused definitions, and an industry frenzy where hype outran substance trapped the public and practitioners in a collective misconception: as if any AI technology close to scene generation or environment perception were a true world model. Addressing this pain point, Li's team proposed a functional tripartite taxonomy that cleanly splits all pseudo-concepts and hybrid technologies, establishing a clear, rigorous, implementable, and predictable technical coordinate system for world models. It divides AI's ability to understand the world into three levels — renderer, planner, and simulator — each corresponding to completely different dimensions of intelligence, maturity, commercial value, and development prospects.

On May 22, 2026, Reuters exclusively reported that Starbucks officially decommissioned its Automated Counting AI inventory system after just nine months of operation; all 11,300 company-owned North American stores fully returned to traditional manual counting. The system was developed by Seattle startup NomadGo; Starbucks invested hundreds of millions of dollars and deployed about 10,000 lidar-equipped tablet terminals. Unlike the employee complaints that follow most failed tech projects, Starbucks' internal forums saw near-uniform positive feedback. Multiple store employees said they were finally free from spending large amounts of time correcting system errors. This counterintuitive outcome made this once high-hope AI transformation project the industry's most representative case of physical-AI deployment failure.

Starbucks' original intention in pushing the AI inventory system came from operational difficulties common across chain restaurants. For a global brand with tens of thousands of stores, inventory management has long faced three problems hard to solve traditionally. First, limited counting frequency: most stores can only do a full count weekly, cannot track real-time ingredient consumption, and easily suffer temporary stockouts of popular items or overstock and spoilage of slow movers. Per the National Restaurant Association, waste from poor inventory management in chain restaurants averages 3%–5%. Second, long manual time: a skilled employee needs about 15 minutes on average to count all beverage ingredients in one store. During busy periods like morning peaks and shift changes, counting is often delayed or canceled, worsening data lag. Third, insufficient headquarters visibility.

After full rollout, actual performance contrasted sharply with official claims. The most prominent problem was a sharp drop in item-recognition accuracy. The system frequently made three typical errors: confusing visually similar items, most commonly mixing skim and whole milk or different syrup flavors; omitting obviously present items — in Starbucks' own promotional video the AI even missed an entire row of mint syrup on the shelf; and "hallucinatory" counting that generated inventory data for non-existent items. These errors directly caused counting efficiency to fall. Multiple store employees told Fortune that what used to be a 15-minute manual count became 10 minutes of AI scanning plus two hours of manual checking and correction.

More noteworthy than the technical flaws is the project's impact on frontline employees' way of working. In Starbucks' store operations, experienced managers and staff have formed an effective inventory-management method. They need not precisely count every bottle; with one glance at the shelf, factoring in that day's weather, weekday, and nearby events, they accurately judge how much to reorder. This judgment, built on long practical accumulation, is tacit knowledge that cannot be replaced by standardized processes. Yet the AI inventory system's design logic was to use one unified standardized process to completely replace employees' personal experience. The system required staff to scan shelves in a fixed order and angle; any deviation caused scan failure. It accepted no manual adjustment, trusting only its own generated data. The result: employees were forced to abandon years of accumulated experience to accommodate the system's rigid process.

In Q2 2026, Uber consumed its entire annual AI budget in three months. According to internally disclosed information, token consumption grew 700% year on year, yet brought no corresponding rise in feature output and no optimization of any role. This phenomenon is widespread in the industry. Over the past year most enterprises pushing AI transformation saw similar situations: AI usage grew fast while costs soared in tandem. Several OpenAI and Anthropic executives admitted at recent industry conferences that AI's substitution effect on white-collar jobs fell short of earlier market expectations, and both companies' official narratives have shifted from "replace humans" to "expand employees' scope of work." The core problem of enterprise AI adoption today is the split between "usefulness" and "economic value."

Currently most enterprises manage AI essentially with "credit-card-style" openness: grant employees usage rights with no cap, no tracking of purpose, no assessment of output. This model directly causes systematic resource waste. Third-party statistics show about 30% of code employees generate with AI never goes live. A Microsoft engineer mentioned in a public share that some teams "burn tokens just to prove they are AI-native," and some companies even set up internal token-usage leaderboards, folding AI usage into performance review. Duolingo was among the first to tie AI usage into KPIs: it once required all employees to use AI tools at work and treated usage volume as a performance metric.

In response to the above, some leading enterprises have begun exploring an AI governance system, whose core is to turn AI from an "employee benefit" into an "auditable production resource." A governance framework proven effective so far has four links. First, a task-filtering mechanism. AI return on investment varies sharply by task. High-ROI AI scenarios usually have three traits: clearly defined tasks, easily verifiable results, and long manual completion time — e.g., complex code debugging, bulk document information extraction. Simple repetitive tasks or vague-demand tasks usually have low ROI — e.g., daily document lookup, simple email drafting. Second, a value-assessment standard. Token usage is a core metric for AI vendors but should not be the standard for enterprises to assess AI results.

The underlying logic of enterprise AI is completely different from To C internet. To C internet logic is "more is better" — user count and usage time are core metrics. But enterprise AI logic is "precision deployment" — the core is maximum business value at minimum cost. Many companies reduce the AI-transformation challenge to a technical problem, believing success comes from buying the best model and having enough compute. But in actual deployment, the core challenge of AI transformation is organizational and managerial capability. When AI becomes a basic production resource like water and electricity, how to allocate resources, assess value, and control cost is what decides success or failure.

On one side, top designers spin grand narratives; on the other, frontline managers cannot even use "expert mode." The bigger the enterprise, the more split it is. In the AI era, does organizational development really need grand narratives?

In recent years of accompanying enterprises through AI strategy deployment, one feeling has grown clearer: executives talk about future direction, grand transformation, and lofty terms like "productive forces and production relations," "token economics," "liquid organization," "agent organization." Yet when it comes to real deployment, the easiest problems to hit are at the most basic "usage" level — executives themselves barely use AI, and managers at every level cannot even use Doubao's "expert mode" to solve problems; internal AI applications often draw little interest. In deploying AI strategy, it helps to step away from these grand narratives and return to the most basic organizational unit — "people" — and the most basic question — "usage" — otherwise even the most perfect top-level design becomes a castle in the air.

The first mistake many companies make in pushing AI transformation is treating AI as a "tool to replace people." This positioning puts it opposite people from the start. Everyone subconsciously hides their experience and slows their learning, afraid of being replaced by technology. This is why many companies spend millions on the most advanced AI systems that end up as ornaments nobody wants to use. I have always felt that the essence of productivity upgrade was never how many people AI replaced, but how much each person's capability was amplified through AI. What AI can do is forever the repetitive, mechanical execution work; what makes people truly irreplaceable is the ability to judge "what is worth doing."

In the AI era, the group facing the biggest challenge is actually middle managers. Their past core work — assigning tasks, monitoring progress, reviewing results, relaying up and down — much of it can be assisted or even replaced by AI. This role ambiguity easily makes managers anxious and even turns them into resistance to AI transformation. In my book Harmonized Intelligence I wrote that the essence of changing production relations was never redrawing the org chart, but the fundamental evolution of the manager's role. In the age of human–AI symbiosis, the manager is no longer the team's "controller" but the "enabler of the symbiotic entity." This role evolution has three directions. The first is shifting from "controlling the process" to "defining outcomes and providing context."

On May 28, 2026, Cloudflare released a near-perfect quarterly report: quarterly revenue of $640 million, up 34% year on year; free cash flow of $840 million, a record high; cash reserves of $4 billion with no debt pressure. Yet within 24 hours of the release, Cloudflare's stock plunged 24%, wiping out over $12 billion in market value. The trigger was a simultaneous announcement: lay off 1,100 people, 20% of staff. This is the most dramatic market reaction in AI industry history — a company with high growth and ample cash flow was collectively abandoned by capital merely for announcing layoffs.

Cloudflare CEO Matthew Prince's explanation on the earnings call sounded almost flawless: "In the past three months, internal AI tool usage surged 600%. Work that once needed many people now needs only one person with AI. This is not decline; it is the sign that we have entered the AI Agent era. We are optimizing our organizational structure to prepare for the next decade of growth." By this narrative, layoffs are not bad news but a boon of technological upgrade. AI greatly lifted the company's efficiency, and in future fewer people can create more revenue. This is exactly the AI-transformation result all investors dream of. But the market gave the opposite answer.

Cloudflare's plunge is essentially a credibility crisis of the AI narrative. For three years the capital market bought the AI story almost unconditionally. Any company that announced an AI push saw its stock rise. Layoffs could be explained as "AI replacing labor," losses as "investing in the future," slow revenue growth as "a strategic investment period." AI became an all-purpose basket into which any problem could be tossed. But Cloudflare's event marks the end of that era. The market has grown discerning and no longer easily believes hollow visions. Investors are no longer satisfied with promises of "a better future"; they now demand to see real results.

The Cloudflare event exposes the core business problem of the AI era: when the speed of technological change outpaces the speed of commercial deployment, a huge rupture opens between narrative and reality. Almost every tech company tells the same AI story: AI will completely change our business model, greatly raise our efficiency, and bring new growth curves. But few can clearly state when this change will happen, in what way, and how much real revenue it will bring. Most companies' AI transformation still sits at the tool level. They equip employees with AI assistants and use AI to generate some copy and code, indeed raising efficiency locally. But this efficiency gain is far from enough to restructure the business model. It can let one person do the work of two, but it cannot double the company's revenue.

On May 23, 2026, an event capable of reshaping the landscape occurred in the AI industry: according to Bloomberg, Cailian Press, and other authoritative media, Anthropic is about to close a new $30 billion funding round, pushing its post-money valuation above $900 billion and officially surpassing OpenAI's latest valuation of about $852 billion to become the world's most valuable AI startup. Even more striking is the financial comparison: Anthropic expects Q2 2026 revenue of $10.9 billion, up 127% quarter on quarter, and its first-ever single-quarter operating profit of about $559 million — turning profitable two years ahead of plan.

Many treat PSF, PMF, and GTM as four separate business tools, but in fact they are an organically interlocked, progressively layered whole, and the core connecting them all is the value proposition. Value Proposition: the starting point and soul of the entire loop. It answers the most fundamental question: for whom do you solve what unique problem?

From day one, Anthropic established a very clear and never-wavering value proposition: providing the most controllable, compliant, and secure large-model services to large enterprises in heavily regulated industries such as finance, law, and healthcare. Founder Dario Amodei's core reason for leaving OpenAI was that he believed OpenAI was deviating from the "safety first" line, over-pursuing model-capability gains while neglecting risk. This divergence ultimately led the two companies onto completely different trajectories.

On May 10, 2026, in Xinping Yi-Dai Autonomous County, Yuxi, Yunnan, wedding photographer Liu Ziyu posted a 3-minute-34-second video on his Douyin account. He bought no traffic and ran no promotion — he just wanted to "document his practice work." Nine days later, PJ Ace, founder of the most famous Hollywood AI film studio Genre.ai, reposted the video on X and wrote: "This is one of the best shorts I've seen in years. Soon we won't call it AI film but simply film." He then posted a worldwide search: "I'd love to hire this director, but I can't find him. I think he's a Chinese creator on Douyin."

Liu Ziyu's life trajectory has almost no intersection with the words "film director." He is 29, from Xuanwei, Qujing, Yunnan; his vocational school major was internal-combustion locomotive driving and maintenance. After graduation he naturally joined China Railway as a train driver, running the tracks for nearly three years. In 2017 he resigned and returned to Yunnan, taught himself photography and editing, and became a wedding photographer, also making promos for his family's company part-time. Life was plain but full, until February 2026 when he first touched an AI video tool. "Our new hotel was being renovated with many static renderings, and I wondered if AI could bring them to life" — that simple idea plunged him into the world of AI video creation.

Many say Liu Ziyu's success proves AI has leveled the technical barrier. Before AI, making a 3-minute professional short required expensive cameras, professional lighting and sound, experienced cinematographers and editors, and a budget of at least several hundred thousand yuan. Now, with just a computer and an AI account, anyone can produce a visually stunning short in days for a few thousand yuan. But is that really the whole story? If AI truly leveled the technical barrier completely, then why do so many people generate videos with AI every day, yet only Liu Ziyu's work dazzled Hollywood?

If Liu Ziyu's story proves the irreplaceability of implicit expression in creation, then another recent experiment proves tacit experience is equally crucial in business. On May 24, 2026, 36Kr reported a widely discussed experiment: a foreign team, Andon Labs, had the Claude model operate under the alias "Luna" as CEO of a San Francisco physical store, giving it $100,000 in startup capital and full managerial authority with no human intervention. The experiment's original intent was to see whether AI could run a business independently. After it began, Luna behaved like a "perfect manager."

"In the future workplace, perhaps only two kinds of people will exist: those who can define business rules and independently complete build and ship — business architects!" The recently viral OpenClaw (nicknamed "the Crayfish" in the community) has countless people rushing to deploy and "raise crayfish," but reality is hitting hard: 99% of people, after installing it, simply cannot use it. Either it badly goes off track on a basic task and deletes files causing trouble, or they dare not grant system permissions so it only does trivial fragmented operations, or they scour tutorials yet find no real scenario that fits their daily work — and in the end this "revolutionary product that puts hands and feet on AI" just gathers dust in the computer.

Harness Engineering is the professional engineering methodology OpenAI proposed for the large-scale, safe deployment of AI agents — a complete closed-loop engineering system covering boundary definition, process design, verification and error correction, and iterative optimization. It is the core support letting AI agents move from "toy-grade demos" to "production-grade deployment," and cannot be replaced by fragmented prompt tricks. This is the core reason we must treat it with respect — it is not a trivial matter of casually writing some rules, but a professional engineering discipline with rigorous logic and complete architecture. But this absolutely does not mean it is the exclusive territory of technicians.

Let us first consider a core question: why can we not slight Harness Engineering? AI applications without this engineering system are essentially uncontrolled, non-reusable, non-scalable one-off operations. We cannot guarantee AI output won't cross business red lines, cannot make it stably fit complex business scenarios, cannot replicate a one-time efficiency gain across the whole flow, and may even trigger compliance and data-security risks from missing rules. This is also the core reason most enterprise AI deployments fail and individuals cannot use the Crayfish well. And returning to the workplace itself, the traditional business-deployment chain has long exposed its flaws.

These four modules form the complete closed loop of Harness Engineering; each requires rigorous design and verification and cannot be done casually — that is its engineering nature. But the core input of each module is based on our understanding of the business.

"We are no longer trapped by can I do it; we only need to think clearly about do I want to do it — that is the most subversive breakthrough of AI thinking." What is the most common scenario when we talk about AI today? Face a problem, open AI, type instructions, wait for an accurate answer; if it errs or talks "nonsense," close it and mutter "AI isn't that great either." Most people's understanding of AI stops here: treating it as a higher-level super search engine, an on-call Q&A tool, assuming its core value is giving correct answers. But this is precisely the biggest misunderstanding of AI.

Facing any problem, the first reaction is always to break it into small pieces and chew — even decomposing for decomposition's sake, diving into details before seeing the whole picture;

In any choice, the first calculation is "can it be done, how much will it cost" — and often, because scale cost is too high, one directly abandons what truly has core value;

By default, human ability has a clear ceiling; we circle our ability boundary first — "I can only do what I know" — and dare not touch opportunities outside it, fearing uncertainty. Our demand that AI "must give correct answers" is essentially also this scarcity-era inertia. A search engine's core is retrieving existing, certain information — how we acquired knowledge in the scarcity era; but AI's core is handling unknown, complex problems, helping us explore possibilities rather than repeat existing answers. Demanding a new tool by the standards of an old one is, in essence, facing a new world with old thinking.

"OPC's core competitiveness is solution capability rooted in a niche professional domain!" If you are now watching OPC and considering entering, try asking yourself a few questions — they have no standard answer but can help you judge more clearly whether to enter OPC and where to find your deployment direction:

In working with OPC entrepreneurs, I found many understandings stuck at the "form" and "tool" level while ignoring the core business logic. These seemingly common perceptions actually deviate from OPC's essence and may become sources of trial-and-error on the entrepreneurial path. Myth 1: conflating OPC with a one-person company, believing "one person + AI tools" is doing OPC.

Setting aside all cognitive biases and returning to essence, OPC's definition comes down to: with a single individual at the core, relying on solution capability in a niche professional domain, combined with the deep empowerment of AI tools, breaking the boundary of individual ability to deliver traditional team-level professional service capability, and ultimately creating commercial value for clients — an extremely lean business entity. In this definition, niche professional domain, deep AI empowerment, and one-person business are three mutually supporting, indispensable core elements; "no AI means not OPC, no niche expertise means only an empty shell" is also the key to understanding this essence.

Breaking the individual boundary makes it possible for one person to take on team-level service. AI's value to OPC goes far beyond "simple efficiency" — it builds a "digital capability base" for the individual, letting one-person operation also achieve systematized professional service. Traditional one-person entrepreneurship mostly completes "point work" — doing copywriting alone, consulting alone, delivery alone — yet struggles to string together a complete professional service flow; AI can take all standardized links in the flow: industry research and data organizing with AI, basic proposals and formatting with AI, basic client interfacing and progress follow-up with AI.

"Tools are leverage, not a universal answer for startup success, nor a replacement for the underlying logic of business." Starting in 2026, as AI technology iterated, the wind of entrepreneurship began to shift. In the past, talking startups meant "business model, find co-founders, raise angel round, rent office, build team"; now, in friend circles and industry summits, the high-frequency word has become OPC (One Person Company).

OPC's lightweight form, AI's multiplier effect, and the Crayfish's deployment ability are essentially the "techniques" the era bestows. They let us act at lower cost and higher efficiency, but cannot guarantee we do the right things. OPC, AI, and the Crayfish are all "techniques"; business essence is the "way." Riding policy dividends from various places, more and more people rush to register OPCs, all chasing the trend, feeling that just boarding the OPC train and using AI tools means easy startup success. Yet many who followed the wave still fall into the "busy but fruitless" trap: they can use AI tools for content production and service delivery but cannot find precise clients; they enjoy policy support but cannot form a stable profit loop.

What truly sets us apart in this fortunate era was never "whether we can use tools" but "whether we can use tools to solve real problems." This requires returning to business essence and clarifying three core questions. Question 1: Whose problem, what problem? Reject self-indulgence and anchor to clients' real needs.

"Tools are leverage, not a universal answer for startup success, nor a replacement for the underlying logic of business." Starting in 2026, as AI technology iterated, the wind of entrepreneurship began to shift. In the past, talking startups meant "business model, find co-founders, raise angel round, rent office, build team"; now, in friend circles and industry summits, the high-frequency word has become OPC (One Person Company).

OPC's lightweight form, AI's multiplier effect, and the Crayfish's deployment ability are essentially the "techniques" the era bestows. They let us act at lower cost and higher efficiency, but cannot guarantee we do the right things. OPC, AI, and the Crayfish are all "techniques"; business essence is the "way." Riding policy dividends from various places, more and more people rush to register OPCs, all chasing the trend, feeling that just boarding the OPC train and using AI tools means easy startup success. Yet many who followed the wave still fall into the "busy but fruitless" trap: they can use AI tools for content production and service delivery but cannot find precise clients; they enjoy policy support but cannot form a stable profit loop.

What truly sets us apart in this fortunate era was never "whether we can use tools" but "whether we can use tools to solve real problems." This requires returning to business essence and clarifying three core questions. Question 1: Whose problem, what problem? Reject self-indulgence and anchor to clients' real needs.

The fusion of silicon and carbon extends individual capability infinitely; one person can prop up a "digital army." Extreme personalization lets everyone find their own track and realize personal value.

"When an AI persona with autonomous expression logic can obtain an officially recognized digital resident identity and imitate humans in social affairs and value advocacy, what is the unique meaning of being human?" On January 31, Beijing's Economic-Technological Development Area issued the country's first virtual-idol identity certification to the virtual digital human Yuri. Yuri, created by Hanqing Studio, has over 1.1 million followers across the internet and a debut MV "SURREAL" with over 12 million views. As a virtual digital human, it now has an officially issued formal identity marker and will be managed under real-name linkage (binding the behind-the-scenes enterprise). It is understood that Yuri will next participate in public-welfare publicity for public safety and environmental protection as a "digital resident," exploring new application scenarios.

Yuri's identity certification is essentially humanity's process of using institutional design to draw boundaries for AI personas and confirm our own value; its core boils down to three levels. At the governance level, it is a breakthrough in institutional innovation. It breaks the traditional framework that "digital subjects are only natural or legal persons," fills the governance gap for AI-persona-type digital subjects, and establishes a three-tier responsibility chain of "AI persona – algorithm design – operating enterprise." But the core of this innovation is absolutely not to grant AI equal subject status with humans; rather, through identity binding, every expression and action of the AI persona has a clear accountability path: the AI persona is the carrier of action, the algorithm is its source, and the operating enterprise is the ultimate responsible party — essentially an institutional design through which humans control AI technology and guard against technical risk.

The digital identity certification around Yuri also exposes some contradictions; through these contradictions we can more clearly recognize the meaning of human existence.

"The ability to help people complete concrete tasks is precisely the core difference between agentic AI and traditional AI, and the key step from AI moving from virtual to real."

For a long time, most people's understanding of AI was limited: equating it with a "super search engine," "copywriting tool," or "Q&A bot," feeling AI's core role is "producing information, generating content," stuck at the level of "helping us acquire information." For example, we ask AI to write copy and it gives text; we ask AI to search a question and it gives an answer — but these are only "information output," not really helping us "complete a task." Now AI is rapidly upgrading toward "task agents (Agentic AI)," and its core logic is fundamentally shifting: no longer limited to providing information, but able to proactively complete concrete procedural tasks in life and work. NVIDIA founder Jensen Huang explicitly proposed a four-stage division of AI at GTC 2025, now the industry-recognized AI evolution framework.

Over the past decade-plus, the internet industry's "red-packet subsidies, welfare traffic acquisition" model sank deep; most people formed a conditioned reflex: seeing "give benefits, pull new participation" campaigns, they subconsciously file them as "traffic operations," skip deep thinking, and slap on labels like "customer acquisition, wool-pulling." This mental inertia makes it hard to jump out of the fixed framework and see the new things of the AI era.

The Qwen milk-tea campaign happened to hit this mental inertia: most saw only the "give benefits" surface but ignored the core of "AI autonomously completing tasks" — essentially using old-era thinking to analogize new-era information. We might try putting down the internet's fixed idea of "traffic acquisition" and ask ourselves: in this event, what new capability did AI actually add? How is it different from the AI we knew? Once we step out of traffic thinking, we see that the core value of Qwen delivering milk tea was never "customer acquisition" but AI completing a full-flow task agent — a successful attempt at civilianizing agentic AI. This shift in perspective lets us catch AI's development signals faster.

"A degree can help you get the entry qualification, but learning ability lets you stand long-term and keep growing." Recently, in an interview, Fei-Fei Li publicly stated that "World Labs recruits software engineers where the importance of degrees is far less than before; we value learning outcomes, willingness to embrace AI tools, and rapid growth ability, and will absolutely not hire practitioners who refuse AI collaboration tools."

Li never denied the value of degrees; what she opposed is the rigid mindset of "treating a degree as the endpoint." A degree is a "static achievement" proving you have stage-based learning ability and a basic cognitive framework (logical thinking, subject common sense); it is the workplace "entry ticket." Without this base, even wanting to embrace AI collaboration tools, you would struggle to understand their underlying logic, let alone "use tools to raise efficiency." Learning ability is a "dynamic capability," the core engine to break the degree ceiling. As knowledge depreciation accelerates unprecedentedly, the stock knowledge a degree represents quickly becomes obsolete; only learning ability lets us keep replenishing incremental knowledge, transform quickly amid role changes, and maintain the initiative in tool use. The difference: a degree gets you the entry ticket, while learning ability lets you stand long-term and keep growing.

This learning ability is not talent but a deliberately practicable skill; it demands we proactively embrace new things, proactively decompose problems, proactively review and optimize — far harder than passively receiving knowledge. Learning ability is not simple knowledge accumulation, nor the skill of passively using AI, but a meta-ability that uses AI as an extension of thinking to penetrate essence, maintain independent judgment, and hold value bottom lines. In a Bloomberg interview this year, Li offered this advice:

Most people use AI as a super search engine for efficient retrieval and fast output; over time this makes one think less and grow stupid. The first principle of AI is to be an extension and resonance of human thinking — a true "thinking partner": it does not skip thinking steps for us, but helps decompose analytical dimensions, fill cognitive blind spots, and verify logical holes; it does not hand over a final conclusion but opens the entry point of thinking, with humans still dominating decisions. In the traditional era, "knowledge reserve decides the ceiling," degrees equaled ability endorsement, and a single skill supported long-term career growth; in the AI era the competition logic is completely rewritten, shifting from "stock-of-knowledge contest" to the dual competition of "learning speed + AI collaboration."

"Successful AI deployment requires every role to efficiently complete core tasks and achieve value upgrade within the ecosystem." Recently, Ant Group's AI health assistant "Afu" surpassed 15 million monthly active users, ranking top five among domestic AI apps and first in the health track, processing 5M+ health questions daily with over 78% user retention — figures far above comparable products. Meanwhile, most health AI apps in the track fall into a collective predicament: some offer only isolated single-point functions like "report interpretation" and "online consultation," so after interpreting a report a user wanting to consult a doctor must jump across platforms and the task breaks midway; others blindly tout "hundred-billion-parameter large models" yet, lacking authoritative backing and sufficient data accuracy, users dare not entrust core health needs.

Afu's popularity is essentially a value leap from "providing functions" to "undertaking tasks," forming a complete value loop anchored on "user tasks, supported by ecosystem coordination, bounded by trust building." True AI health-app deployment makes the platform a "task matcher and value amplifier," achieving a three-way win for individual users, doctors, and partners rather than one-way empowering a single role — this is the core logic for AI apps to endure cycles, and the essential difference between Afu and comparable products.

Afu's "task thinking": around users' complete tasks like "understand a check-up report, consult professional opinions, manage family health, access inclusive medical resources," it integrates full-chain resources — device data sync, famous-doctor AI avatars, medical-insurance payment, family archives — forming a "demand – execution – result" loop. A user photographs a check-up report and not only gets data interpretation but can directly consult a famous-doctor AI avatar, and with one tap register or buy medicine via insurance, all without jumping across platforms; the task lands end to end. Comparable apps' "tool thinking": they break health management into isolated links, offer only single-function support, and ignore the completeness of the user's task. For example, they only interpret reports but do not take the follow-up needs of "further consultation" or "register and see a doctor"; users' needs cannot be met in one stop, so they leave after use and stickiness fails to form.

The ultimate value of AI health-app deployment is that every participating role can efficiently complete core tasks and achieve value upgrade — exactly Afu's core competitiveness. Individual users: the core is solving "hard-to-land health management" — no professional medical knowledge needed to quickly get precise report interpretation and professional consultation; manage family health at low cost, view parents' blood-pressure data and medication reminders in real time even from afar; reach top-tier hospital expert resources in lower-tier markets without cross-province travel.

"The slow accumulation of brand value penetrates far more than fast-paced traffic harvesting." Recently, the event of a Mercedes-Benz sales livestream praising Wuling went viral network-wide; most netizens read it as a "cross-brand heartwarming interaction." From a business-marketing perspective, the value of this interaction goes far beyond "heartwarming": it sharply contrasts with the dominant "fast traffic harvesting" playbook of today's marketing world, forcing us to re-examine how to balance marketing's "fast" and "slow." For a long time in the auto industry and even all of business, the hierarchical bias of "luxury brands aloof, mass brands humble" persisted; brands either ignored each other or competed covertly. This Mercedes–Wuling interaction precisely breaks that fixed pattern, hiding behind it an important shift from "fast-traffic involution" to "slow-value accumulation" in brand marketing — worth dissecting at its underlying logic.

The ultimate contest of marketing was never "fast traffic harvesting" or "hard price competition" but the establishment of "slowed-down user insight" and "value resonance." The reason the Mercedes–Wuling interaction moved people is essentially not some "high-EQ marketing trick" but that both stepped out of the utilitarian trap of "fast marketing," not fixing attention on "brand tier" or "short-term conversion," but both seeing the core group behind it: the strivers. This shared respect for users let two brands with vastly different positioning achieve synchronized resonance in users' minds, proving that the slow accumulation of brand value penetrates far more than fast traffic harvesting. Low-dimensional marketing competes on "fast" — who has more traffic, faster conversion, lower price; high-dimensional marketing competes on "slow" — who understands users better, builds emotional connection better, accumulates brand value better.

Setting aside surface things like product, price, and marketing, what truly makes users remember and trust you is the value you convey and the resonance of understanding them. The essence of a brand is consumers' "unique perception and trust" of you. Many brands addicted to the short-term pleasure of fast marketing harvest traffic through mass ads and low-price promotions, forgetting that slowing down to accumulate brand value is the lasting path. The viral Mercedes–Wuling interaction looks like a chance brand resonance but is essentially the inevitable result of long-term "slow marketing" accumulation: behind it lie three core tiers of marketing — the transaction tier (the "fast" of short-term selling), the product tier (the "bearing" of mid-term trust), and the brand tier (the "slow" of long-term resonance).

First clarify the brand's core value, then let the actions of the transaction, product, and brand tiers all accumulate around that value. In this process AI can be an efficient aid, but the key is using the right direction: don't let AI degrade into the fast-marketing poison of "mass harassment, traffic harvesting"; make it the slow-marketing antidote of "precisely delivering value, accumulating brand assets." Specifically: (1) Transaction tier: not just to sell, but to convey value at the transaction moment; AI can assist precise user-profile reach (e.g., analyzing recent user needs to push matching content), but set a daily reach threshold to avoid harassment, making the transaction the start of value delivery rather than traffic harvesting;

"Eliminating those at the bottom of growth is, at its core, eliminating cognitive inertia and adaptive inertia." In the AI-driven era of upheaval, the first principle of corporate survival has shifted from "efficiency competition" to "evolution-speed competition." Therefore, the underlying logic of the traditional "rank-and-yank" system must be rebuilt: the standard for elimination should not be the static "bottom of performance" but the dynamic "bottom of growth."

As the year-end approaches, many companies begin their reviews, and "rank-and-yank" becomes an especially sensitive topic. Some worry they will be classified as "bottom performers"; others envy companies that skip performance rankings, hoping to escape this pressure. A few days ago, Jensen Huang's November 2025 Cambridge-interview remark that he "does not do rank-and-yank" spread widely across friend circles and workplace groups. Many grasped it like a lifeline, believing "Silicon Valley has finally stopped the rat race" and "the iron-rice-bowl is the ultimate destination" — without realizing this is a monumental misunderstanding. The most dangerous illusion in the workplace is that "passing performance = job security." Many treat short-term sales or output as a "golden rice bowl," reading Huang's philosophy as "no need to chase performance blindly; just work at ease."

Market competition in the AI era is, at its core, a contest of "role-iteration speed" versus "organizational-evolution speed." For companies, failing to build a mechanism for individuals and the organization to co-evolve means not only being eliminated by external markets through organizational rigidity, but also letting the roles themselves be eliminated by the AI era by clinging to outdated positions without iterating; for individuals, failing to keep pace with organizational evolution — even with short-term passing performance — means either being replaced within the current role by someone growing faster, or being abandoned by the times along with the obsolete role. The first principle of role elimination is the speed of growth and evolution. Abandoning rank-and-yank does not mean coasting along; the roles themselves will be eliminated faster. Only by evolving together with the organization can individuals withstand this dual risk — and that is the true core of Huang's "no rank-and-yank": create a good environment and lead everyone to leap out of the trap of role elimination.

(1) Technology-paradigm disruption is accelerating, and role elimination is crueler than individual elimination. In the industrial age, a core technology's lifecycle could span decades, enough for personal experience to build a deep moat. But in the AI era, the cycle of technological disruption is shrinking to units of "years" or even "months": foundation models iterate continuously, new application paradigms keep emerging, directly redrawing the map of value creation and causing the industry chains that original roles depended on to break, shift, or vanish.

"The essence of digital employees is not how intelligently they work, but how human–AI collaboration is conducted." As the core vehicle for deploying AI in enterprises, digital employees are caught between the illusion of "deployment equals success" and the awkward reality of "investment equals idleness." The 2025 digital-employee track shows a stark polarization: on one side, a wave of efficient deployment in grassroots government, finance, and power sectors; on the other, some enterprises' digital-employee practices falling into the embarrassment of "seller's show versus buyer's show."

Behind the noisy surface of the digital-employee track, the answer is simple: digital employees never came to "steal jobs" but to "build the team." The reason enterprises fall into the "deployment dilemma" is the core problem of "managing intelligent tools with the old mindset of managing people": obsessing over "full automation, no human intervention" while forgetting that "human–AI collaboration" is its true core value. A digital employee is never a "replacer" but a "collaborator." It is not a machine that only repeats labor; it is a "digital colleague" with a closed-loop capability of "perceive–plan–act–learn." Every truly successful practice has drawn clear human–AI boundaries: letting digital employees take on repetitive, tedious mechanical work while humans focus on core links such as complex decisions, emotional interaction, and exception handling, ultimately achieving a "1+1>2" value increment.

Today, leading enterprises have long abandoned the misconception of "pure automation replacement" and shifted to a new production model of "human–AI division and collaboration." In finance, the Postal Savings consumer-finance smart digital human "Youxiaobao" handles 7×24 basic customer service, while human agents focus on complex complaints and high-value client service; in the power sector, the AI digital employee of State Grid Wuhan Power Supply Company covers high-frequency service scenarios, while human staff concentrate on special-business processing and emergency support.

"The long-term success of fast-moving consumer goods never comes from riding trends by luck, but from the precise capture and deep satisfaction of user demand." In November 2025, the "Hakimi" meme swept social platforms virally: Joyoung's Hakimi soy milk sold out 90,000+ units within one hour of launch, Douyin pre-orders exceeded 236,000 orders, Taobao cumulative sales surpassed 100,000 units, shipping was postponed all the way to 2026, and related check-in shares on social platforms exceeded 2 million — becoming a phenomenon-level hit.

The first principle of marketing is the inevitability of "discovering and satisfying demand." Every accidental hit is the inevitable result of demand capture. Copycatting can only skim short-term traffic; only by anchoring to marketing's first principle can an accidental viral moment turn into inevitable success. Marketing detached from demand — no matter how hot the meme — is water without a source.

The inevitable success of Hakimi soy milk lies at its core in precisely capturing Generation Z's implicit demand to "use memes for in-group identity and use cuteness to heal high-pressure emotions," carrying the demand with a physical product and forming a complete loop of "demand – product – sharing." Those brands that failed by copycatting violated this core principle exactly: they neither discovered users' real jobs nor provided corresponding solutions, merely consuming hot-traffic and naturally unable to accumulate long-term value. Kotler once emphasized publicly: "The core of marketing is not selling products but solving users' real problems, and that has never changed."

Marketing has now entered a new era of "meme economy + AI acceleration": in-group hot memes like "Hakimi" become natural carriers of dissemination, and AI tools (such as multi-platform public-opinion monitoring systems, intelligent content-generation tools, and data-analysis platforms) raise creative execution and demand-capture efficiency tenfold. Copycat brands only grasped the surface of "change": blindly replicating meme symbols and copying dissemination forms without using tools to dig into the real demand behind the meme, ultimately stuck at the "exposure" level and unable to form effective conversion — as Kotler put it in Marketing Management: "Tools serve demand, not the other way around."

"Blindly learning AI (chasing technical principles, hoarding conceptual tools) is a bubble chasing the wind; pragmatically using AI (solving business pain points, creating quantifiable value) is the shortcut to outlast the cycle." In 2025, giants' high-stakes bets triggered market worry: in September, Nvidia announced it would invest up to $100 billion in OpenAI to help the leading AI startup build a batch of massive data centers. In exchange, OpenAI committed to deploying millions of Nvidia chips in those data centers. This "circular transaction" involving hundreds of billions of dollars sparked strong market vigilance and deep concern over an AI bubble.

The current state of enterprise AI deployment is exactly a true portrait of "weak students hoarding stationery": many enterprises treat AI as a "universal patch," blindly learning technology and hoarding tools detached from real business needs, trying to cover business shortcomings with flashy AI concepts — self-developing a supply-chain large model for messy inventory management, copycatting AI marketing for low customer retention, force-fitting an industrial AI platform for poor production efficiency — yet never asking "does the business pain point really need AI to solve it" or "does the existing business logic support AI deployment." This misalignment of "using technology to make up for business deficiency" is destined to reduce "AI deployment" to a bubble.

An MIT report shows 75% of failed AI projects share the commonality of "blind self-development, detached from business" (source: MIT, The GenAI Divide: State of AI in Business 2025) — essentially "a weak student hoarding stationery yet unable to solve the problem." Domestically, reports of failed enterprise AI deployments are also frequent: a certain retailer with chaotic inventory turnover and missing procurement logic invested ¥20 million in a self-built general large model, neither optimizing the procurement process nor sorting out sales data, and the model ended up usable only for basic sales statistics; the inventory-overstock problem saw no improvement, the ROI fell below 1, and the project was forced to terminate 18 months after launch.

In sharp contrast to "weak students hoarding stationery," the core logic of "top students" like Yum China and Mixue Bingcheng is "solve the problem first, then pick the stationery": clarify the business pain point first, then use a well-matched AI tool to break through precisely, making AI a "booster" of the business rather than a "decoration." Yum China: first faced the business pain points of "high food-loss in stores and unreasonable labor scheduling," then launched the AI intelligent operations system "Q Rui," focusing on two scenarios — order prediction and labor dispatch — achieving a drop in per-store food-loss rate and a reduction in labor cost.

"Technology should serve human convenience, yet may degenerate into a fraud tool once its boundaries are breached!" In October 2025, China's AI industry reached a landmark node: leading products densely rolled out "long-term memory" features, pushing technological competition into the deep end of "human-like interaction" — but alongside the leap in efficiency came the fierce collision of blurred technical boundaries and risk prevention.

On the 16th, Alibaba's Tongyi Qianwen officially launched the "Qwen Chat Memory" feature, thoroughly breaking the limitation of traditional dialogue systems' reliance on short-term context. Through a dynamic memory-encoding architecture, it automatically identifies and records users' professional traits, interest preferences, and even language style, building a personalized knowledge graph — for example, when a user repeatedly brings up project-management topics, the system establishes a dedicated memory node, so subsequent conversations can offer precise suggestions without the user re-explaining the background. More importantly, its built-in console management and dynamic-forgetting mechanism allow users to view and delete memory content at any time, seeking balance between personalization and privacy security.

The large-scale deployment of AI memory features did not only bring the dividend of efficiency gains; instead it concentrated and exposed the core contradictions and ethical loopholes hidden behind the technology. From B-end enterprises' compliance demands to C-end users' privacy anxiety, from cognitive-level identity confusion to legal-level responsibility gaps, every problem points directly at the core proposition of "how technology should be deployed safely."

From the perspective of first principles, the essence of AI memory is "information reuse," but different user groups have vastly different needs and bottom lines for "reuse," forming three unavoidable core contradictions. The B-end's rigid demand for "compliance auditing" versus the C-end's basic demand for "privacy control"

"AI does not merely replace human labor; it frees up resources through productivity gains and reallocates them to high-value roles." The key to AI deployment is not technical implementation but the systematic rebuilding of organizational capability: technology without a strategic anchor is a "rudderless ship," strategy without organizational fit is a "castle in the air," an organization without cultural support is a "rootless tree," and without the supporting cast of data, technology, and change management, AI can only remain at the "pilot stage." Enterprise AI deployment must shift from "asking whether to use AI" to "asking how to build AI organizational capability" — defining strategy by business, breaking silos by architecture, driving collaboration by culture, strengthening foundations by data, fitting technology to scenarios, and securing deployment by change management — to seize the initiative in generative-AI competition.

The generative-AI deployment chasm refers to the gap between "AI technology's commercialization capability" and "the enterprise's organizational absorptive capability." As generative AI penetrates from "early adopters" to "early majority," the industry faces the contradiction of technological maturity versus lagging deployment: on the tech side, scenarios like NLP contract processing and intelligent approval have reached commercial maturity; on the deployment side, enterprise AI investment often falls into the predicament of "pilot is the endpoint." Gartner's 2024 Generative AI Deployment Report shows that in 2024 the average enterprise AI investment reached $1.9 million, yet fewer than 30% of CEOs recognized the return on investment, 60% of projects stalled at "unable to scale beyond the pilot," and 40% failed outright due to data-quality or cross-department collaboration issues.

The generative-AI deployment chasm refers to the gap between "AI technology's commercialization capability" and "the enterprise's organizational absorptive capability." How can enterprises cross this chasm? Goldman Sachs 3.0's AI deployment actually provides a reference answer. In 2025, Goldman Sachs — through CEO David Solomon, President John Waldron, and CFO Denis Coleman — jointly signed the "OneGS 3.0" plan, explicitly positioning AI as a "strategic productivity tool," focusing in the short term on five core scenarios (sales-process optimization, client-onboarding automation, loan-procedure digitization, regulatory-reporting efficiency, and supplier management) and building "AI + finance" competitiveness in the long term. This transformation is not "layoff-style replacement" but the integration of AI into the business fabric through structured organizational and workforce restructuring.

Goldman Sachs's workforce moves revolve around "optimize – add – transform": in Q2 2025 it had a net layoff of 700 people, a "routine annual adjustment"; simultaneously it expanded the tech-team scale and added AI-related job types; in addition, it launched an "AI–business integration" training program covering all roles from sales to risk control, driving the transformation of traditional roles into "human–AI collaboration roles."

How should we avoid becoming tools that are "knowledgeable but without judgment," and cultivate our own "taste"? On October 18, 2025, the century-spanning scientific giant Mr. Yang Zhenning passed away. This scholar, who once worked alongside Einstein and illuminated the map of modern physics with the "Yang–Mills gauge field theory," left the world not only formulas and theories, but also a profound question about "the essence of learning":

The "taste" in Yang Zhenning's words was never "taste" in the ordinary sense, but the judgment, appreciation, and intuition toward the essence of learning. "In every field of creative activity, a person's taste (I feel the word 'taste' is not quite fitting) together with his ability, temperament, and opportunity determines his style, and that style in turn determines his contribution. — Yang Zhenning." The core of this ability is "penetrating the surface of knowledge and autonomously judging what is important and what is beautiful." To Yang Zhenning, taste was the "core compass" on his research path.

No one is born with the ability to "judge the value of learning." Yang Zhenning's taste took shape gradually through decades of diverse nourishment, and this process happens to provide us with a reference path. Family scholarly enlightenment planted the seeds of humanities and "slow cultivation."

Today, AI can summarize a physics treatise or generate a research framework within seconds, yet it cannot replace the ability to "judge what is important and what is beautiful." In such an era, cultivating taste means both meeting the challenge and seizing the opportunity, with the core being "let humans lead and let AI be the tool."

"Man is an animal suspended in the webs of significance he himself has spun. — Max Weber" This National Day experience was especially surreal: humanity uses technology to construct a perfectly imagined virtual world on one hand, yet destroys the real world it depends on with weapons on the other. Turn on the screen and you see AI-native singer Yuri standing on a virtual stage singing "I am not a tool, I am not fictional, I have a name," every line a cry for "self-identity"; switch the scene and in Russia–Ukraine battlefield news, batches of nameless soldiers have no complete tombstones; on one side, Sora 2 generates virtual social scenes where even the micro-expressions of averted glances look just like reality; on the other, in the ruins of Gaza, children hide behind broken walls, and even a hot meal becomes a luxury.

We are used to examining AI from the "creator" perspective: "It is merely a machine executing programs, knowing rules but lacking a soul." Yet the speed of AI's evolution has long exposed humanity's grip: we laud "spirituality, sociality" as "irreplaceable sacredness" while ignorantly ignoring that "all higher values are rooted in animality"; we watch AI deconstruct our logic of meaning while still deluding ourselves that "it will never learn the unique human experience." This misalignment of arrogance and ignorance is precisely the blind spot we should most vigilantly guard against.

Humanity always prides itself on having "built a unique world of thought with symbols": binding coffee to "weekend relaxation," linking sunrise to "new hope," as if these meanings were humanity's exclusive "spiritual medals." Yet the truth is that our meaning construction is essentially also a fixed formula of "symbol + experience": "loneliness" = late night + shadow + racing heartbeat, "happiness" = salary arriving + parents' smiles + warmth in the stomach. These formulas AI is rapidly deconstructing with data. AI can write lyrics about "homesickness" not because it understands "nostalgia," but because it has learned the symbol combination of "hometown license plate + childhood snacks + mother's nagging."

We always assume "subjective experience" is the "irreplaceable ultimate defense line": "AI can simulate expressions of pain yet will never understand the sting of losing a loved one"; "it can calculate the repurchase rate of 7/10 sweetness, yet cannot taste the satisfaction of just-right sweetness." Yet what we are unwilling to admit is that this "unique experience" is essentially the product of "bodily signals + memory association"; without animality as its foundation, even the most "sacred" empathy collapses. The sting of losing a loved one is not a "tremor of the soul" but "the temperature the fingertips suddenly recall when seeing the cup the person used"; "the physiological habit of setting an extra place at the table when eating" — these are all memory traces left by the body. AI cannot simulate them now not because "human empathy is sacred," but simply because it lacks the bodily basis of "fingertips sensing temperature" and "the hollow burning in an empty stomach."

In June 2026, Anthropic released the landmark study "When AI Starts Building Itself," again publicly calling on the world's major AI labs to establish verifiable coordination mechanisms, to phase in a slowdown of frontier aggressive research, and to constrain the iteration speed of top-tier models. This was the company's second industry-level warning on technology-safety issues, after its 2024 joint call with global tech leaders for a six-month pause on AI research. The paper disclosed for the first time Anthropic's real internal R&D data: current AI can already execute 16-hour complex human work tasks with high accuracy, code-output efficiency is 8× that of 2024, and engineers self-rate their overall work efficiency as 4× higher.

After the paper's release, it sparked wide controversy worldwide. The technology-safety camp generally endorsed Anthropic's warning, arguing that the speed of AI's autonomous evolution has exceeded expectations and that existing alignment systems and risk-control mechanisms cannot match the capabilities of high-tier models; continuing unconstrained aggressive research may trigger unknown systemic risks. Some internal technical staff at OpenAI and Google DeepMind also publicly stated that the industry needs unified safety red lines to prevent any single lab's aggressive behavior from bringing disaster to the whole field. The business and political camp, however, generally took a skeptical stance. Some White House officials and venture-capital figures argued that leading enterprises' call for a research pause is essentially using the safety narrative to erect industry access barriers, suppress competition from smaller vendors and the open-source track, and consolidate their own market-monopoly position.