Summary

China’s foundation-model wave converged from dispersion to focus: 2023 was the year of capital mania, with “catching up to OpenAI” as the defining theme; Llama 3’s release in 2024 eliminated a batch of startups and companies claiming to build foundation models; DeepSeek’s breakout in 2025 convinced the market that open source had a future, followed by Manus’s explosion and the arrival of the Agent era. 卫诗婕 remains cautious: broader consensus does not mean the confusion has cleared. Some in the industry still miss the period before the frenzy, when things felt “more grounded, more conducive to actually getting things done, and less likely to distort the way people moved.”
Koji sees two defining traits in this wave of AI startups: monetization from day one and globalization from day one. Founders are thinking about commercialization almost immediately, often operating domestically with one hand and overseas with the other, or targeting overseas markets from day one because B2B and B2C willingness to pay, as well as acceptance of AI, are stronger abroad. As an angel investor in Manus’s predecessor and its current advisor, Koji sees Manus as having moved AI beyond one-question-one-answer exchanges into long-chain, complex tasks. In an internal test, a design Agent built out an entire visual system for a co-branded campaign, from the logo to delivery-worker caps and highway billboards.
One key difference between Big Tech and startups is whether they can survive the stair-step plateaus in model capability. 卫诗婕 believes DeepSeek’s willingness to keep investing without rushing to monetize matters: the real decision point is whether a company can continue spending through a period when no progress is visible. China’s business culture favors saturation attacks—“Manus took off, and ByteDance had four teams working on it internally at the same time”—but giants often lack concentrated attention. A Big Tech company that has built audio products may still fail to beat 小宇宙, leaving room for startups.
Vertical AI companies that raise little money may still be quietly profitable. Koji believes the image-recognition app 形色 should have very high revenue and must already be above RMB100M. 刘飞 adds that it has raised very little and may not need a large model at all, having built an excellent user experience on earlier vision models. Podwise, an AI podcast-summary service with a subscription model, has reached breakeven; Hangzhou-based Tripo3D is a global leader in turning photos into printable 3D models. A strength becomes a moat only after years of accumulation.
The core personal-strategy judgment is: “AI will not bring intellectual equality; it will only bring information equality.” 卫诗婕 argues that people should use AI to extend their strengths and shore up their weaknesses, while avoiding becoming a “screw” that depends on AI to do its work. Once you only let AI act on your behalf, you are effectively letting it replace you. 刘飞 sees helping people write papers, generating public-account articles, and teaching courses on how to make money with AI as short-lived information-asymmetry businesses; eventually, people with fewer scruples may make the most money. A more durable path starts from one’s own skills and needs. The real moat is the feel and judgment built by doing the dirty, difficult, exhausting work.
People from the previous era can still get a seat at the table in this one if they actively embrace the wave. 谢阳, the 1996-born founder of Fellou, had previously built a B2B project into one of China’s leading companies in its category. After setting aside two periods to read papers systematically and think, he produced an early version of an AI Agent product. Koji uses Dyson’s leap from construction-site carts to hair dryers to show that creativity can transfer across domains. Open source and tools are lowering multiple barriers across technology, operations, and talent.
A method for taking action matters more than an attitude: name the project, make a 5% change, and find peers and physical space. Koji says ChatGPT did not give him the “delight” he felt when he first used an iPhone, but rather “shock,” giving him his first genuine belief that AI would reshape his future. Many founders may not have seen a future invisible to everyone else; they may simply have happened to occupy the right position, atmosphere, and physical space. On reinforcement learning, Koji summarizes it as “learning by doing,” while 卫诗婕 quotes Alibaba Cloud CTO 周靖人: “Everything is RL”—action means building an environment, receiving feedback, and continuously adjusting.

Deep dive

1. Four Peers on One Stage: Action Itself Is a Way to Fight Anxiety

Koji’s origin story begins with a month of unemployment, sprawled on a large white sofa in a classmate’s living room in Foster City, “surrounded on three sides by the lake and all floor-to-ceiling windows.” He touched his phone and found that “no one was looking for me,” leaving him with the feeling that “the world had abandoned me.” Then 张伟 sent a message: “Come back. Let’s start a company together.” Koji agreed without asking what they would do. “Lying there was often when I felt the most anxious.”
Koji says both optimists and pessimists are common after the arrival of AI, and “there is nothing wrong with either attitude in itself.” The problem is having an attitude without taking action—remaining in the realm of speculation. “The key is how you act. That is what we are here to discuss today.”
程曼祺 separates “being proactive” into two layers: visible, immediate activity, and positive contributions to an industry or society that may take time to emerge.

2. The Prelude: BAAI and Scientists Moving First on Technology

卫诗婕’s key story is that, after GPT-3 was released, a small group of Chinese researchers had already seen large language models complete the zero-to-one phase. They believed the remaining work was a one-to-one-hundred problem. Like OpenAI concentrating resources on large models, the Beijing Academy of Artificial Intelligence put everything behind the direction—three years later than OpenAI.
The most encouraging detail was that 面壁智能 CTO 刘知远 had not even been promoted to associate professor at the time. Yet within a research system organized around seniority, he was able to marshal computing resources worth tens of millions of RMB for frontier research. 王小川 has also said that his earliest introduction to large models came through events at BAAI.
The melancholy in hindsight is that this non-utilitarian research environment, devoted purely to exploration and cultivating technological innovation, “was actually closer to the ideal.” After ChatGPT appeared and the capital frenzy began, BAAI was “almost emptied out,” its people scattered like sparks.

3. 2023–2025: From Capital Mania to Convergence and Focus

The timeline starts with 2023’s “large-model capital mania.” Scientists with experience training large models or with a theoretical foundation in related technologies could raise money, while few had time to think about what AI capabilities might eventually be used for. Competition turned white-hot in 2024. After Llama 3 launched, a batch of founders and companies in China’s large-model startup scene were eliminated; once open-source models reached a certain capability threshold, teams claiming to build foundation models were automatically pushed out.
DeepSeek broke out at the start of 2025. 卫诗婕 sees its greatest significance as making people “start to believe in a future for open source,” which could greatly accelerate the AI industry. Manus then went viral, ushering in the Agent era, and the market began looking for a Killer App this year. The threshold for entering the game is also falling: in 2023, founders needed startup resources or credibility; from the second half of 2024, experience from the mobile-internet era could be transferred to the application and framework layers.
Her caveat is that a small consensus expanding into a broad consensus “does not necessarily mean moving from confusion to clarity, because the confusion is still there.”

4. The National Dimension: Modu Space and the Two U.S.–China Powers

Shanghai’s Modu Space is less than 500 meters from where the recording took place. On April 29, CCTV’s Xinwen Lianbo said that “artificial intelligence is a cause for young people.” A national leader visited Modu Space and spoke with founders and scientists. The Caohejing area is home to more than 500 AI-related companies, with the Shanghai AI Laboratory nearby.
卫诗婕’s national-level assessment is that “the most proactive actors in the world right now may be China and the United States.” Europe has some small companies and open-source labs, while Japan also wants to seize the opportunity.

5. The Debate Over Belief: Koji’s Startup History and AI’s Shock

Koji observes that former colleagues from ByteDance and Alibaba were “all pretty depressed” two years ago, saving money with plans to return to their hometowns. Over the past two years, they have moved back to Beijing, Shanghai, Guangzhou, and Hangzhou to see whether they can build something. That confidence reminds him of the old belief in LBS, which eventually became the delivery, ride-hailing, and review platforms that are now basic infrastructure. In Web3, by contrast, many people were simply “confident that it could make money.”
Looking back at his 3 startups, Koji says he did not feel he was betting on the mobile internet when he started 街旁 in 2010. When he started 新世相, he initially thought, “How can a public account become a startup?”—until the first ad was sold to Chanel for RMB200,000. 糖岛 and 猫肚皮枕 were not exercises in believing in new consumption either: “There was no belief, and I was not deliberately chasing the wave.”
This time is different. “My first iPhone gave me a sense of delight; ChatGPT gave me a sense of shock.” For the first time, he developed a belief that AI would bring earth-shaking changes to his personal life and career, “but I have no idea how that change will happen.”

6. Peers and Physical Space: Why Huaqing Jiayuan Was a Cradle

Koji’s subconscious motive for setting up the Shanghai AI Hacker House was to recreate the atmosphere of Huaqing Jiayuan in Wudaokou. When he interned with 王兴 through 饭否 and 海内, he was “non-mainstream”: his classmates were preparing to join Microsoft Research Asia or apply to CMU, while he found a strong sense of recognition there. 宿华, 陈一笑, 张一鸣, and 王兴 were there as well.
His argument is that many founders did not truly see a future that no one else could see. “They simply happened to be in a certain position, a certain atmosphere, a certain physical space.” Silicon Valley remains Silicon Valley because of positive feedback from peers, not just colleagues; online communication cannot replace an offline space of this kind.
The space has existed for only 3-4 months and has not yet produced a big story. But 杨永智—the founder of Dolphin Browser, an angel investor in Li Auto, and the low-profile investor whose fund invested in ByteDance’s first round—has already spent many hours there talking with unfamiliar young people.

7. The Posture of Action: Naming Things and Making a 5% Change

Koji’s small trick is simple: “When you are about to do something, give the project a name. It comes alive.” He initially thought organizing a salon of around 200 people was exhausting. Later, he named gatherings of roughly 30 people “Sit on the Floor” (“席地而坐”). Sitting upright makes people act out a persona; sitting on the floor makes it easier to relax and speak honestly.
The idea behind it is 李松蔚’s “5% change”: ask someone to do just 5% more than their existing habit, and the request is more likely to be accepted.
For ordinary people trying to get closer to proactive builders, 刘飞 recommends building relationships through altruism: first study your own specific strengths and value, identify what information, feedback, or insight you can offer the other person, and then gradually enter that community.

8. How Ordinary People Can Make Money With AI: Do Not Sell Information Asymmetry

刘飞 warns that writing papers for others, generating public-account articles, and teaching courses on how to make money with AI are all short-term businesses built on information asymmetry. “If it is that easy to make money, it means the barrier is very low.” In the end, people with the fewest scruples may make the most money, so there is no need to pile into this direction.
The durable path is not to discard one’s existing skills. A content creator can use AI to lower costs, then open another account or increase posting frequency and quality. A programmer or developer should start with a small use case or narrow direction. 刘飞 cites 花生, someone with no coding background who left a Big Tech company, built the 小猫补光灯 app independently, and rose to No. 1 in the paid rankings. There is currently no highly certain money-making formula that works for everyone.
卫诗婕 adds examples of her own: people she knows who built an AI podcast-translation app or used AI to take on design work all started from personal needs, combining existing capabilities with AI.

9. The Strength–Weakness Framework: Experts Must First Confront Their Own Arrogance

卫诗婕 urged a former colleague who wanted to abandon business development and become an AI content creator not to do it. “AI’s significance lies in strengthening everyone’s strengths and filling everyone’s weaknesses.” Giving up accumulated experience to chase a hot trend may run against AI’s actual value.
She also admits that when a friend told her he was building “an AI that replaces reporters in writing articles,” her “first reaction was intense aversion,” driven by “the resistance and arrogance of an expert.” But “you have to fight that emotion.” Continuous use of AI helps one feel the boundary of its capabilities and the speed of its improvement, and anticipate what it might be able to write in 2-3 years or even 5 years. AI can, in turn, force people to think about what their real strengths are.
Google DeepMind scientist 卢一峰 gave her a useful frame: treat AI as “a magical tool that has fallen from the sky,” and ask what you can use it to do in your life today to make life better.

10. Humans and AI: The Horse Metaphor Is Still Too Human-Centric

The familiar metaphor is that AI and humans are like humans and horses: you do not need to beat a horse in a race, but you can ride one. 卫诗婕 thinks this remains too human-centric. Looking further out, it may be a fusion of silicon-based and carbon-based civilizations, and “it may not be human-centered.”
Her alternative metaphor is that AI is a beast or a new life-form. Humanity’s way of coexisting with it may be to “designate a nature preserve for it”: it grows, humans grow, and together they build a biological civilization.
The practical anchor is that AI was replacing human customer-service agents and lower-end labor long before generative AI. Yet when people get angry at AI, a human customer-service agent is still needed to calm them down. As for “PR cannot be replaced, because someone has to take the blame,” that was a saying she relayed, not a standalone factual claim she was making here.

11. What Happened to Go After AlphaGo: The Answer for AI in Education

程曼祺 observes that the Go industry survived after AI beat humans and was instead reshaped by AI in its entirety. Many sports now use AI to support decision-making and training, and “human teamwork once defeated AI.”
She cites an old 《人物》 profile of Go player 芮乃伟: the vitality of Go “has nothing to do with winning or losing; it lies in the experience of the process.” Human experience cannot be replaced by AI. Perhaps one day, “AI will become the teacher that teaches us to beat it, or to become better versions of ourselves.”
卫诗婕 adds a concrete direction for what happens to teachers: constructing long chains of thought may still require industry experts to write out each step. Large companies may also invest extra resources so that programmers can focus on writing high-quality code. Teachers can help build high-quality datasets for large models.

12. The Agent Aha Moment: From One Question and One Answer to Complex Tasks

As an angel investor in Manus’s predecessor and its current advisor, Koji sees Manus’s shock factor in the transition from a world where “AI could still only do one question and one answer” to one where it could handle long-chain, complex tasks.
In an internal test of a design Agent, he imagined a McDonald’s and giant-panda collaboration. The Agent produced a logo and packaging extensions, then proactively considered what cap delivery workers should wear and how a highway billboard could use the highway setting for a humorous concept. The result was a complete visual system.
He draws out 2 trends. Founders in this wave have “an unusually mature commercialization mindset,” with almost everyone monetizing from day one. They are also highly global from the outset, often operating domestically with one hand and overseas with the other, or targeting overseas markets from day one because B2B and B2C payment environments and enthusiasm for AI are both stronger abroad.

13. Deep Research Changes the Workflow: Do Not Use Only Free Products

刘飞 says Deep Research in ChatGPT and Gemini cut the time needed to write 《半拿铁》 episodes by half. In the past, a business biography could involve tens or even hundreds of thousands of words; when older AI systems compressed that into 2,000 words, too many details were lost. Deep Research can now produce 20,000-30,000-word documents and restrict itself to trusted sources supplied by the user.
卫诗婕 adds that after ChatGPT launched global memory in April, it will gradually learn about the user. “Style is only a matter of time.”
Koji’s core advice is not to judge AI based only on free products. They may leave users thinking “is that all?” because users have not yet reached the scenarios AI can genuinely solve. Those who can afford it should try the leading AI products globally, including o3’s reasoning capabilities.

14. The Tool Split: 豆包’s Compute vs. DeepSeek’s “AI Divinity Moment”

卫诗婕 makes the case for 豆包: only a company as large and well-resourced as ByteDance can let everyone use advanced intelligence for free. With abundant GPUs, 豆包 can quickly generate 20 posters in different styles; comparable functionality in overseas products might cost $100 per month.
程曼祺 is less fond of 豆包, finding that its anthropomorphic interactions still feel “too AI.” For emotional support and talking through her own thoughts, she prefers DeepSeek, which she says teaches her to change perspectives and express herself, producing what she calls an “AI divinity moment.” She also uses 可灵 for image generation, and ultimately used a 可灵-generated design for her own trademark.
Her tool stack combines domestic and overseas products for research, information retrieval, and understanding problems, with OpenAI’s Deep Research, DeepSeek, and 秘塔 as the main tools. She uses Midjourney and 可灵 for image generation; 秘塔写作猫 for proofreading and fact-checking; and 通义 for synthesizing thoughts and generating mind maps. 刘飞 also recommends Monica as a one-stop way to use multiple models.
卫诗婕 cautions that 元宝 and DeepSeek can sometimes over-“interpret” when used for creative work. The so-called “DeepSeek flavor” is not suitable for every setting. Different products have different use cases; users should try several before deciding.

15. AI Brings Information Equality, Not Intellectual Equality

卫诗婕’s central claim is that “AI will not bring intellectual equality; AI will only bring information equality.” People with knowledge structures and hands-on experience use AI more effectively. Cognition rises “like climbing a staircase,” and one may get 2-3 steps ahead of others, but cannot skip the process.
The most dangerous trap is the workplace “screw” who has just left school and has been assigned repetitive work, then expects AI to do everything. “Once you depend on AI to help you do things, you are essentially letting AI replace you.” Maintaining agency means “I use AI to help myself grow”: let AI handle basic labor while humans do more creative and deeper thinking.
Her hierarchy is that people who have worked for years and observe sharply will adapt to AI fastest; the AI-native generation still receiving compulsory education is luckier; those whose learning journey has ended while workplace pressure squeezes them hardest are in the most awkward position.

16. Who Can Still Get a Seat: 谢阳, Dyson, and “Learning by Doing”

卫诗婕 interviewed Fellou founder 谢阳, born in 1996, who had built a B2B project into one of China’s leaders in the category. After setting aside 2 periods to read papers systematically and think, he developed the early form of an AI Agent product. Her 2 takeaways are that a moat is the feel and judgment a person or team accumulates by doing dirty, difficult, exhausting work, and that people from the previous era can still get a seat in this one if they actively embrace the wave.
Koji uses Dyson as evidence that creativity transfers. Dyson started with construction-site carts, replacing ordinary wheels with sturdy, brightly colored balls, and later crossed into products such as hair dryers. “Creating” does not depend on whether you work in the internet industry or AI.
Koji summarizes reinforcement learning as “learning by doing”: build an environment, receive feedback, and keep adjusting. 卫诗婕 adds that large models can also be viewed, at a macro level, as reinforcement-learning models, quoting Alibaba Cloud CTO 周靖人: “Everything is RL.” “Action is one of the core principles of reinforcement learning.”

17. Liangzhu’s Old-School Product Managers and ByteDance’s Yes or No

At an AI Demo Day in Liangzhu—also called 启师傅会客厅—Koji sat in a backyard lawn circle discussing AI with everyone, transported back to the early mobile-internet days of Zhongguancun Street and Garage Coffee. The conversation was not about building a company with RMB10B in market value, but “why I want to build this product.”
Podwise, an AI podcast-summary service with a subscription model, has reached breakeven. Tripo3D turns photos into 3D-printable models and is a global leader in the category. These examples show that in the AI era, founders do not have to start by asking how large a company they want to build; they can begin with a personal need or something they enjoy.
“Classical” refers to a return to the scene after ByteDance and Pinduoduo popularized data-driven and growth-driven methods. The consensus is that large models will produce new Killer Apps, but no one knows what they will look like. GUI may not be AI’s most rational interface, and LUI may not be the future either; AI may have a new interaction model of its own.
程曼祺 relays a distinctive view from a Shanghai AI investor: growth-oriented product people from core businesses such as Douyin may not be the best fit because their systems are complete and feedback is abundant, so they do not need to think through the scenario in such detail. People from places such as Feishu, where B2B feedback is slower and operations cannot rely solely on a strong system, may be better suited.
卫诗婕’s answer is: No. ByteDance talent should not be deemed unsuitable simply because of a growth background. ByteDance has attracted excellent people, is pragmatic and goal-oriented, and its core business lines require employees to maintain observation and hands-on feel for AI products. But she also stresses that foundational innovation requires the kind of all-in commitment seen at BAAI or OpenAI. Model capability grows in steps, and investment must continue through the plateaus. Whether Big Tech can survive the bottleneck is the decision point—and a key reason China lacks an environment for foundational innovation.

18. Giant Blind Spots and the Finale: Small, Beautiful Businesses Still Have a Shot

形色 offers a telling example. Koji believes the plant-identification app should have very high revenue and must already be above RMB100M. 刘飞 adds that it has raised very little and can also identify rocks, birds, and insects. It may not need a large model at all, using earlier vision models to deliver an excellent experience. Koji further notes that it went through years of accumulation and hardship early on; its strength eventually became a moat.
The business cultures differ. 卫诗婕 says the U.S. innovation ecosystem often contains an instinct of “you built this and built it well, so I will not build it.” China instead favors saturation attacks by larger players: “Manus took off, and ByteDance now has 4 teams working on it internally.” But AI’s space for imagination is enormous. Giants lack concentrated attention, and a Big Tech company that has built audio products may still fail to beat 小宇宙, leaving startups with room to win.
程曼祺 mentions that the school at HKU where 马毅 teaches will launch an AI literacy course open to students from all disciplines this September. She also recommends the 《数字生命卡兹克》 episode on how ordinary people can use AI. A useful starting question is: “Which repetitive tasks do I want to hand over to AI?”
Koji closes with Paul Graham’s “How to Do Great Work”: doing great work is not about doing what other people do, but doing what is yours; more specifically, finding things that feel effortless to you but may be difficult for others. 卫诗婕 adds that the work need not be AI-related. It can be something connected to your own feelings and memories that machines cannot replace. “Everyone who lives actively is an actor in this era.”

Rippling vs. Deel Lawsuit: WTF Happens Now? The Future of the Late Stage Private Market The AI Reset Is Here: Search, Jobs, and Everything Else w/ Anish Acharya, Dave Blundin, Salim Ismail