Summary

AI’s clearest returns today are not in the “universal assistant,” but in driving the marginal cost of code, sales, and customer service low enough. 任鑫’s incubator now has AI writing 100% of its backend code, cutting process redesign from two months in the GitHub Copilot era to one or two weeks with Cursor. 曲凯’s auto-sales funnel runs from identifying intent in Xiaohongshu comments, to nurturing leads in private chats, to handing only high-intent buyers to humans for closing. “One thing that clearly works today” is still improving internal sales and customer service productivity.
AI is pushing products that once could not support a 3-5-person team into a “strange-point zone” where 1-2 people can operate them sustainably. Using AI to write code, test, and review voice interactions, 赵纯想 finally brought to life a “Stranger Alarm” idea shelved for 7 years; 任鑫 has likewise seen teams move from repeatedly debating marginal projects to saying, “Fine, let’s just build one.” The real investment shift is not that everyone can build a large company, but that more small needs now have viable unit economics for the first time.
As code, testing, operations, and deployment become nearly worthless, the solo developer’s final moat is creativity, design, and the product feel that leaves users “genuinely delighted.” 赵纯想 believes AI can fill gaps in engineering knowledge, but it “doesn’t understand what it means to move people, or what craftsmanship means.” His advice is not to grind harder on general-purpose technology, but to study design extensively and invoke a model only when “this model can do it and nobody else can.” Human taste and product sensibility are moving from nice-to-haves to core factors of production.
Job losses are already concentrated in art, content pipelines, and the “translation-and-errands” middle layer, but it is still too early to call this mass unemployment. 庄明浩 sees around 80% of the production process for 2D scene art and online murder-mystery games becoming AI-ready, while a friend’s company has already eliminated every sales-assistant role. At the same time, low-cost legal services, visual assistance for the blind, and pronunciation training for deaf people are opening markets that previously could not be served. 赵纯想’s macro warning is that the AI industry remains “a money-losing business” overall and has not yet reached the point where industrial restructuring can be declared complete.
AI companionship has exposed a huge gap between model capability, user perception, and ROI; for now, the most effective product is often “AI acquisition, human delivery.” At one end is EVE, a 3D AI girlfriend built by a dating-game team; at the other are Southeast Asian teams using “AI companionship” purely as a user-acquisition slogan while low-cost human guilds handle the chats. Some products use AI to replace the “pretty face” while keeping humans for the “interesting soul.” 曲凯 offers an even sharper counterexample: “The good AI products all say they aren’t AI,” because the AI label itself can signal cheapness and a lack of willingness to pay.
This AI cycle has delayed the retreat of China’s venture market while forcing the old model to answer its endgame questions. 庄明浩 sees the cycle from being ignored, to attracting everyone, to being rapidly abandoned compressing from 2-3 years to quarters; AI reached consensus and deployed capital within 3 months, like “giving everyone one last chance to go crazy.” Capital and institutions have shrunk materially, but for credible projects with solid people and teams, an explorable direction, and disciplined fundraising, the odds of a deal may be lower than at the peak of the frenzy when facing the few funds still acting seriously—though actually closing may not be that difficult.
The short-term opportunity may be multimodal “wrappers,” but the bigger long-term opportunity is turning people from production tools back into ends in themselves. 任鑫 is currently focused on combining models for faces, movement, 3D, dance, and voice into highly entertaining products, while admitting they may be “a flash in the pan” rather than an investment opportunity. Over the longer term, he is betting on health, mind-body-spirit, self-understanding, and becoming oneself. 庄明浩 believes AI will certainly widen the wealth gap, but may also bring “emotional equality,” allowing everyone’s emotions to be seen, answered, and cared for.
The key gap between Chinese and US AI startups is not just models or talent, but whether the To B ecosystem can provide a floor for entrepreneurship and an exit loop. 曲凯 believes short Silicon Valley study tours often fall into a fixed hospitality industry; unless visitors stay and enter the core circles, simply “going around once” has limited value. What matters is that specialized US To B companies can be acquired at a premium by larger companies, allowing founders to start again and reinvest, whereas domestic acquisitions often amount to “shut down your company and we’ll give you a salary.” China spent years betting on To C while the US worked through cloud, SaaS, and enterprise services; now that the paths are converging in AI, China still has to rebuild its To B capabilities.

Deep dive

1. AI’s best role in medicine is still a second opinion

任鑫 sent a family member’s serious illness case simultaneously to the attending physician, ChatGPT, and Claude. At the follow-up visit, faced with dozens of screenshots of old and new reports, he had ChatGPT consolidate them into a CSV, identify 6 key indicators, and mark which had improved or deteriorated. AI was not a chat toy here; it turned medical records into a comparable time series.
After several experiences involving a sprained ankle, his child’s hand-foot-and-mouth disease, and his mother’s liver disease, 任鑫 felt the models’ judgments were broadly indistinguishable from those of a “department-director-level doctor.” But his boundary was equally clear: “Of course I don’t dare trust it completely.” In practice, he always asked Huashan Hospital to double-check rather than replacing the doctor.
The main difference between the models and Chinese doctors is that the models often imagine a more serious situation. 任鑫 suspects this may reflect alignment or overseas medical workflows: devices routinely used in China for efficiency can be described by models as signs of a major problem, while doctors around him say, “It’s fine to proceed; there’s no issue.” His conclusion is not that the models are infallible, but that they are “just overly cautious, not harmful.”

2. The same generation of AI users has split into dependents, explorers, and skeptics

赵纯想 says AI has become like “electricity” and “energy”: code, architecture, and automated testing all depend on it. The more direct side effect is that coding has become completely addictive. He feels like a conductor constantly assigning work to multiple AIs, with another AI operating the database in the background. “You just can’t stop, like an infinite game,” so the lifestyle change is almost the disappearance of personal life.
庄明浩’s child turned an AI plush toy into an experiment in the boundaries of language. When the robot could not correctly play Red Light, Green Light, the child repeatedly corrected its counts and pauses, only to run into the limits of its lack of vision and inability to tell whether it had moved. What 庄明浩 finds most striking is that nobody taught the child about models or prompts; the child had instinctively learned to “hack the model with language.”
曲凯 was the most important counterexample in the room. Despite having seen a large number of AI companies and products, he barely uses AI in daily life and believes the industry “is still at a very early stage.” Around two-thirds of the audience had interacted with AI in the previous 48 hours, but his own life had changed mainly in two ways: he had met more listed-company chairmen anxious about AI, and he had begun warning Xiaohongshu commenters not to be fooled by AI-generated fake people.
The window for detecting fakes is also closing quickly. 曲凯 says fake-human accounts have improved dramatically from a year ago: “In another 6 months, I won’t be able to criticize them anymore.” 庄明浩 adds that images in the Stable Diffusion era were “obviously fake at a glance,” while Flux refinements have made them extremely difficult to distinguish.

3. The bottleneck in AI coding is no longer just the model, but whether organizations will change themselves

任鑫’s incubator began its AI transformation roughly 6 months ago. Today, “not a single line of code is written by us” on the backend; only the frontend is outsourced because it requires extensive communication and makes it difficult to judge “whether it actually looks good.” The backend can support multiple projects in parallel with almost no increase in headcount.
The process redesign took 2 months during the GitHub Copilot phase. The technical lead now estimates that Cursor would reduce it to 1-2 weeks. The key change is not one-off code generation, but rewriting task decomposition, acceptance, rework, and collaboration around AI.
The technical lead’s principle is extreme: “If AI writes that one plus one equals three, you should assume you are wrong, not AI.” He does not mean models do not make factual mistakes. The point is that if a team declares a tool unusable first, it will never investigate whether the task was underspecified, the architecture was wrong, or the process itself made no sense.
任鑫 put a price on this patience: “When he charges you RMB20,000 a month, you can be picky; when he charges you RMB20 a month, if the work is poor, you should blame yourself.” The core skill in managing AI is therefore shifting from knowing how to write prompts to knowing how to domesticate cheap capability inside a workflow.

4. Content production gets structured first; reasoning models turn “talent” into something that can be sampled at scale

A television director 赵纯想 knows used to hire large numbers of interns to break down shows. The basic task was copying plots; the more advanced task was mapping character relationships; the truly time-consuming work was extracting reusable structures. Later, the team built a workflow on Coze and optimized the prompts, allowing dozens of classic TV series to be broken into spreadsheets or mind maps in a single afternoon.
To an ordinary viewer, these are just summaries. To a director, the macro structure stripped of dialogue and specific characters can become the skeleton of the next show. He then revises the structure and hands it to intern screenwriters and copywriters to expand, compressing the most labor-intensive part of pre-production into selection and editing.
赵纯想 believes the “strawberry model,” o1, pushed this process another step: first specify an unexpected ending, then force the model to answer “why,” producing 100 logical chains through “confidently making things up.” One might be both feasible and beyond the director’s expectations. Humans handle selection and refinement; the model mass-produces potential talent.
A related web series has already been made, but remains under filing review and has not yet launched. It shows that the creative workflow has been tested; it does not prove that audience acceptance or commercial results have been validated.

5. AI companionship is splitting into perfect virtual humans and human services wearing an AI costume

庄明浩 observes that after GPT launched, the social industry effectively assumed responsibility for To C experimentation. AI companionship, dating, and chat projects surged, and “add some AI” became the sector’s slogan. 2 years later, the result is not a unified product paradigm but a split into two extremes.
At one end is EVE, a 3D AI girlfriend built by a dating-game team. The team’s accumulated expertise in characters, storylines, and 3D presentation makes the demo look nearly perfect. A large number of other projects, however, remain stuck between emotional value, engineering stability, operations, and cost, without a clear end state.
An entrepreneur from YY familiar with Southeast Asian guild resources simply treated “AI companionship” as a slogan and user-acquisition hook, while humans continued chatting behind the scenes. After testing solutions on the market, his conclusion was that “the kind of experience that maxes out emotional value can only be delivered by humans,” and local human supply was more cost-effective because the model economics did not work.
Another team split apart the “pretty face” and the “interesting soul”: AI drives the front-end video avatar while a human speaks from behind the scenes. 曲凯 offers the reverse signal—some strong-performing products deliberately avoid saying they are AI, because users hear “AI,” think cheap, and refuse to pay. Technical identity and marketing identity are not the same.

6. The sales funnel is the first commercial interface this AI cycle has truly made work

A Southeast Asian chat-companion business may receive around 10,000 users a day. Previously, humans had to chat with every one of them until they paid. The operating experience is that the first 10 or so rounds are enough to judge ability to pay. AI screens first, and once a lead is worth pursuing, immediately hands the conversation to a human by phone or chat without restarting from zero.
曲凯 saw a more complete auto-sales funnel on Xiaohongshu. AI first generates soft-sell content, then identifies weak intent in comments such as, “What do you think of this color?” Another bot leaves a message inviting the user to DM and offering a discount; AI continues nurturing the lead until purchase intent is strong enough for a human to close the high-ticket sale.
This also cools 曲凯’s enthusiasm for many “AI-native application” stories. What ultimately works is often not an independent new species, but better internal sales or customer service. The value comes from reducing human waste on low-intent users while keeping humans in the trust, negotiation, and closing stages.

7. The first jobs to be compressed are all in the translatable, template-driven middle layer

The first clearly affected occupation 庄明浩 sees is art. Stable Diffusion entered production pipelines roughly 1 year before ChatGPT, especially for 2D scenes without characters or fine-grained detail. Substitution has already occurred at game and product-design companies.
The shift in online murder-mystery games is more thorough. From logos, frameworks, characters, and plots to monetization hooks and deliverable packages in the editor, around 80% of the process can now be handed to AI. A content-heavy operating business that once involved perhaps 10 occupations is becoming a business with a “technology-oriented” core.
A friend’s company has eliminated every sales-assistant position. Sales assistants used to translate customer needs into internal materials and assemble them into proposals. But the role neither connected directly with customers nor knew the product best; it was essentially “translation and errands.” 任鑫 therefore believes all similar translation-oriented jobs will be hit early.

8. Productivity gains eliminate old tasks while turning uneconomic needs into markets

任鑫 does not equate higher efficiency directly with fewer total jobs, because “there is always more work than people can do.” Once the cost of building a product becomes low enough, teams stop holding long meetings for every launch and say, “Fine, let’s just build one.” Experiments that were previously optional begin entering production in batches.
Legal services are his clearest non-consumer-market example. Nobody would pay a lawyer $1,800 or even $3,000 to handle a $180 speeding ticket or gym deposit. Companies now specialize in these small disputes and take a share of the amount recovered. They serve people lawyers previously did not consider worth serving.
The same expansion logic appears in Be My Eyes and projects helping deaf people who can speak train their pronunciation. Human coaches used to be too expensive, so the demand did not appear in commercial statistics. Once costs fall, a market for personalized guidance can emerge.
赵纯想 retains a macro objection: capital invested in the AI industry remains wildly disproportionate to its aggregate returns. It is still “a very money-losing business”; only once it approaches break-even does it make sense to discuss stable industrial replacement and widespread unemployment. Even if programmers are affected, they are also the people most likely to learn AI—if they can learn to code, natural-language tools should be easier to master.

9. AI is China’s venture-market lifeline—and the old narrative’s final acceleration

任鑫 says his friends were already unemployed, and AI instead rescued many of them. In previous years they had planned to look at new energy and chips; after AI appeared, China’s venture market once again had a direction worth examining. 庄明浩 offers a colder explanation for this lifeline: venture capital has to be permanently long, otherwise the job cannot function institutionally.
Beginning around 2015-2016, the cycle from nobody knowing a theme, to a few people watching it, to everyone rushing in, to nobody looking at it again compressed from 2-3 years into quarters. 庄明浩 believes AI is the endgame of this narrative: consensus formed in roughly 3 months, and money was deployed in an extremely short window.
His signature judgment is that AI is “giving everyone one last chance to go crazy,” pushing the old model to an extreme and forcing the industry to ask whether it can continue. Dividends, buybacks, RMB and dollar-fund strategies, LP psychology, and listed-company participation are all, in essence, searches for a new answer. Without AI, this self-examination might not have become so urgent.

10. There is less money and fewer funds, but deals for good teams have not necessarily disappeared

庄明浩 mentions that the Financial Times once cited dollar-institution financing data from PitchBook alongside data from IT Juzi, which showed only 260 Chinese startups in the first half of 2024. He considers that number “absolutely impossible the moment you hear it,” but says it also shows that public data during a downturn may not capture the full market.
Even if active institutions have shrunk from a very large number to a metaphorical 20, deals have not disappeared as long as those 20 are still working seriously. For To C AI projects with credible people, an explorable direction, a complete team, and disciplined fundraising, 庄明浩 thinks the odds of a deal may be lower than during the peak frenzy—but closing may not be so difficult under those conditions.
任鑫 explains from his own founder experience why raising less money is not necessarily worse. Both times he raised at the tail end of a bubble, and much of the capital went not to solving user problems but to subsidizing order volume to catch competitors and avoid falling behind. Fundraising is harder and amounts are smaller now, but competitors cannot raise much either, so competitive spending can fall together.

11. AI has not chosen 任鑫’s investments for him, but it has begun questioning him and manufacturing perspectives

任鑫 describes his main use of AI when facing genuinely difficult problems. He does not ask directly for an answer; he first lays out the situation, sticking points, and conflicts in full, then asks AI to “ask me only 1 question at a time.” After each answer, the model gives feedback and asks the next question. After dozens of rounds, he can usually think the problem through himself.
His second use is to have AI play different roles—Ultraman Seven, Steve Jobs, Lei Jun—and answer what each would see, what emotions they would feel, why they would feel them, and what they would ultimately do. The value is not imitation of celebrity one-liners, but the reminder that his current view and emotions are just one perspective, not the only response dictated by the situation.
The incubator’s positioning illustrates the point. A Y Combinator-style incubator, a venture studio, and Bending Spoons’ model of acquiring and rebuilding Evernote have completely different goals and organizational logic. In reality, it is hard to find a friend who understands these paths at any given moment; an AI role library can at least help him see multiple facets of a problem faster.
Koji mentions the “AI will” as a consumer application of the same mechanism. Early versions could only ask users 10 fixed questions; AI can now dig deeper like a journalist, change angles when the user cannot answer, and track the faintest emotional signals. More than 10,000 people have tried it. It has no legal force, and its purpose is not to pressure young people into writing wills, but to use death as an extreme state for identifying what truly matters today.

12. Multimodal wrappers are a short-term window; treating people as ends is 任鑫’s long-term bet

任鑫 and his partner at AI Alchemy once publicly looked down on “wrappers,” and later regretted it repeatedly: “Wrappers are so useful. Wrappers are an opportunity.” The early window for wrapping pure large language models may already have passed, but the misjudgment made him take rapid productization more seriously.
He believes multimodal wrappers still offer short-term opportunities. New papers and small models for face recognition, facial driving, 3D modeling, dance movement, and voice provide a rich weapons library; lightly assembling them can produce highly entertaining products. But he is careful with his language: they may merely be “a flash in the pan,” it is unclear how to make money, and they may not constitute investment opportunities.
Over the longer term, 任鑫 believes modern people have been alienated into “tool-person status,” introducing themselves through companies and job titles and treating productivity as their value. But in a productivity contest with AI, humans can “at most eke out another 10 years.” His conclusion is not pessimism, but that people should become ends again: health, mind-body-spirit, self-understanding, and becoming oneself may matter more than every efficiency tool built to hammer nails.

13. AI may widen the wealth gap while industrializing “emotional equality”

庄明浩 believes AI “will definitely widen the wealth gap,” while wealth distribution must be handled through policy, government, and public services. The other form of equality AI may deliver is allowing emotions to be seen, understood, answered, and cared for. In the past, that depended on being born into a good family, meeting a good partner, or being able to afford the best therapy.
庄明浩 relays 陈天桥’s view that mental illness is the condition AI can most reliably identify today through language and facial expressions; other diseases often still require instruments. After identification, AI may also provide targeted interventions. 陈天桥 hopes to collect more relevant data and apply it in the real world, while 庄明浩 sees this as China’s “spark” for using AI to improve emotional health.
赵纯想 defends “industrialized emotional support.” Industrial products do not necessarily corrupt humanity; they first make benefits broadly accessible. His wealth gap with Bill Gates is enormous, yet he can still drink the same Coca-Cola and use the same iPhone. If everything were handmade, the gap in experience would be even greater.
Koji extends the analogy with IKEA’s RMB29.9 LACK side table. Capital and industrial production can bring beds, dining tables, and TV cabinets once reserved for the rich to everyone. AI emotional support may sound like a cheap industrial product, but precisely because of that, it may first benefit those with the fewest resources.

14. The biggest risk of a short Silicon Valley trip is mistaking a fixed hospitality route for the core ecosystem

曲凯 says the California sun is excellent and the time difference with China is completely inverted: nobody contacts him during the day, and by the time work begins in China at night, he is already asleep. It is therefore ideal for vacation and resetting. But if the trip is treated as industry research, he sees “not that much need.”
He has seen visitors marvel at Silicon Valley’s talent density after meeting several Apple AI employees in a row. He considers that no different in substance from an Indian visiting China, meeting Baidu and ByteDance employees consecutively, and taking photos outside TikTok. “Stupid people are stupid in different ways; capable people ultimately look similar.” In his subjective view, Chinese practitioners may even have a higher average level because they train more and compete harder.
The more practical issue is that a group of people already specializes in hosting Chinese visitors. 曲凯 says the people they met this time were basically the same group, while 任鑫 adds that Japan has a similar industry chain. Either stay long enough to genuinely enter the core circles, or admit that you are simply “going around once” on a mature hospitality route. Do not mistake visit density for information gain.

15. The To B ecosystem sets the floor for entrepreneurship and explains the resilience gap between Chinese and US venture markets

曲凯 believes To B demand is generally easier to find than To C demand. Enterprise pain points can be validated one by one, while consumer PMF is harder and network effects often push markets toward winner-take-all outcomes. To B allows many companies to succeed in different verticals; not every company has to hit a super-platform opportunity.
A US founder can operate a niche To B company for several years, sell it to a listed or large company, then use the capital to start again or invest. That loop supports a floor under entrepreneurship and the venture market. The typical offer from a domestic Big Tech company to a Chinese AI founder may instead be: “Shut down your old company and we’ll give you a salary,” with almost no acquisition premium.
庄明浩 takes the timeline back to 2010. China then sprinted sideways into To C across TMT, while the US moved through cloud, SaaS, and enterprise services. The paths are converging again at the AI junction. Capabilities China did not build in the past may now have to be rebuilt, but the room offered no universal answer on which To B scenarios fit China or how capital exits. 曲凯’s closing line remains: “Everything is cyclical.”

16. The ultimate asset of a one-person company is product sensibility that cannot be generated at scale

赵纯想 believes AI has largely erased differences in labor, intelligence, engineering experience, and efficiency. The one thing it cannot eliminate is human creativity. He advises one- or two-person companies to spend substantial time on design: AI can build a page, but it does not know what will leave users “genuinely delighted,” nor does it understand being moved, craftsmanship, or the traces of a developer taking the work seriously.
His next product, Action Park, follows the principle of choosing tasks “only this model can do.” Users take one photo before cleaning and another afterward; o1 uses timestamps and visual changes to determine whether the task was completed, where items were moved, and how to organize next time, then provides praise and points. A sleep-early challenge combines device data to determine whether the user is still playing with their phone and assess a “deception index.”
“Stranger Alarm” was an idea he had 7 years ago but could never operate within the cost of a 3-5-person team. AI can now handle voice review, testing, databases, deployment, and operations, pushing the product into a “1-2 people can make money, but hiring 2 people immediately makes it lose money” strange-point zone.
When Meituan launched a highly similar food-logging feature, 赵纯想 responded: “That is the privilege of the winner: copy me—what does it have to do with you?” Meituan naturally has delivery data and human-entered calorie data, so it should be more accurate. What “未知数” can retain as an unknown is the soft moat of design and feel. Cal AI generates more than $400K in monthly revenue overseas, while “未知数” may make only a few dozen times less. But 赵纯想 admits he enjoys the process of “building one, selling one”—more like a designer and craftsman than a CEO skilled at long-term operations.

Yet Another Value's 2025 Idea of the Year: Full House Resorts $FLL Kering: It’s Gucci - [Business Breakdowns, EP.199]