Vol.93 In the Face of AI, the Internet Is Actually a Traditional Industry—A Cross-Show Appearance, Unusually Serious
Summary
- AI did not wipe out traditional industries first; it turned the internet itself into one. 庄明浩’s view is that short video, advertising, e-commerce and content platforms have accumulated deep know-how and become locked into rules set by algorithms and traffic buying. Large models can move quickly from 0 to 60–80 points, but businesses need to grind from 99.5 to 99.6. The key variable in early 2026 is therefore not how much general capability improves, but whether each vertical can complete a long, tedious and potentially unsuccessful final-mile integration.
- Six months, RMB10B and 8% capture the brutal profit structure of the platform-era hot streaks. Mini-games, short-form dramas, manhua dramas and live-stream chat games can all reach roughly RMB10B in about 6 months, but platforms and traffic acquisition absorb most of the value, leaving producers with perhaps 8%—or even less in manhua dramas. A hot streak is more like squeezing a pimple: technology or a business model creates a temporary outlier, then algorithms and a flood of competitors quickly flatten ROI, leaving behind “illusory fireworks.”
- AI’s real commercial breakthrough is the possibility of making some services both product-scalable and individually tailored. A minute of traditional animation costs about RMB10K, while AI manhua dramas can bring that down to a few hundred yuan; wedding designs, music, video and other POD products might be delivered by 2 or 3 people through a workflow. But humanities content allows more error, scientific fields such as healthcare demand precision, and physical POD remains constrained by transfer printing, UV printing, digital spray-painting and flexible supply chains. “Can be done” and “worth taking to 100 points” remain two different questions.
- As content production inflates without limit, algorithmic recommendation may give way to personality, trust and a retro return to editorial curation. Douyin already receives roughly 100M video uploads a day, and AI may increase supply not by 10x or 100x but by an order of magnitude that is harder to imagine; a uniform algorithm optimized for the lowest common denominator will only produce more “AI slop” and “prepackaged food with no wok hei.” 庄明浩 is therefore betting on polarization: one end will be super-giants no one can catch, while the other will be a super-long tail of super-individuals serving small groups of loyal users.
- When the answer is unknown and the waterline keeps rising, feedback speed is worth more than a complete plan. “Even if what you build is shit, get it out before everyone else” is not a call for shoddy work; it means securing excess attention from users, capital, media and competitors, then using that feedback to compound the next move. A team hiding underwater to build a grand reveal may find its project obsolete 3 months later. PLAUD shows that launching first can also catch an external boost from a new GPT model and stronger multimodal capabilities—luck that cannot be written into a plan.
- AI hardware is an investment consensus, but also an inventory trap between models that iterate monthly and the physical world’s 6- to 18-month cycle. The intersection of China’s supply chain, advanced manufacturing and AI makes glasses, toys, companion devices and learning hardware worth testing, but users imagine “Iron Man” and may receive something they experience as “idiotic”; some products are described as having return rates close to women’s apparel. CK’s operating rule is blunt: the product can be a little worse, the order smaller, the cost higher and the loss acceptable—“but never, ever carry inventory.”
- China’s To B opportunity may not be to copy US SaaS, but to productize the 80–90-point expertise in industry veterans’ heads and sell outcomes directly. The development of agents and coding in the US may let individuals and companies build more customized software themselves; China’s legacy SaaS market never fully worked, which may mean less historical baggage. The real gap is the data cleanup, knowledge bases, ERP retrofits and on-site implementation required before traditional businesses can use models directly. These unglamorous jobs may be what turns AI revenue from wrapper demos into sustainable delivery.
- AI is making giants bigger for now, but it has not eliminated niche markets; it has only raised the cost of competing for the crown jewels. ByteDance can invest in video models and distribution, while Tencent can spend on AI gaming in ways many startups could not even imagine; other companies will struggle to close the gap head-on. Smaller companies are better off targeting specialized needs the giants cannot cover. Just as On, Hoka, Salomon and Arc’teryx grew beyond Nike and Adidas, the future may feature giant trees pulling further away while vast amounts of “moss” fill the long tail.
Deep dive
1. AI Turned the Internet Itself into a Traditional Industry First
庄明浩 started the podcast because technology is changing too fast. He needs to consolidate each phase of his thinking and “say goodbye to it properly.” The joke is also true: “If you don’t learn, you don’t have to keep learning”—because a concept understood today may already be obsolete a little later.
His core observation is that social media, entertainment, short video, gaming and e-commerce—the “new economies” of the past—now have complex supply chains, equipment, experience and know-how invisible to outsiders. In the face of AI, they are no different in kind from any traditional industry.
The tension is that large models can move from 0 past 60 points, even to 80, in a very short time, while online businesses are already competing at the limit and optimizing from 99.5 to 99.6. Closing the gap between general capability and vertical requirements may require years of work from practitioners in the old economy; some sectors may ultimately conclude that “no one can save me.”
2. Six Months, RMB10B and 8% Reveal Where the Hot-Streak Economics Accrue
After short video consumed users’ time, platforms’ advertising, recommendation and traffic-buying systems largely fixed the ROI at every layer. Content, e-commerce and brands have all been thrown into the same heavily fortified furnace, leaving them to hunt for short-lived outliers that algorithms have not yet flattened.
Mini-games, short-form dramas, manhua dramas and live-stream chat games keep following the same path: the moment the market hears that someone is testing one, the category can roll to roughly RMB10B in about 6 months, followed by industry conferences, forums and a rush of entrants. 庄明浩 compares these peaks, forced higher by mature platforms and traffic-buying systems, to squeezing a pimple.
The people actually making the content may end up with only about 8%, and in this round of manhua dramas, even less than 8%. Platforms take a large cut, while producers still have to buy traffic. Once algorithms identify an outlier, they flatten its returns; latecomers accelerate the process like snowflakes in an avalanche. What remains is “illusory fireworks.”
3. AI Compressed 3 Quarters into 2 Eras
Manhua dramas heated up in the second half of last year, yet by the time the program was recorded, after only 3 quarters, they already felt like a mature project. 庄明浩 says that in the AI world, “3 quarters is a long time—it is already a story from 2 different eras.” DeepSeek 2.1 had also been around for only a little more than a year.
That speed has directly changed organizational cadence. A sector presentation updated once every 3 months might have been adequate before; now, materials finished before Chinese New Year can be obsolete after the holiday. By March, new topics, model vendors and version numbers push everyone into the next phase.
4. AI Is Productizing Services, but Physical Delivery Still Runs into Supply Chains
张伟伟’s framework is that physical products rely on economies of scale and standardized production, while services require local, personalized delivery. The 2 attributes were historically hard to combine. AI creates the possibility of giving some services both the replication efficiency of a product and the customization of a service.
Animation is the clearest example. As 张伟伟 recounts it, 1 minute of traditional animation costs about RMB10K, while AI manhua dramas can be automated end-to-end and bring the cost down to a few hundred yuan per minute. The change is not just lower cost; it “turns a service into a product, or enables low-cost service delivery.”
POD products such as wedding designs, personalized music and videos might be produced by 2 or 3 people using a simple workflow. But once virtual content becomes a physical product such as a T-shirt, the economics run into transfer printing, UV printing, digital spray-painting and flexible manufacturing. AI can generate a unique design; that does not mean the supply chain can produce it economically on demand.
5. The Closer Content Supply Gets to Infinite, the More Human Trust May Become the Entry Point
Once 60-point videos are readily available, the market will split toward 2 extremes: “super-giants” on one side, and a “super-long tail” built around taste, aesthetics and individual expression on the other. Most producers below the common waterline will drown, while a few localized capabilities above it will become more visible.
Douyin currently receives roughly 100M video uploads a day. If AI pushes supply not to 10x or 100x but to a scale that is harder to imagine, algorithms may be overwhelmed and no longer able to solve user fatigue or discovery.
Information distribution moved from portals, yellow pages and search to algorithmic recommendation. 庄明浩 expects the next phase may bring a retro return to “editor’s picks.” Users will rely on particular creators, brands or small communities because long-term personality, emotion and trust are primitive signals people can still process amid an ocean of content.
张伟伟’s “AI slop” is the negative case: content appears fluent and confident, but in retrospect says nothing, like prepackaged food with no wok hei. Phrases such as “a chill down your back” and “scared into a cold sweat” are replicated at scale, exposing the industrial sameness produced when algorithms chase the lowest common denominator.
6. Super-Individuals Win by Serving a Few People, Not by Pleasing Everyone
AI’s leverage on individuals and small teams is already visible: projects one person would once have considered impossible are now feasible, and small teams can take on work that previously belonged to mid-sized teams. Value does not need to come from reaching everyone; if a small group “really loves you,” it can support deeper trust and a stronger commercial relationship.
庄明浩 summarizes Pinduoduo’s original story as aggregating fragmented demand and using demand density to push back on prices. What is being aggregated now is not just product orders, but services such as music, images and companionship—though the model depends simultaneously on AI, online channels and supply-chain integration.
7. Healthcare Reminds the Market: Humanities Tasks Tolerate Error, Scientific Tasks Demand Precision
Recalling her mother-and-baby private-domain community, 张伟伟’s ideal was to cover the full pregnancy, infant and child lifecycle. In reality, maintaining constant interaction was intensely labor-heavy; employees who had never had children could easily be exposed when users asked follow-up questions. She had to move high-value users into groups run around her “straight-A student” persona and even tutor children’s homework herself.
AI makes personalized companionship imaginable again, but 庄明浩 stresses that health requires handling chronic conditions, one-off symptoms, medication, diet, hospital data and medical records, among countless other details. A model may complete the first step, but moving from 80 points to a reliable 100 remains a case where “every point is painfully hard.”
庄明浩 separates tasks into those closer to the humanities and those closer to science. A user ordering a custom illustration can tolerate multiple acceptable outcomes; healthcare demands precision, while AI fundamentally produces probabilities and therefore will inevitably be wrong. 张伟伟’s weight-loss coaching, diet photo tracking and AI reminders may find an initial foothold outside medical settings.
8. Build in Public Rewards Whoever Surfaces First
庄明浩 translates the shift from plan-driven to feedback-driven execution in deliberately crude terms: “Even if what you build is shit, get it out before everyone else.” No one knows the correct answer above the waterline, so the person who does even a little more first can receive attention many times larger than the lead itself.
The boost may come from users, capital, media or even competitors trying to copy the answer. It does not fully belong to the current product, but it can help the team attract talent, funding and another round of feedback, then compound its small lead.
“Fake it till you make it” does not guarantee success. Most projects may still be drowned by the rising waterline, and something commissioned 3 months ago may suddenly become obsolete. Staying underwater to build a grand reveal only means losing confidence as the water continues to rise.
9. Greatness Cannot Be Planned; Products Must Start with the Nearest Stepping Stone
One smart-hardware founder conceived a similar product before PLAUD but held back because the technology did not seem ready. PLAUD went to market first and happened to catch the release of a new GPT model and stronger multimodal capabilities; all the external boosts landed on a product that already existed.
Participants describe the environment using an unfamiliar game map: you may know you should head roughly east, but not what terrain lies ahead, making a complete plan meaningless. The more workable approach is to step onto the nearest stepping stone, let users test the product and let feedback reveal the map. The outcome remains uncertain, but “if you’re going to play, you have to show up first.”
10. AI Hardware Consensus Comes from China’s Supply Chain and Software-Payment Reality
Early-stage investors cannot know which product is right, so they increase the weight assigned to the people involved. First-round bets naturally favor big-tech specialists, algorithm engineers and CTOs; soon, however, investors discover that pure technical teams may lack commercialization, organizational and user capabilities.
The US first demonstrated that users would pay directly for software, while Chinese software monetization is considered harder to make work. At the same time, advanced manufacturing, electric vehicles, Huaqiangbei’s supply chain and global execution have been repeatedly validated by DJI, Bambu Lab, Insta360 and Xiaomi’s supply chain. AI and hardware therefore become the natural consensus intersection.
The field is quickly splitting among glasses, plush or plastic toys, companion devices and learning hardware. Shenzhen’s supply chain already has middleware vendors packaging models into locally running, low-power, small-parameter modules. Add 3D printing, and the original idea of “putting a model inside something” may connect with all of these directions.
11. Models Upgrade Monthly; Hardware Must Carry a 6-Month-Plus Physical Cycle
Even simple hardware takes at least 6 months from idea to tooling, supply chain, stocking and live-commerce sales. A more complex device such as a guitar took a team about 18 months. Today’s hot product may have been started in 2024, when DeepSeek 2.1 did not even exist.
Users see model demos and media hype every day, but the product they receive is constrained by the battery, endurance, compute, display, screen and AR architecture. Glasses buyers imagine “Iron Man”; after putting them on, they may feel they have received something “idiotic.” The expectation gap directly pushes up returns.
Return rates for some AI hardware are described as close to those of women’s apparel. Meanwhile, conflict can move component costs, while online acquisition ROI and gross margin are already mathematically pinned down. Suppliers demand larger orders to lower prices, but a brand may sell more and lose more because of returns.
CK’s advice is to accept low early efficiency: the product can be a little worse, orders smaller, costs and return rates higher, and losses tolerable, “but never, ever carry inventory.” By the time a large order finishes production, the product may already be obsolete; inventory risk is far greater than the savings from optimizing unit cost.
12. Real Users Often Have Nothing in Common with a Founder’s Geek Fantasy
Rabbit R1 is essentially a square Android device running a model, and the first generation was criticized mercilessly. But the team did not build massive inventory first, and the product was later mentioned again because of Open Cloud. The experiment may not have been right, but the team continues to explore new forms.
“Little Lobster” and COCO are described as sitting on a status ladder. 庄明浩 compares them with Records of the Three Kingdoms and Romance of the Three Kingdoms: ordinary users prefer things that are accessible, fixed, tangible and human. Plot targeted doctors, lawyers and teachers rather than internet workers because the experience delta may be more obvious to first-time users.
Exoskeletons expose the mismatch even more sharply. A geek founder wants to become Iron Man, while the actual buyer may be someone purchasing the product for an elderly parent. The company must choose between its identity as a “cool company” and the hard demand of elder care; the market will not change to accommodate the founder’s aesthetic.
13. As Traffic Playbooks Depreciate, Consumer Brands Return to Product Experience
张伟伟 says a traffic team known as the “Xiamen gang” once ran more than 30,000 creative variants at the same time, relying on young employees’ relentless effort, platform instincts and Douyin algorithm expertise. Today, AI generates nearly all of the performance creatives, and the old combat culture and playbook are suddenly no longer scarce.
The team has therefore gone back to fix the product experience. If NPS and repurchase do not improve, the business still cannot turn. AI has first flattened the efficiency of traffic production; it has not answered the old question of why a product deserves to be kept.
14. China’s To B Opportunity Is the Dirty Work Between Models and Enterprises
The mainstream US view is that AI agents and coding capabilities will let individuals and companies build more genuinely customized software themselves. Software engineering, legal work and financial analysis were also core targets for SaaS companies, so the logic holds in the US.
China’s enterprise services and the software layer of To B SaaS never truly took off, producing 2 opposing views: AI may make an already weak market even harder, or China may have less historical baggage and be able to rebuild services around local payment and delivery habits.
Over the past 1–2 quarters, leading AI SaaS companies listed in Hong Kong and the US have all been talking about “AI-driven revenue growth,” but the definition remains vague. Many knowledge bases, digital employees, contextual memory and chat-box products look like crude wrappers, yet early implementation cases have generated market feedback far beyond their product maturity.
When large-model companies work with traditional enterprises such as SAIC and Shanghai Chemical, the real obstacle is not demo capability but the customer’s data condition, level of digitization and other prerequisites for using models directly. People still have to retrofit ERP systems, connect data and implement on site. Palantir is seen as a “new-era arms dealer” precisely because it fills this interface between governments and technology vendors.
15. Giants Pull Away, but the Long Tail Still Grows around Special Needs
AI lets individuals do the work of small teams, but it also allows very large teams to widen the gap. At least for now, the phase-level consensus is that “big companies will get bigger”: existing battlefields, platforms and data amplify the difference, and more small teams cannot close it linearly.
After ByteDance released Seedance 2.0 around Chinese New Year, 庄明浩 says it told the market that the company had found the formula “good data can produce a good model.” Data is not a sufficient condition, at most one necessary condition. Video is also a core battlefield on which ByteDance will not waver strategically; startups pursuing the “crown jewels” must absorb that resource gap head-on.
When Tencent applies AI to gaming, it can take on work that many other companies “would not even dare to imagine.” This is no longer a problem solvable by adding headcount, increasing volume or relying on traditional craft. Companies trying to reach the same depth with only a general-purpose model will move more slowly.
Large companies still cannot cover every need. On, Hoka, Salomon, Arc’teryx and lululemon all show that brands matching new settings, audiences and social emotions can emerge outside Nike and Adidas. The realistic path for a small company is not to become another largest tree, but to fill the long tail as “moss” around the giants.
张伟伟 sees the same long-tail value in To B. Many domestic SaaS companies do not understand the underlying business, while genuine 80-, 85- and 90-point capabilities sit in the heads of industry experts. If that experience can be AI-enabled and productized, the expensive intelligence once priced by McKinsey at RMB50K an hour may finally be made broadly accessible through Deep Research.