How Power Users Use AI, How Everyone Else Learns AI, and How Investors Invest in AI | 立正 Talks with 课代表

Summary

• 课代表’s core call: stop using ChatGPT for chat and switch to Agent tools such as Claude Code, Codex, and Cursor. Chat has three cardinal sins—context must be reintroduced every time, output does not persist, and the tool set is limited—so “it cannot compound,” trapping users at 30%–100% productivity gains; as with electric motors replacing steam engines, real productivity comes from rebuilding the workflow around the new tool, which is the prerequisite for the “10x path.”
• The tools and models are the same for everyone; context is what creates the gap. Skill—the codification of tacit knowledge—along with quality standards and accumulated materials all count. Skill itself is hard to commercialize: “It’s too open-source; once you buy it, you can sell it,” but “skill as a service” works. The revenue proof point: the 2-person team of 课代表 and 亚哥 generated about $500k in 2024 and just over $1M in 2025.
• OpenAI and Anthropic take 90% of US AI-related software revenue, including models. Cursor, Manus, and every other application share the remaining 10%—don’t pick up pennies in front of a bulldozer. At the same time, “the stone you used to cross the river”—the previous generation’s business model—has disappeared: after a customer discusses an A/B test with Codex, the results are already in, so why buy your platform? There are 2 escape routes: jump upward into AI for science and embodied intelligence, or go deep underground and transform industries in an FDE-style model.
• The valuation of embodied intelligence is not about near-term commercialization. Performance, leasing, data collection, and sales to research institutions “are not a large market in aggregate”; today’s valuation is the discounted value of a future stretching to the stars. China already has 10–20 companies valued above RMB10B, with one approaching RMB30B. The bubble is enormous, but don’t rush to call it a scam: “9 out of 10 bubbles in history ultimately produced something of lasting value.” Agent deployment should be measured in decades, per Karpathy, while demand for FDE roles is set to “explode.”
• 课代表’s portfolio logic, explicitly not investment advice: hold Tencent with conviction on a WeChat Agent thesis. It has China’s largest user base, the context of every social relationship, and the “one-of-a-kind OS in the world” that is the mini-program ecosystem; “AI has a great deal of time left to give Tencent.” The same logic applies to Kuaishou. Video models are already making money: Kling’s ARR has reached $500M, while “C Dance” (possibly Seedance) is at $1.5B. Short-drama creators who know how to use AI will “instantly crush” peers who do not.
• To anxious big-tech workers, his message is: “What will you do after you’ve finished worrying? There has never been a savior.” Take 3 small steps: save money—“this money is your future freedom”; stop using ChatGPT so the brain can enter the new reality—“if you remain in the old reality, how could you know how to make money in the new one?”; and learn basic business, because the big-tech belief that “harder things are more valuable” is commercially false.
• Software is not dead, but it has to be reborn. The smallest unit is workflow plus knowledge base plus interface, and any one of the 3 can be made valuable—主持人 would pay $100 per use to study Mark Andreessen’s and Peter Thiel’s process for evaluating founders, “at least 10 times”; a polished interface such as Vibe Island can still command payment. Yet 课代表 turned his own book into a free skill on his website and “hardly anyone used it, hehe”—the market still needs to be educated on new ways of delivering value.
• The contrarian idea: 课代表 wants to be an “unproductive role model.” With AI, people do not need to earn 10x more; they could do the same amount of work and spend more than half their time playing, “so everyone would stop competing so hard.” The experiment failed: free time became Douyin and Xiaohongshu; even though Claude Code became what gamers called “the most fun game,” he remained anxious while building—“I’m wasting tokens.”

Deep dive

1. How Power Users Use AI: Always Asking What They Can Do for AI

• Start with the sample: 课代表 is 37, holds a PhD in economics from Cornell, and has worked as an economist at Amazon, a data scientist at Meta, deputy director of game data science at Tencent, and Principal Data Scientist at Stastic, later acquired by OpenAI. In March 2024, he and 亚哥 started an AI education community and corporate training business as a 2-person team; revenue was about $500k in 2024, just over $1M in 2025, and about $500k in 2026, while the team has only recently expanded to 5.
• In 课代表’s view, 亚哥’s 2 defining traits are fundamentally different. He is always thinking about what he can build for AI—creating scaffolding around context, tools, preferences, and standards of work, so that a single prompt can route to a large library of existing tools and materials. He also asks AI about everything, big or small, anywhere: taking a spin in a Cullinan using Tesla Autopilot, completing 10 tasks on his phone, and having research automatically written into a Google Doc and sent to 课代表 via Gmail.
• The host adds a case study: a friend wanted to learn about world models, so he fed Claude all the authoritative papers and podcasts and asked it to turn them into a 30,000-character melodramatic short story about office romance and startup competition. After reading it, “I felt like I understood world models a little better.”

2. The One Thing Everyone Must Do: Stop Chatting and Switch to Agent Tools

• 课代表’s sole must-do: “You have to stop using ChatGPT.” Pick one of Claude Code, Codex, or Cursor; Antigravity also works, while “Little Lobster” is not recommended. The analogy is the replacement of steam engines by electric motors: the tool itself delivers only 30% productivity gains, while the real leap comes from rebuilding the workflow around the advanced tool. Chat remains a workflow built around the old mode of production—using AI to replace a search engine or an intern.
• Chat has 3 cardinal sins: context must be reintroduced every time, output has to be manually copied and pasted because it does not persist, and the available tool calls are limited. “Those 3 things mean it cannot compound.” AI has hundreds of times the brainpower of a human, but chat requires a human to sit alongside it, turning the human into the bottleneck. Chat leaves users stuck at 30%–100% gains; Agent tools are a prerequisite for 10x productivity.
• His own migration path: “Before 4.6, I used Claude Code the most; after 4.6, I used Codex the most.” GPT-5.5 was “an extremely impressive update,” especially for programming. 课代表 is using Codex more and more for simpler reasons: its GUI feels friendlier, it can access ChatGPT history, and its connection success rate with Notion is higher.

3. The Anti-Productivity Manifesto: Becoming an Unproductive Role Model

• 课代表 says his growth over the past 6 months has not been about capability but about figuring out a lot of things, especially that “productivity is not such a righteous pursuit.” Could people do the same amount of work and earn the same amount of money after AI arrives, then spend more than half their time playing? “Then everyone would stop competing so hard. I don’t need to earn 10x more.”
• The experiment has not gone well: all of his unproductive time went to scrolling Douyin and Xiaohongshu. “That’s not right. I should find things that are unproductive but more meaningful, like going hiking.”
• The host offers a counterpoint: when a friend in gaming asks people around him what the most fun game they have played recently is, many answer Claude Code. Creation itself is entertainment. 课代表, however, struggles to enjoy it fully: “On the one hand I think it’s very fun; on the other I’m extremely anxious—I’m wasting tokens. What’s the point of building this? No one will buy it.”

4. Skill and Context: 2 Compounding Advantages

• A good skill makes tacit knowledge explicit: define the boundaries clearly, provide the necessary tools, specify what good and bad look like, and supply the context. 课代表 criticizes the current trend of writing style-only skills such as a “Socratic skill” or a “departed colleague skill”; real skills are about process, tools, and standards. His own examples include video post-production—connecting Qwen3 ASR for transcription, Claude Code for cleanup, and automatic generation of highlights, articles, titles, and editing briefs—and automatically rebroadcasting YouTube videos as podcasts and uploading them to Transistor.
• Can skill be commercialized? “It’s difficult—it’s too open-source. Once you buy it, you can sell it.” As products make less and less sense, services still make sense: “skill as a service may be more logical.”
• The third step is to accumulate and refine context: “In the end, everyone’s tools are the same and everyone’s models are the same. What actually separates people is their context.” The compounding is already visible: in corporate training, “give me an even higher bar and I can deliver within 2 days,” because enough context has accumulated.

5. The Stone Used to Cross the River Has Disappeared

• The collective state among frontline founders: ComfyUI is facing a “huge strategic decision” over what to abandon and where to concentrate resources, while Opus Clip’s attempt to build an AI Agent is also full of question marks. 课代表’s diagnosis is that the previous generation’s business model has disappeared: the old playbook of identifying a common large demand, building a company to deliver a product, and taking the market through go-to-market no longer works.
• The demand-side example comes from his former company, Stastic, which sold an A/B testing platform. “I could give you 100 reasons why you should use Stastic,” but after a customer talked to Codex, “the A/B test results were already there,” leaving him unable to explain why the customer should buy the platform. On the supply side, “copying is too fast,” making it difficult for anyone to build enough accumulation to reach escape velocity.

6. Big-Tech Workers: There Has Never Been a Savior

• He laughs when he hears big-tech workers talk about anxiety, calling it a “pathological revenge thrill.” In early 2023, he wrote “The 5 Most Important Questions About ChatGPT” and made many bold predictions, all of which were right, including warnings that jobs would be replaced in 5 or 10 years. The response was: “You’re an idiot. You’re being a Xianglin Sao.” Now his question is: “What will you do after you’ve finished worrying? There has never been a savior.”
• The host cites psychologist 李松蔚: grand principles and advice are often useless; what works is “0.5% action”—one tiny step that generates positive feedback and gets the positive loop started.
• 课代表 offers 3 small steps: save money—“this money is your future freedom,” so do not spend it on meaningless things; stop using ChatGPT and switch to an Agent—“if you cannot adapt to the new productive forces, your brain has no way to live in a new reality, and you cannot plan for a future”; and learn basic business and sales. Big tech instills the mistaken belief that “harder things are more valuable”; watch 勇哥’s restaurant-location videos and ask the restaurant downstairs how you can help increase its sales.

7. The Second Renaissance: The New Generalist Learns AI as Verbs

• 课代表 is optimistic about the next generation but pessimistic about our generation’s transition. AI will bring a “second Renaissance”: education trained us to be tools and screws, and once AI fully replaces those roles, “people will be forced to discover themselves.” His parenting approach is to build more intrinsic motivation, teach less, and protect creativity.
• The host’s education vision for the AI era is not to master one specialty but to “know a thing or two about more than 100 fields.” As long as you can ask a question, AI can keep expanding the answer; if you cannot ask one, “the entire world is darkness to you.”
• 课代表 agrees that a Da Vinci-style new generalist will become important again. His framework: “You think learning AI means learning nouns, but actually learning AI means learning verbs.” Swimming, golf, and making a podcast are all verbs. In Latin, Manus means “hand.”

8. The 2026 Reality: Foundation Models Swallow 90% of the Value; Go Up or Go Deep

• After the balance between offense and defense has shifted, 课代表 asks the question; the host’s frontline view is that 2026 is a year of unprecedented startup heat and capital enthusiasm. Frenzied capital narratives and obvious bubbles coexist: “The other side of the bubble is that confidence in technological progress is at an unprecedented high.”
• The line that stayed with him most came from the founder of Paper Boy: “The best way for humans and AI to collaborate may not have been invented yet.” We are still copying and pasting, AI still does not understand us, and friction is everywhere, which leaves room for founders. “Pessimists are often right, but only optimists have a chance to succeed.”
• The structural fact is that OpenAI and Anthropic account for 90% of US AI-related software revenue, including models. Cursor, Manus, J Spark (possibly Genspark), HeyGen, and every other application divide the remaining 10%—“don’t pick up pennies in front of a bulldozer.” There are 2 ways around the bulldozer: jump upward into moonshots such as AI for science and embodied robots, or go down to earth and embed yourself in an industry.

9. What AI for Science Means, and the Honest Truth About Embodied Valuations

• 课代表’s standard is that science means advanced knowledge work that an ordinary white-collar worker cannot participate in today: AI-driven chip design, new-material discovery, mineral exploration, and quantum-computer design. Ordinary white-collar work is already being handled well by models and general Agents; the opportunity is to tackle problems solvable only by smarter minds. AI for math is not an industry but a method: requirements must be expressed with mathematical rigor to enable rigorous verification and reinforcement learning. The boom is also reinforced by professors and PhDs starting companies, alongside strong capital demand.
• 课代表 asks whether the valuation of embodied intelligence is an investment question or a demand question. The host’s honest answer is that current commercialization consists of performances, leasing, data collection, some industrial applications, and sales to research institutions—“taken together, it is not a very large market.” China already has 10–20 companies valued above RMB10B, with one approaching RMB30B. “It is not that these companies have already generated this commercial value and are therefore worth this much; it is the discounted value today of the future stars and seas.”

10. FDE: The New Role That Acts as HR for Digital Employees

• About 1 month ago, OpenAI and Anthropic issued announcements independently but at the same time, partnering with or acquiring PE and consulting firms. The common thread was FDE, or forward-deployed engineers: people who map an enterprise’s processes, culture, context, and databases, then embed AI into every part of the work. “It’s not enough to give every employee a DeepSeek.” The role is closer to HR onboarding digital employees; it used to be called pre-sales, customer success, or an on-site engineer.
• There is no ready talent pool and no shortcut that lets someone switch careers after 1 week of training. “There should be an explosion in demand for this role,” making it a good direction for big-tech workers. Why has it not happened yet? Innovation diffusion follows its own timeline: Andrew Karpas (possibly Karpathy) posted at year-end that Agent deployment should be measured in decades. Today is only the starting point; even counting from Manus’s launch, it has been just 1 year and 4 months. Electricity also took a very long time to transform factories.

11. Who Is Making Money With AI Today: WeChat Agent Is 课代表’s Conviction Holding

• 课代表 lists the winners: applications that keep disclosing ARR, including Cursor, Lovable, and Manus; model companies; and people who bought shares in 2 Chinese publicly listed model companies. Short-drama and slow-drama writers who know how to use AI “instantly crush” peers who do not, while foreign-trade operators use AI for product selection and ad placement.
• His own holding, “not investment advice,” is Tencent, held with conviction not because of its scale but because of the WeChat Agent thesis. China’s largest user base has the context of every social relationship—“your homeroom teacher, your ex”—while the mini-program ecosystem is an “extraordinary, one-of-a-kind OS in the world” that can handle insurance purchases, visa applications, and SF Express tracking. “Give another product 5 years—can it build up all this accumulated context? No. AI has a great deal of time left to give Tencent.” The same applies to Kuaishou, while Kling has value as a business in its own right.

12. 课代表’s Current State: Compounding Without a Goal and the New VC 3-Piece Set

• His current setup is the 十字路口 podcast alongside a role as an investment partner at ZhenFund—a strong synergy. After 3 zero-to-one startups, he is waiting for “native passion” to return before starting another company: “Starting a company impulsively, or starting one just for the sake of starting one, definitely won’t work.” Entrepreneurship is always a tug-of-war between real capability and a sufficiently high goal, which “translated means anxiety.” For now, “I can have no goal”: do a few things he is certain are worth doing, and “they will become a friend of time.”
• The host’s new VC framework is narrative connector plus connector as a service plus a new kind of financial partner: help founders find a narrative and amplify it, make horizontal connections by finding customers, meeting clients at dinners, and persuading hesitant candidates, and provide hiring and financial infrastructure. His own comparison is Stastic: Sequoia’s resources in the early days “didn’t feel that strong,” while the iconic investor in the Series C took the company to major customers and dinners, creating far more opportunities for portfolio-company collaboration.
• 课代表’s view of investors is “self-reliance makes everyone strong”: the best founders do not need a VC telling them what to do. If a founder particularly needs a VC to handle ABCDEFG, that is cause for concern. A good investor is a copilot who adds value when necessary, not someone constantly telling you to watch the traffic lights. As for community, “community is like water—it can be large or small”; he has no energy to cultivate one, preferring small-circle socializing through dinners and pizza gatherings.

13. For Big-Tech Workers Who Want to Start a Company: Solve One Small Problem of Your Own

• The 3 principles remain unchanged: find something you want to build so badly that you would not sleep; if you have not found it, do not use entrepreneurship as an escape from an unhappy job; and find something effortless for you but difficult for others. For example, researching tea for 3 hours feels easy and absorbing, and after you explain it, friends are persuaded to go buy tea. Products rooted in real demand are more likely to find PMF.
• The zero-to-one stories of 2 companies that recently went viral worldwide—Open Cloud (possibly OpenClaw) and “Hermès” (phonetic)—share the same starting point: both solved a small problem of their own. The former began with: “Web Coding is so fun I don’t want to eat. What do I do? I have to eat.” So it built a small tool that lets people continue coding on their phones.

14. 5 Directions, Part 1: Everything Agent and Humility About Bubbles

• The first major direction 课代表 is watching is everything Agent: Agents will penetrate the capillaries of white-collar work. Infrastructure opportunities include sandboxes, memory, communications networks, and payments, although “Agent payments may very well not be an opportunity for startups.”
• A related opportunity is designing products for Agents. Feishu’s CLI is built for Agents, and its recent surge in word of mouth is related to its active embrace of the Agent ecosystem. The warning is clear: “If an Agent cannot find your tool, your tool does not exist for it,” just as your app does not exist if it cannot be found by searching for Meitu in the App Store.
• The second direction is physical AI. The bubble is enormous, and the reaction that “it’s too crazy, everyone is here to fleece investors” is understandable, but 课代表 regularly reminds himself not to jump to that conclusion. With so many smart people involved, it is impossible that everyone is wrong. “9 out of 10 bubbles in history ultimately produced something of lasting value,” so it is still worth spending time on embodied intelligence and world models.

15. Video Models: A Renaissance for Small-Town Youth and the Short-Drama Debate

• The third direction is the commercial value unlocked by video models. Kling’s ARR has reached $500M, while “C Dance” (possibly Seedance) has reached $1.5B. In a small town in Yunnan, a young man who shoots wedding portraits made the globally viral short film Zombie Cleaner by himself—“cinematic quality, an incredible story.” This is a productivity revolution and a second Renaissance in which everyone can release what is in their head as a finished work.
• On the short-drama hierarchy, the host invokes Zhou Enlai’s words to cultural workers: “If you don’t like it, who do you think you are? What the people enjoy is what all of us should be making.” The debate between high art and grassroots work has existed since antiquity and is difficult to settle. 课代表’s self-mocking conclusion is worth preserving: “I may still want to watch xianxia more. I don’t stigmatize short dramas, but I despise the version of myself who likes short dramas.”

16. Software Is Not Dead: Nail the Workflow, Knowledge Base, or Interface; Voice Is the Native Interface

• The fourth direction: against the backdrop of collapsing SaaS stocks, 课代表 believes software will not disappear but must be reborn. The smallest unit of software is workflow plus knowledge base plus interface; a well-built Excel file with formulas contains all 3. Master any one and it has value: Vibe Island concentrates Claude Code task status in the Mac Dynamic Island, and “when it is polished as exquisitely as a piece of art, you are still willing to pay for it”; data can be valuable when it is private and unavailable through public-market tools.
• The workflow itself can be sold: “How do Mark Andreessen and Peter Thiel talk to founders, and how do they evaluate startup teams? Charge $100 per session. I would buy it at least 10 times to learn.” 课代表’s own data point is a cold shower: he distilled his book The Real Thing into a skill and put it on his personal website for free, but “hardly anyone used it, hehe.” New forms of value delivery still need market education; “maybe one day a super app will appear.”
• The fifth direction is everything voice-related: “Typing is not the most comfortable interface for human communication, and Chinese is genuinely difficult to type.” WeChat’s zero-to-one story began with voice chat inspired by Talk Box; speech carries far more information and nuance than text. There is also voice-interactive content: 亚哥 built a Flappy Bird controlled by the pitch of your voice, “with the whole family just sitting there making donkey noises.”