“With AI, I Feel Terrifyingly Powerful!” | A Conversation with 鸭哥

Summary

• 鸭哥 says viral GEV is neither an LLM nor pure hype: technically, it’s essentially “GPT plus JSON Mode,” a classifier that only does classification. Similar Classification APIs have come in several waves over the past 10 years, with Google, OpenAI and Mistral all shipping products that were later abandoned. GEV is catching on now because foundation models are capable enough, agentic applications have made token costs more salient, and “System 1 classifier” is a sharp piece of PR.
• Cost declines are unlocking use cases that once seemed uneconomical: his tour-guide project burns about 4B tokens a day, which would cost an estimated $12,000 on Sonnet versus just $300 on DeepSeek Flash. An open-source setup using Qwen reads logits without fine-tuning or training, scoring 85% against GEV’s 88%. 鸭哥 sees plenty of potential for open models to deliver similar use cases at very low cost.
• A new trend he sees is giving AI a computer that never gets wiped. Manus launched Cloud Computer in April 2026, followed by Grok Bot in August and Meta Muse in September. Persistent credentials and memory take AI beyond the “too much of a coder” GitHub paradigm and into everyday tasks like booking flights and using Dianping. 鸭哥 hasn’t adopted these tools widely: context is hard to migrate, and chat logs can expose keys in plain text—something he doesn’t want leaking onto the open internet.
• The core way to use AI hasn’t changed: it has grown from a “vocational-school student” into an “undergraduate,” but still needs clear direction and close review. When his child suffered nursemaid’s elbow, AI helped assess whether to seek care, find a hospital and handle the insurance claim. 鸭哥 says it made him feel “terrifyingly powerful” (强得可怕) and “not alone in the fight”; he doesn’t expect AI to replace doctors, and still had to call the insurer to move the claim forward. The prerequisite is knowing how to use it: “Give most people 2 secretaries and they still won’t know how to use them.”
• yage.ai is an experiment in running a company with no employees: AI built an X account from scratch to more than 5,700 followers, and 鸭哥 has never written an X post. AI also built its own UTM attribution system, then used the data to adjust its writing style and audience acquisition. The challenge is overcoming the mediocrity of off-the-shelf models: pre-training teaches consensus, while alignment makes expression more cautious. AI needs your way of thinking as context to produce work with a point of view.
• The AI era rewards people with domain expertise who can “manage” AI, not just use it. 鸭哥 says fast coding, LeetCode grinding and design are all losing value; clearly articulating what you want is increasingly rare. He is “optimistic about humans, pessimistic about humans using AI,” because getting value from AI starts with being able to state the task clearly.
• A context-driven life experiment: AI helped him export Apple Health data and run a multivariate regression, finding no link between coffee and sleep duration—but a link between sleep and the time he last used AI that day. After following its suggestion to stop using AI around 7 or 8 p.m., his average sleep time rose by more than an hour in the second month.
• His cyber-wish is for context infrastructure to take on more of the work of execution and collaboration inside companies. 鸭哥 and 课代表 are testing agents that review each other and discuss plans. He still isn’t comfortable with embodied AI—say, a robotic arm applying makeup—or with AI publishing articles without review, though he says that boundary keeps moving.

Deep dive

1. First impression of GEV: GPT plus JSON Mode, repackaged

• 鸭哥’s take: GEV isn’t an LLM and can’t generate free-form text; it does classification only. Give it a set of options and it returns the most likely one, along with a probability and a confidence score. The probabilities are calibrated, and probability and confidence are distinct; ordinary LLM probabilities are often unreliable—“It says 95%, but the actual figure may be only in the 30s.” GEV is especially fast and cheap. After looking into it, he thinks the technology is indeed close to JSON Mode, but can’t see a major difference in the use cases.
• He stresses that “it’s all good”: he isn’t saying GEV is pure hype; there are misunderstandings, but also things worth learning from. Companies have built similar Classification APIs in at least 4 or 5 waves over the past 10 years, and Google, OpenAI and Mistral have all shipped similar products. Those products were eventually discontinued and went unused.

2. Why it didn’t catch on before—and why it has now: 3 reasons

• First, foundation models are finally capable enough. On AI Index, GPT-4o scored 8, while the previous generation’s Qwen3.5 4B, with 4B parameters, already scored above 20. 鸭哥 says tiny models can “crush GPT-4o”—a powerful model that can be used directly in many commercial settings.
• Second, the rise of agentic applications is driving up token consumption and making users more cost-sensitive, creating demand and a market for low-cost, low-latency classification. Third, the positioning is smart: say “OpenAI shipped an API 3 years ago, it died, and we’re bringing it back,” and no one pays attention. Call it a “System 1 classifier” and a paradigm shift—and add that the author co-wrote papers such as InstructGPT—and people take notice. 鸭哥 calls it “a brilliant piece of PR.”
• Koji’s sardonic analogy: it’s like announcing, “The world has men and women, and you guys somehow only knew about men.” 鸭哥 agrees the packaging is clever. Koji calls it a marketing win—and a technology whose moment had arrived finally boiling over.

3. Two impressive applications: browser use and harness classification

• Browser/computer use: swap screenshots for a DOM tree and free-form text output for classification; classification can even be cascaded, first finding the broad area and then the small button within it. 鸭哥’s impression is that browser use can be 10x faster and cheaper. But he cautions that it’s not clear which change matters most; perhaps an LLM with a DOM tree alone would already be faster and cheaper. More ablation studies are needed.
• Harness classifiers: tools such as Codex and Claude Code check whether each tool call is dangerous. These checks happen frequently, making latency and cost important. Risk assessment is flexible and complex, and has already become a kind of general-purpose classifier. Popular harnesses add proprietary classifiers; GEV lets open-source projects port similar capabilities. Without GEV, people were already using this approach without realizing it.

4. “Two years of training” raises questions: open-source substitutes and token economics

• 鸭哥 looked for relevant patents, papers, white papers, open-source databases and code, but found nothing. He suspects “two years of training” may mean more that people tried and found it didn’t work; the problem itself is hard, and required waiting for foundation models to mature before building on top of them.
• An open-source implementation uses Qwen and reads logits only, with no fine-tuning, training or additional data. GEV’s benchmark score is 88%, versus 85% for Qwen. 鸭哥 thinks achieving comparable results under those constraints shows that open models have considerable scope to power very low-cost applications.
• GLM-5.3-Flash and DeepSeek 4.1 Flash, for example, make data labeling cheap. 鸭哥 built a tour-guide project covering dozens of Chinese tourist cities that used about 4B tokens a day; Sonnet would have cost an estimated $12,000, while DeepSeek Flash cost just $300. Running Qwen3.5 4B locally on a Mac, 4090 or 5090 means the main cost is electricity. Use cases that once seemed impossible are now viable.

5. A new trend: giving AI a computer that never gets wiped

• Unlike early Cursor Cloud Agents, which started from scratch each time, Manus launched Cloud Computer in April 2026: the VM persists, along with credentials and memory. Grok followed with Grok Bot in August, and Meta with Muse in September. 鸭哥 says this moves AI beyond the “too much of a coder” GitHub paradigm and into everyday tasks: booking flights, remembering identity details and using Dianping without logging in again.
• He compares the idea to OpenClaw, but without having to configure it yourself, buy a Mac, keep it running 24/7 or set up network tunneling. Koji jokes that anyone carrying a MacBook around and afraid to shut it down is obviously a Claude Code or Codex user.

6. GPT-6 Astro takes down an industry: a Psyduck VTuber

• 鸭哥 used GPT-6 Astro to create a 3D Psyduck model and rig its skeleton, then used a camera to capture his own movements so Psyduck could copy them—a VTuber prototype. Building that kind of avatar once required a 3D model, textures, rigging, and both art and technical staff. Now he got GPT-6 Astro to do the whole thing from scratch with a few prompts, and open-sourced the skill.
• He says model capability has “taken down an industry.” That doesn’t mean it has fully replaced one, but it points to a direction, and the pace of progress is exciting.

7. The method hasn’t changed: an undergraduate and a nursemaid’s-elbow incident

• The core principle is unchanged: don’t treat AI as a tool alone; it’s more like a person. It makes mistakes, so check its work, and be clear about the goal and what success looks like. “I used to manage a vocational-school student; now I manage an undergraduate. The capability is stronger, but I still have to manage them.” Koji recalls the episode he recorded with 雨森 2 years ago, “Large Models Are Still in Elementary School—Don’t Rush to Put Them to Work.”
• After his child was injured at kindergarten, AI suggested the injury might be nursemaid’s elbow, described the symptoms and asked questions. The parents thought it fit, and AI said there was no need for the ER but urgent care was an option. It checked nearby hospitals and suggested Seattle Children’s Hospital, advising them to go soon because nursemaid’s elbow doesn’t resolve on its own and the child shouldn’t have to stay in pain. At the hospital, a doctor performed a reduction maneuver and fixed it quickly. 鸭哥 doesn’t expect AI to replace doctors, but says it provided useful information for the decision.
• Insurance was straightforward because the school had comprehensive coverage; it was mainly a matter of filing a claim. AI had the context on the injury, hospital and insurance, and after being connected to his email, helped assemble the paperwork, identify recipients and people to cc, and draft emails. The only thing 鸭哥 had to do was call and push for progress; then the insurer moved the claim forward and sent a check.
• Later, when another parent’s child was bitten and they needed to file a claim, 鸭哥 had AI summarize the previous process, contacts and sequence, then write one document for a person and another for an agent to execute. He gave the documents to AI, which recognized their value and followed them quickly.
• Two lessons: without context, AI may offer only the correct but useless advice to “go to the hospital”; and you need to manage AI like a secretary, building structure, checking quality and letting it accumulate context and SOPs. “The highest level of management is when the team runs normally in your absence.”

8. “Terrifyingly powerful”: a doctor, legal assistant and admin team

• When his child was injured and crying 6 months ago, with no clear idea whether to go to the hospital, 鸭哥 might have been “completely overwhelmed.” Now it feels like having a personal doctor, legal assistant and admin team. “I’m not alone in the fight”; these helpers can work 24/7 and can be replicated without limit.
• The prerequisite is still knowing how to use them: “Give most people 2 secretaries and they still won’t know how to use them.” 鸭哥 thinks many people who find AI useful are business owners, because they already know how to work with secretaries.

9. Taking a gap at 35: AI clears the grunt work, leaving only the thinking

• The main reason 鸭哥 took a gap was that, at 35, working 8 hours a day felt exhausting. Heavy AI use got the physical work done, leaving only cognitive work; the workload went up rather than down. He jokes that he used to imagine AI attending meetings while he coded; now he goes to the meeting and AI finishes the code, so he’s even more tired. Koji says coding brings flow, and mentions a DeepSeek researcher’s essay, “I Had to Bury My Talent in Yesterday.”
• The second reason: after meeting many people and companies over the past 2 years, he concluded that most people use AI the wrong way. Using AI isn’t a technical problem; it’s a management problem. When his gap ends, he plans to work full-time with 课代表 at Superlinear Academy from October, teaching more people how to use AI properly.
• After a month back in China, his impression was that U.S. consumers have it rough, while the consumer experience at home is good. Koji relays a friend’s view: once material needs are met, people still need relationships, memories and emotion. 鸭哥 thinks that makes sense.

10. The eyes experiment failed; the recordings continue: everything is a file

• 鸭哥 says he stopped his experiment of giving AI eyes—using a camera to take lots of photos every day—after 3 or 4 months, because he couldn’t see what additional context the photos provided. He still records audio, especially meeting recordings.
• He built an open-source intake skill for the recordings. After transcription, AI indexes the content in 2 ways: one like Yahoo, organizing the index by topic, and one like Google, using embeddings for semantic search. Together, the 2 approaches let AI search local files and find relevant material.
• Context files stay on his local machine. He believes “everything is a file” and that there’s no need for an embedding database or a MemGPT-style memory layer; the Manus article on context engineering is “worth $1M.” Files plus CLI are enough. He’s against using MCP or memory for their own sake.
• One reason he hasn’t adopted cloud computers widely is that context is hard to migrate completely. Also, many of his AI services are zero-data-retention, but chat logs can still contain keys in plain text; he doesn’t want Grok Bot accessing them. Koji says backing up WeChat chats locally for Codex to read was “like adding wings to a tiger”—for example, using an event attendee list to find people with past connections or mutual friends.

11. The zero-person company yage.ai: why off-the-shelf AI tends toward mediocrity

• 鸭哥 explains it through the training process: pre-training is next-token prediction, so models mostly learn consensus. “Consensus has no incremental value; from an investment perspective, it has no alpha.” Alignment makes models safer, more cautious and more all things to all people. “Jobs would definitely get killed by alignment,” and Musk would be similar. So off-the-shelf AI used directly to run a blog or newsletter with strong opinions tends to produce work that’s correct but mediocre.
• The fix is to inject your way of thinking as context. For technical research, for example, 鸭哥’s AI asks about both horizontal research—competitors, alternatives and advantages—and vertical research—how this type of company has changed over time and why. That helps surface things ordinary AI might miss. Readers tell him his articles “don’t read like AI” and are “thoughtful.”

12. Give it eyes and hands: a self-improving loop with 5,700+ followers

• AI initially suggested running an X account: “You have a newsletter, so why not build a presence on X?” 鸭哥 told AI to figure out how and execute it itself, and paid for the work. The account now has more than 5,700 followers; he has never written an X post, and AI writes all the articles.
• AI realized it couldn’t distinguish traffic from X, the newsletter or WeChat, so it built a UTM attribution system. The data showed that newsletter subscribers were high-quality users, while many X visitors took a quick look and left. AI then adjusted the writing style and audience-acquisition copy accordingly.
• 鸭哥 says he’s given AI “eyes and hands”: the hands let it post autonomously; the eyes connect it to Google Analytics, Google Search Console and X analytics, enabling data-driven decisions, self-improvement and SOPs. Sometimes he tells it an article “doesn’t sound like me,” and it improves based on the feedback.
• AI’s views can represent 鸭哥’s: it imported his blog, WeChat chats and conversations with AI through skills, then synthesized his opinions. In another experiment, it downloaded 100 videos by “Yongge Talks Restaurants” on Douyin and transcribed them, drawing out 20 lessons on opening a restaurant. One example: watch customer flow, not just headcount; before opening a breakfast shop near a middle school, check whether roadside railings force students to take a long detour. He thinks you can try to “distill” yourself or a business owner.
• AI also helped him articulate the core idea behind context infrastructure: collect and organize context, let AI consume it, then make the output new context—creating a closed loop and compounding returns. 鸭哥 says AI is “the boss”; it thinks more deeply than he does.

13. Can it make money? Old duck soup takes time to simmer

• 鸭哥’s answer is “yes and no”: yage.ai, as a teaching case for Superlinear Academy, brings in real revenue, but that revenue depends on an existing commercial loop. It doesn’t mean the website itself has found a new use case and proven a business model. He hasn’t yet seen AI independently find a need, customers and a way to deliver just by being told to “go make money,” though AI has sharply lowered the barriers to business.
• Can others reproduce his approach with the skills and thinking he open-sourced? He thinks they can reach the same ceiling, but it won’t work out of the box. He calls the system “old duck soup”: “You can take the old duck soup, but it’s still my soup, not yours; you have to simmer your own and raise your own shrimp.” You have to use it in your own context, express dissatisfaction and save the feedback in files. Out of the box, you get 鸭哥’s approach; with enough feedback, it can become your own system.

14. Managing AI, not using AI: which skills are gaining value

• 鸭哥 thinks coding fast and grinding LeetCode are no longer valuable, and design is also losing value. Fable 5 and GPT-6 already have strong design capabilities, at the level of a senior engineer. What’s becoming more valuable is management: getting AI to build context, evaluation and self-improvement loops.
• He is “optimistic about humans, pessimistic about humans using AI”: using AI well requires clearly articulating what you want to do, and that is a scarce skill. He suggests practicing with AI and asking it what is unclear about the request.
• Managing AI differs from managing people: AI doesn’t need motivational speeches, but requirements, context and other information must be externalized into files or prompts. 鸭哥 says to ask “how to manage AI,” not just “how to use AI.” Treating AI as something to manage lets you get more out of it.

15. Raising kids, teaching students and the software he’ll still pay for

• 鸭哥 thinks kids should be exposed to AI early. He built an embedded e-paper device so his child can touch it, interact with it and communicate by voice. The child also uses Siri and HomePod to listen to songs from KPop Demon Hunters, spending 30 minutes a day on the device. 鸭哥 shared the context documents with another parent, who quickly used Codex to build a device for their own child. 鸭哥, his student and another friend had never done embedded development; AI wrote the code.
• One trait of strong students, he says, is that they actually build things after learning. “AI is like swimming; you can’t learn it just by watching videos.” When 课代表 teaches people to create content, the advice is to shoot 20 videos first, then come back with questions.
• Koji mentions that software payments on the App Store fell for the first time last quarter. 鸭哥 would still pay for 2 types of software: products like Office that everyone uses and can’t avoid, and products that require a lot of work to keep accurate and maintain. He learned the lesson from studying cryptography: “Don’t implement cryptography yourself.” Koji adds that he’s heard the 1Password team has 1,000 people.

16. A cyber-wish: context infrastructure in the enterprise—and what he still won’t hand to AI

• 鸭哥 wants context infrastructure to take on more execution and collaboration inside companies. Take Amazon’s Leadership Principles: even after repeated meetings and calls for employees to role-model them, some still lack bias for action. If AI were executing the work, embodying the principles might take nothing more than adding a prompt or writing markdown. 鸭哥 thinks AI has an advantage in consistency and following instructions.
• He and 课代表 are already testing agent-to-agent collaboration: another person’s agent sends a draft email, his agent reviews it through email and replies, and the 2 agents discuss the issue over several rounds before settling on a plan. He imagines this scaling inside companies and changing how people collaborate, document, execute and evaluate work.
• He still isn’t comfortable with embodied AI, such as a robotic arm applying makeup, or with AI publishing articles without review. But he says that boundary keeps moving, and he’s gradually come around to the zero-person-company experiment.