Vol.90: AI’s 2026 Acceleration Breakout—How Should We Position Ourselves? — A Conversation with 潘乱 + Koji

Summary

• 庄明浩用“拦不住了”概括2026年春节后的AI态势：模型跨过能力阈值，Agent、Skill、Cowork、OpenClaw（“龙虾”）和Seedance 2.0同时打开新场景。 GPT has released 3 versions in the past 3 months; the discussion noted that GPT-5.4 had just launched, following GPT-5.3. The apparent February release peak was more coincidence than pattern, because going forward “every period will be a dense period.” For investors, the question is no longer which launch to bet on, but which part of the value chain will show real usage first after the capability step-change.
• 算力和Token压力已在几十万OpenClaw部署者、数万名“日烧一亿SOTA Token”用户及限购的Coding Plan上显形。 A cloud provider’s R&D head initially estimated that more than 100,000 people were burning 100M Token a day, then revised the figure down to tens of thousands; 智谱, MiniMax, Alibaba Cloud, and others have also begun imposing purchase limits or tightening discounts. 庄明浩’s question is whether, if a technology revolution requires infrastructure to be overbuilt, the true scale may still need to be 1,000x or even 10,000x today’s level: “China lacks chips, the US lacks power, and everyone lacks Token.”
• SaaS是第一批被重估的资产，因为按席位收费本质上就是对白领座位收费。 The “2028 doomsday” thesis links white-collar displacement, weaker consumption and household balance sheets, zero platform friction, and corporate debt defaults into a self-reinforcing loop, but 庄明浩 keeps the caveat that the intensity, timing, and degree of government intervention are unknown. There is also no consensus on whether China’s lack of legacy SaaS can give To B Agent a shortcut; public-sector procurement and the government-enterprise cloud structure mean “no legacy systems” does not mean “no institutional friction.”
• AI把生产变廉价，却没有同步解决获客、信任与价值分配，真正稀缺的反而变成分发。 Small teams can use Vibe Coding to handle product, engineering, design, and operations, potentially shrinking organizations to “a handful of superplatforms plus a mass of sole proprietors.” But a product burning $1,000 a day on Token can still get stuck at cold start, while a cross-border company with $7M-$8M in annual revenue may still spend $20K a day buying traffic from Facebook. 潘乱’s summary of the brutal economics of content is “six months, RMB10B, 8%”: the market scales rapidly, while creators may receive only 8%, or less.
• 当前较硬的应用证据仍包括Coding和被拆解的人机协作流程，而非泛化的“一人公司”口号。 Karpathy said his workflow flipped within a year from 80% handwritten code and 20% autocomplete to 20% handwritten code and 80% autocomplete. But Koji stresses that burning Token only in Coding without redesigning QA, Infra, and PRD processes does not make the organization more efficient; more code is not the same as higher productivity. Manhua dramas offer another case: a pipeline broken into 5 or 6 clearly specified deliverables, with a few people plus “抽卡师,” converting video and image model capabilities into production efficiency.
• AI的普惠价值可能比白领提效更早触达老人和孩子。 潘乱 installed Doubao for his illiterate mother, who only knew how to use WeChat and Douyin, so she could speak with it in her dialect; his father, meanwhile, proactively asked AI for a speech for the groom’s parents at his son’s engagement. 庄明浩 believes Doubao, Yuanbao, and Qwen may be the highest intelligence many Chinese people have ever encountered, while AI is also trying to absorb the downstream friction left by search and deliver answers in one place.
• 面对焦虑，潘乱和Koji保留了一组有价值的张力：不要追逐每个短命技巧，但也不能忽视边界处正在形成的创业机会。 潘乱’s strategy is, “If I learn slowly enough, I don’t have to learn,” starting with the question of what he actually needs to solve and training himself to be the boss of AI. Koji’s counterpoint is that as the frontier of intelligence keeps expanding, version and product developments remain directly relevant to entrepreneurship and investing. 庄明浩 uses AlphaGo to describe the mindset shift: from dismissal and skepticism to having one’s conviction shattered (“道心碎了”), after which the task is to rebuild one’s mental model rather than pretend nothing has changed.
• AI硬件与终身陪伴Agent会继续推进，但入口、隐私和组织形态远未定型。 庄明浩 relayed a policy framework that identifies phones, computers, embodied intelligence, and new-energy vehicles as priority integration areas; big tech sees glasses as the next entry point, while Xiaomi’s glasses chief believes they cannot mature within 10 years. 潘乱 believes an AI companion that records a person from birth to death will inevitably emerge, but who builds it, how it operates, and how humans participate and constrain it remain open questions.

Deep dive

1. The 2026 Lunar New Year Was Not a Release Peak; Every Period Will Be Dense

• 庄明浩 said the Lunar New Year barely felt like a holiday: 智谱 and MiniMax stocks “shot through the roof,” OpenClaw’s momentum carried from before the holiday through the break, and ByteDance’s Seedance 2.0 continued improving on an already powerful video model.

• He saw late January through the run-up to the holiday as a cluster of major releases from China’s leading model companies. Seedance 2.0 had already launched, while DeepSeek V4 was merely “rumored to be coming in the next couple of days” at the time of recording; the two should not be treated as equally certain.

• 庄明浩 rejects the idea that February has a special release pattern: “That was an accident.” GPT released 3 versions in the past 3 months; GPT-5.4 had just launched, following GPT-5.3. The real shift is that pretraining did not end as many expected in 2024; the Transformer path is still accelerating.

2. OpenClaw’s Breakthrough Is Not a New Capability but a New Way of Imagining Demand

• 潘乱 believes OpenClaw has not actually done much that previous tools could not do, especially for people who already used AI tools extensively. What it does is force users to rethink what else AI might be able to do for them.

• Koji spent 2 hours installing OpenClaw and only discovered the next day that people on Xianyu were charging RMB400 for on-site installation. He had the Agent scan the OpenClaw ecosystem every day for new products and startups, identify Chinese founders, fill in their backgrounds and contact details, and draft outreach emails. That process led him to VisionClaw, an open-source infrastructure project connecting Meta glasses to OpenClaw that ranked first among similar projects on GitHub; its author was a Chinese student at New York University.

• 潘乱’s first OpenClaw task was more mundane: check every morning at 9 a.m. whether a long-out-of-stock pair of Li Auto glasses was available, and buy it if stock appeared before the discount code expired on March 31. Scheduled tasks were possible before, but “the lobster” packages configuration, persistent execution, and task delegation into a product users can feel.

3. Personalized Production Revives the Long Tail but Turns Many Products into One-Off Content

• After installing a Bambu Lab 3D printer and printing himself a fidget spinner, 庄明浩 felt that “communism had arrived early”: instead of buying a standardized industrial product, “I manufactured an object for myself,” with material supply beginning to serve tiny groups or even a single individual.

• 潘乱 extends the same logic to Vibe Coding: one person can take on product, engineering, design, and operations, without having to prove a massive market and a fully staffed team on page 1 of a BP. They only need to solve a problem for themselves, people around them, or a small user base. “The true revival of the long tail” becomes possible.

• He then qualifies the optimism: when applications can be generated quickly, they may simply be a new form of content. A product like “死了么” is more an online curiosity and check-in destination than an application on the scale of Ele.me; the more abundant supply becomes, the less valuable each individual product is.

4. The Lunar New Year Red-Envelope War Was a Prisoner’s Dilemma That Hid the Releases Behind a Traffic Narrative

• 庄明浩 argues that Chinese companies were not merely “handing out red envelopes and ordering milk tea” during the holiday. The genuinely important event was the release of major model versions, but several internet companies entered a prisoner’s dilemma, allowing the marketing war to receive “far more attention than it should have.”

• Sponsorship of the Spring Festival Gala reportedly started at around RMB100M, quickly rose to RMB300M and RMB500M, and may ultimately have approached RMB1B. Dreame was said to have spent RMB800M-RMB1B; Doubao may have paid more because it received more exposure, while humanoid-robot companies each spent several hundred million yuan. On the second day of the Lunar New Year, the State Administration for Market Regulation told several companies to stop escalating, and the frenzy “came to an abrupt halt.”

• 庄明浩 believes the final ranking held no surprises: Doubao remained ahead, MiniMax had a surge, and the third player was broadly the same as before. 潘乱 adds that the Spring Festival Gala represents one centralized burst of nationwide traffic each year, though Douyin employees told him that the peak audience for Germany vs. Japan in the 2022 World Cup group stage was even higher than the Gala; the figure was not officially disclosed.

5. The “2028 Doomsday” Thesis Turns White Collars, Platforms, and Debt into a Self-Reinforcing Loop

• 潘乱, borrowing 王慧文’s framing, offers a reversal: Chinese SaaS fundraising historically used Salesforce as a valuation benchmark, but after AI arrives, it is not that Chinese SaaS companies become as valuable as American ones; “American SaaS companies become as worthless as Chinese SaaS companies.”

• 庄明浩 explains the first mechanism: US SaaS charges by seat, and a “seat” is a white-collar job. Once AI replaces white-collar workers, companies use AI more aggressively to support their share prices and raise efficiency, which accelerates AI development and leads to more layoffs. Unemployment weighs on consumption and puts mortgages, credit cards, and other household debt under pressure.

• The second mechanism targets businesses built on friction: real estate, travel, distribution, intermediaries, and two-sided platforms all monetize information and transaction costs. If Agents reduce friction to near zero, users can handle recommendations and purchases directly, putting platform margins under severe pressure. The third is that companies cannot repay acquisition and expansion debt, eventually triggering a debt crisis.

• 庄明浩 stresses that this is merely a “very internally coherent” story; nobody knows whether events will reach the scale described or happen on that timeline. The irony is that the author priced the article’s Substack subscription at $999 and reportedly added tens of thousands of paid subscribers in a week: the doomsday narrative first paid off in the creator’s own tens of millions of dollars in subscription revenue.

6. China’s Lack of Legacy SaaS Does Not Automatically Give To B Agents a Mobile-Payments-Style Tailwind

• The optimistic analogy comes from payments: China lacked a credit-card system, allowing WeChat Pay and Alipay to sweep through the market, while legacy SaaS is deeply embedded in US enterprise workflows. Many founders therefore believe Chinese To B Agent companies can bypass old systems and reignite enterprise-services entrepreneurship.

• 庄明浩 says he is “as skeptical as I was 10 years ago.” In China’s predominantly public-ownership system, government and large institutions can choose Tianyi Cloud, China Unicom Cloud, or Huawei Cloud “without needing any explanation”; choosing Alibaba Cloud requires an explanation. He cites the stagnation of Alibaba Cloud’s share 5 years ago alongside growth at those other providers as a counterexample to the simple analogy.

• His conclusion is not that Agents do not matter, but that resistance in US SaaS should not be mechanically translated into an advantage for China. Critics of the doomsday thesis also point out that the original article underestimates government: its responsibilities around unemployment and financial stability cannot be ignored.

7. The AI Capex Paradox: The Upstream Takes Most of the Value, but Nobody Can Explain the End-User ROI

• 庄明浩 says no one can currently answer the questions of Chinese companies’ monetization pace or capex ROI. Everyone is “selling shovels,” but history shows that Qiandao Lake did not become the most valuable company simply by supplying water; the biggest gains usually went to the companies that delivered products to consumers.

• That makes this AI cycle the inverse of the mobile-internet era. Chip, compute, component, and infrastructure suppliers are temporarily taking a larger share of the value pool, while it remains unclear whether end-user applications can absorb the investment. 庄明浩 sees this as the opposite of mobile internet’s value-distribution logic.

• 庄明浩 revisits Nvidia: it fell 17% on the day DeepSeek R1 launched, traded down into the $80s during the US-China tariff war, then rose to around $180, and has been almost flat over the past 6 months. “Has it already peaked, or is it holding back something bigger?” He has no answer.

8. Genuine Infrastructure Overbuilding May Still Need to Be 1,000x to 10,000x Today’s Scale

• 庄明浩 and the AI discussed the numbers repeatedly; the number of Chinese users who have actually deployed OpenClaw may be only in the hundreds of thousands, spread across different model APIs. Even so, many model companies other than the largest few are already facing significant compute pressure.

• Every industrial revolution went through infrastructure overinvestment, with excess capacity later giving rise to downstream applications. If hundreds of thousands of users are nowhere near mass adoption, 庄明浩 estimates that infrastructure spending large enough to qualify as “overbuilding” may need to reach 1,000x or even 10,000x today’s level. The tension is that current compute construction is already approaching the limits of what humanity can build.

• 庄明浩 cites the “100M Token club”: based on an internal observation by an R&D head at a cloud provider, the initial estimate was that more than 100,000 people were consuming 100M Token a day; after further consideration, the conservative revision was tens of thousands. And they were not burning 7B small models, but the most intelligent SOTA models.

• 智谱, MiniMax, Alibaba Cloud, and other AI Coding Plans have begun imposing purchase or quota limits and tightening discounts. 庄明浩 offers a supply-and-demand summary that can be monitored: “China lacks chips, the US lacks power, and everyone lacks Token.”

9. Burning Token Does Not Automatically Raise Productivity; Organizations Must Be Rebuilt Like Electrified Factories

• 庄明浩 uses the history of electricity to explain the lag in adoption: it took roughly 40 years for electricity to materially change factories after its invention. Early factories were constrained by water sources, power transmission, and multistory layouts; only after factories moved to open-plan floors and production lines were reorganized did electricity change overall efficiency.

• Coding works the same way. If a company burns huge amounts of Token only at the code-writing stage without redesigning QA, Infra, and PRD definition, the result may simply be “more code than at any point in history,” with no improvement in overall organizational efficiency. AI productivity comes from rebuilding the entire workflow, not from increasing calls at a single point.

• Software iteration is not subject to the same atomic-world constraints, so this transformation may happen faster than 40 years. The point of the electricity analogy is not to predict a timeline, but to remind investors that real returns depend on complementary processes being rebuilt; Token consumption alone is not enough.

10. Agents Need a 24/7 Host, Bringing the Private Cloud Back into the Consumer Imagination

• 潘乱 believes phones are designed to capture attention and display screens, making them poor devices for running tasks over long periods. Computers may not be the final form either. What people fundamentally need is a private server that stays online and can run an Agent.

• He mentions a new company founded by another ByteDance angel investor that reverses ByteDance’s old story of centralized recommendation engines, giving every person a “private Alibaba Cloud” or “private Volcano Cloud.” The device’s crowdfunding price was close to RMB20K, and sales performed well after OpenClaw took off.

• Manus also briefly described itself as “a cloud virtual machine for everyone,” though users found the phrasing difficult to understand. 潘乱 would personally rather pay for 1TB of local storage than the same amount of cloud storage: as Agents gain access to more permissions and personal data, security, privacy, and lower Token subscription costs will become reasons to buy local compute.

11. The Model War Is Hitting the Physical Limits of Power, Capital, and Talent Density

• 庄明浩 observes that model capabilities did not slow in early 2025; they accelerated continuously from 2025 through early 2026. Every capability improvement continued opening Agent, content-generation, and workflow scenarios, even if some temporarily lacked commercial viability.

• The battle is simultaneously approaching “the limit of electricity, the limit of money, and the limit of talent density.” He says he does not know who can keep going after linear expansion continues, what the endgame will look like, or whether a higher-order actor might emerge to force the competitors to stop.

• His concern is not merely whether one company can raise capital, but that technological progress is approaching nontechnical problems involving unemployment, wealth distribution, and the functioning of society. It is easy for observers to state a position; model companies, application founders, and ordinary employees will all struggle to remain untouched.

12. If AI Becomes a Weapon, Regulation May Shift from Product Governance to Arms-Control Logic

• A major rebuttal to the “2028 doomsday” thesis is that it ignores the government’s role. 庄明浩 cites the title-like question of an article: “If AI becomes a ‘weapon’ in the future, should we manage AI the way we manage weapons?”

• 潘乱 remains deeply skeptical. Ten years ago, Netflix attributed the success of House of Cards to big data, but the subsequent wave of “data-guided creative production” failed to deliver; today’s claims about AI guiding warfare may contain similar exaggeration. He says plainly that he does not understand AI’s actual contribution to targeted arrests in Venezuela or the war in Iran.

• What he can verify with his own eyes is that drones mass-produced in China are being used in warfare with high cost-effectiveness. The disagreement remains unresolved: 庄明浩 argues that AI may have the attributes of a weapon, while 潘乱 insists on distinguishing algorithmic narratives from proven advantages in hardware and cost.

13. AI Is the New Camera, but Capability Growth Is Outrunning Ordinary Human Perception

• 潘乱 places Seedance 2.0 in the history of creators: the mobile internet made more people creators through the camera, and the two-sided platforms with the most creators ultimately captured the most value. AI is “the new camera,” capable of generating content without even taking a real photograph.

• 潘乱 uses height to describe model upgrades: once a creature is already far smarter than you, it is hard to tell whether it grew from 2.20 meters to 2.21 meters or to 2.50 meters. Later, answering an audience question, 庄明浩 switched to IQ: “If AI is already at 150, it is hard for people at 120 to feel the difference between 150 and 160 or 170.”

• Karpathy supplied an observable result: a year ago, he wrote roughly 80% of his code by hand and relied on autocomplete for 20%; a year later, the ratio had completely reversed. 潘乱 therefore argues that most people feeling nothing when a model launches does not mean capability has not changed; they should still actively test what today’s AI can do.

• Koji adds that scoring AI has itself become an area of entrepreneurship and research. Mathematics and programming have clear answers, but open-ended questions are difficult to evaluate; much of the work of post-training and reinforcement-learning teams at model companies consists of designing new evaluation methods.

14. AI Is Giving Illiterate Elderly People Access to the Smartest “Person” Around Them

• During the Lunar New Year, 潘乱 installed Doubao for his mother. She is a rural woman who cannot read and previously knew only WeChat’s text-to-speech function and Douyin; she can now speak and video-chat with AI in her dialect without the burden of typing or the psychological pressure of “asking the wrong question.”

• He also discovered that his father was asking AI what the groom’s parents should say at his younger brother’s engagement. The moment made him reconsider the assumption that older people lacked information needs; search products were simply so poorly designed that complex interfaces kept large numbers of people away.

• 庄明浩 believes Doubao, Yuanbao, and Qwen “should be the highest intelligence Chinese people have ever encountered,” and may even be “the smartest person” many will meet in their lifetime. Friends who once used Douyin to keep their children quiet can now have them carry on conversations with Doubao.

15. Search Pushes Friction Downstream; Generative AI Tries to Put the Answer in Front of the User

• 庄明浩 recalls that Google once used a counterintuitive metric: the shorter users stayed on a page after searching, the better, because it meant they had found the answer. In retrospect, that may simply have shifted the burden of reading, comparing, and judging to downstream websites.

• 庄明浩 believes AI products are trying to put the answer directly in front of users in one step. That is why Agents threaten not only the search entry point but the entire commercial chain that depends on the friction of clicks, comparison shopping, and handoffs.

• 潘乱 further challenges Google’s historical halo: its mission was to organize the world’s information, but before the AI era it primarily organized webpages; its attempts at recommendations, UGC, and current-events news were not especially strong. AI changes the picture by turning the web it accumulated over decades back into an advantage in training and answering.

16. This Is the Best Era for Founders, but Also the Era of the Most Copycats and the Hardest Pricing

• 庄明浩 believes Manus, OpenClaw, and Lovable prove that small teams do not need to be in Silicon Valley to build global To C products from China, Singapore, or Europe and compete with the world’s best companies. As intelligence keeps unlocking new boundaries, the moats of old incumbents are being eroded.

• The worst part is that production costs have collapsed. Once OpenClaw took off, the most obvious opportunity was to lower the installation barrier, and thousands of one-click deployment products appeared overnight in China, Japan, Europe, and Singapore. Every business needs differentiated value before it can set a price, but differentiation itself is becoming shorter-lived.

• 潘乱 therefore sees a return of “a hundred flowers blooming” and “Dream big, dream crazy.” Even a flying companion robot that existed only as a video demonstration could attract discussion and capital enthusiasm. His rebuttal is that TikTok was the first foreign company to influence Americans’ thinking at scale; a social product need not be less consequential than a flying car.

17. Space Compute and Flapping Aircraft Are Nonlinear Bets Forced by Physical Limits

• 潘乱 mentions a group of teams proposing to build data centers in space. Many problems are already approaching the limits of human physics, and continuing to solve them through conventional linear methods may not be enough. The real uncertainty is whether delivery takes 5 years, 10 years, or 30 or 50 years.

• Flapping Plane raised hundreds of millions of dollars in its seed round. The name refers to a bird flapping its wings: the founder believes the Transformer path can produce only visible, linear gains in intelligence, and that pursuing AGI requires trying stranger approaches.

• 潘乱 says Silicon Valley calls these companies “Neo Labs,” represented by new laboratories founded by people such as 李飞飞 and willing to put large sums behind foundational-path innovation.

• 庄明浩 says the difficulty of rebuilding China-style ultra-high-voltage power grids in the US may exceed Elon Musk’s challenge of sending Starship into space and bringing it back. He adds an IKEA analogy: beyond raw materials and instructions lies “process knowledge,” embedded in workers’ muscles and neurons. Without that knowledge, US manufacturing costs could be 10x or 20x higher, or the products might not be manufacturable at all.

18. The Most Effective Antidote to Anxiety Is to Stop Chasing Every New Thing and First Identify the Problem

• 潘乱’s middle-aged version of the answer is: “If I learn slowly enough, I don’t have to learn.” When he was younger, he would clear out Google Reader, fall down Wikipedia rabbit holes, and follow the entire Fanfou timeline. Now, even after subscribing to several AI products, he often forgets to use them because many updates are merely “fast-food news and fast-food applications.”

• He once studied prompting courses, video keyframe controls, and other techniques intensively, only to see them become obsolete as soon as a new model arrived. He compares it with independently proving the Pythagorean theorem in elementary school, excitedly showing the result to his teacher, and learning the following week that the class had already covered a better established proof. The shorter the knowledge half-life, the more questionable the payoff from being one step ahead.

• His filter is to ask first, “What do I want to do? What problem do I want to solve?” and then make AI serve that objective. Instead of learning to become a better worker-bee, learn market judgment, task decomposition, and resource deployment — learn “how to be the boss of AI.”

• Koji offers the opposing view: if 6G or 7G arrives, most people only need to use it and do not need to study the underlying technology. But the AI frontier matters because it is still where entrepreneurial and investment opportunities are forming. Ordinary users can reduce the noise; founders and investors cannot treat every development as noise.

19. The 林俊旸 Episode Shows How Peripheral Technical Talent Can Become a Company-Level Risk Factor in Days

• 庄明浩 conducted an informal survey on the spot and found that few people had heard of 林俊旸 a few days earlier, while almost everyone in the room knew the name after the incident. He sees this as a revaluation of talent in the AI era: a seemingly unremarkable 90s-born person can suddenly acquire global influence.

• According to 庄明浩’s account at the event, this person’s moves were enough to send Alibaba down 4%-5% in a single day and attract the core leaders of the 3 best laboratories in the world to compete for him. Whether he would be allowed to leave the country might even become a subject of discussion at a higher level.

• This is not a conclusion about one incident, but an observation about how the media and capital price people: when a small number of model experts can alter technical direction and organizational confidence, people once buried deep inside teams can rapidly become public figures, valuation variables, and national talent-policy issues.

20. AlphaGo’s Real Lesson Is Not Losing, but Rebuilding After “道心碎了”

• 庄明浩 recommends revisiting the documentary on AlphaGo’s match against 李世石 9 years ago and listening to 樊麾’s recollections. As a European Go champion, 樊麾 initially dismissed the machine. He attributed the first game to a personal mistake, could no longer understand what was happening in the second, was quickly defeated in the third, and did not continue with the final 2 games of the scheduled 5-game match.

• By the time AlphaGo challenged 李世石, 樊麾 already expected a 5-0 result, but professional players around the world did not believe it. After the first game, 李世石 likewise said, “I didn’t play well.” In the second, he and the commentators initially laughed at a move that looked absurd, then gradually realized something was wrong as his expression shifted from relaxed to tense.

• After 3 straight losses, the on-site team should have been celebrating proof that machines had surpassed humans, but “no one was happy.” 樊麾’s description was: “李世石’s conviction was shattered” (“道心碎了”). In the fourth game, 李世石 played what later became known as the “God move,” and 樊麾’s eyes lit up: “His conviction had come back.”

• 李世石 withstood the new pressure created by the possibility of winning and took the fourth game. 庄明浩 calls it humanity’s last victory over AlphaGo and says 李世石 was unbeatable in the world for several years afterward. He maps the experience onto 4 stages: dismissal and incomprehension; doubt and perception; being surpassed and shattered; then accepting reality and finding a new position.

21. The Actionable Starting Point Is Not Studying AGI but Reworking Tasks Repeated 5 Times a Week

• Koji’s first piece of advice for anxious people is extremely simple: list the tasks in work, entertainment, and daily life that recur more than 5 times a week, then ask whether AI can make them faster. This is more likely to generate consistent feedback than vaguely “learning Agents.”

• Koji explains that much of the Token burn still comes from Coding, but users are not necessarily “programming for the sake of programming”; the model implements tasks in code. For ordinary people who simply want to make their first RMB1,000 or RMB10K, he jokes that they could install OpenClaw for others on Xianyu, then seriously suggests starting with problems in their own industry.

• The vertical opportunities he sees include AI-assisted breeding, genuine mineral exploration, new-material discovery, medical records, dental front desks, and solar-installation crew management. When he reviewed the 150 companies funded by YC 2 years ago, more than 70% were already “AI plus an industry”; China is now beginning to produce more granular, localized combinations.

• Koji sees manhua dramas as an efficiency case enabled by stronger video and image models: a few people split production into 5 or 6 stages, each with explicit inputs, outputs, and delivery standards, while humans handle “抽卡” or connect the stages. 庄明浩 pushes the point further, criticizing Youku, iQIYI, Tencent Video, and Mango TV for remaining like traditional television stations moved online, pursuing only a few premium productions rather than building platforms that let more creators enter.

22. AI Hardware Will First Land in Phones, Computers, Embodied Intelligence, and New-Energy Vehicles, Not an All-Purpose New Terminal

• 庄明浩 relays that the hardware section of the State Council’s “AI Plus” policy documents highlights 4 categories: phones, computers, embodied intelligence, and new-energy vehicles. He believes the definition has a practical basis: these products either already have enormous installed bases or have demonstrated demand for combining AI with the physical world.

• Phones are the most numerous, but they are designed around screens and attention, not OpenClaw-style background workflows. Computers are better suited to task execution, but may still be transitional hosts. The distinction between the two will determine whether an Agent is a chat assistant or a digital employee that works continuously.

• Glasses should be viewed separately from other smart hardware because big tech sees them as a new entry point and hopes to escape the constraints of Android and Apple. 庄明浩 relays the Xiaomi glasses chief’s judgment: “Impossible within 10 years.” At most, the product is currently at the iPhone 1 stage, with no short-term solution for ecosystem, battery life, or configuration.

• Niche products such as voice recorders and music devices have already produced successful cases, but 庄明浩 offers a rough screen: watch whether Apple enters. Apple has long limited its expansion beyond computers, phones, and tablets to products such as AirPods and Apple Watch, with glasses still exploratory; shipment volume and high-frequency demand remain hard thresholds for hardware viability.

23. Organizations Will Split Between Superplatforms and Sole Proprietors, While Lifelong AI Companions Push Privacy to the Extreme

• Asked whether “one-person companies plus ultra-large companies” could become the end state, 潘乱 says the founders most worth backing at the angel stage will still need exceptional leadership and charisma. The difference is that they may organize “a group of OPCs” to collaborate rather than maintain traditional hierarchies.

• 潘乱’s structural judgment is sharper: “Big platforms plus a mass of sole proprietors,” with mid-sized companies collapsing systematically. Sohu-, Zhihu-, and Sogou-style organizations have neither the concentrated resources of superplatforms nor the cost and speed advantages of individual teams.

• One entrepreneur imagines an AI that accompanies and records a person continuously from birth to death, ultimately understanding that person’s talents, health, and decision preferences better than the person does. The questioner worries that this would be “unethical” from a privacy standpoint. 潘乱 does not reject the vision; he instead says “this will definitely happen.” The real questions are who builds it, how it is built, how people participate, and how humanity establishes checks and balances.

• Asked what he would never let AI touch, 潘乱 cannot think of anything for now. 庄明浩 jokes about using it in the bathroom, and 潘乱 immediately concedes that he would still ask AI when physically unwell. He places his hope in humanity continuing to invent ways to unload, limit, and fight algorithms, rather than imagining that technological progress will stop.

24. The Cheaper the Supply of Products, the More Expensive Distribution, Trust, and Credibility Signals Become

• One founder describes the cold-start problem: an OpenClaw product consumes $1,000 a day and still struggles to recover its costs after launch. Another cross-border company generates $7M-$8M in annual revenue but spends $20K a day buying traffic from Facebook, with most of the revenue ultimately flowing back to social media. AI has lowered production costs without creating a new attention platform.

• 潘乱 therefore says “building an IP is still very important.” As content and applications multiply without limit, the people, brands, and communities that can serve as trust intermediaries become more valuable.

• 庄明浩 adds that if a project is sufficiently cutting-edge, podcasts are a relatively good way to build trust and distribution. Unlike most promotion channels, they do not offer only a clearly priced performance outcome and can still generate unexpected word of mouth.

• 潘乱 summarizes the recent economics of content: “six months, RMB10B, 8%.” Casual games, bullet-screen games, short dramas, and manhua dramas can all scale to RMB10B within 6 months, while content producers may receive only about 8%; manhua dramas may get even less. This may be the normal outcome once the infrastructure is highly developed.

• Traditional revenue and profit no longer explain the full valuation picture. Why 寒武纪, MiniMax, and 智谱 are valuable may have more to do with narrative and strategic position than current earnings. Asked how to assess an algorithm lead, 潘乱 suggests starting with paper count and citations, while candidly admitting that he does not understand algorithms well enough to give a high-quality answer — knowing the limits of one’s signals matters too.