Original · Unique Research · 2026-04-09
Editor's note: This English edition retains the original author's first-person narrative and the panelists' statements; these are not the translator's firsthand experiences. The source describes six cuts from a team of ten, but later quotes Ray as saying four people were cut at year-end. Both passages are preserved without inventing a reconciliation. It also calls Bella a one-person company while describing two interns. Business metrics are the speakers' reports, not independently audited results. The moderator's wording about Google and Character.ai, the statement attributed to Musk, and specific model references require verification before release. “This year,” “last year” and predictions retain the original April 2026 context. Assertions about AI superiority, employment, human emotions and future autonomous companies are the source's opinions or scenarios, not independently established facts or medical findings.
Unique Awards
In the AI era, the real barrier to entry is not technology or data. It is people.
In the AI era, founders' organizational structures, approaches to efficiency, and judgments about competitive barriers are being rewritten.
At a roundtable during Hangzhou AI WEEK, the moderator asked the four founders onstage: How many people are in your companies now?
The first said four. The second said four. The third said just me and two interns. The fourth said ten.
I immediately thought this was interesting. Those four answers essentially represented three ways founders are operating in the AI era.
Even more striking was what one of the people who said four added: We had ten people last year. We have now cut six—two UI designers, two front-end developers, and two back-end developers.
Why the cuts? Because OpenClaw and Cursor came along.
From Ten People to Four
This founder is Ray, CEO of EZsite AI. He previously oversaw used-car e-commerce at ByteDance.
He laid out a very direct rationale: After OpenClaw and Cursor came along, they realized they could replace people entirely with AI. So they took that approach to its limit and cut six programmers.
His new brand is ClawLite AI. From customer acquisition and website building to lead conversion and closing sales, he built everything himself using OpenClaw and Claude Code.
No one else was needed.
I wanted to ask right then: What about the six people who were cut? But I did not ask. The answer seemed obvious—they had been displaced by the market.
Not by Ray. By AI.
Ray said that recently, he had been handing 80% of his work to AI and doing only two things himself: First, set goals for AI. Second, Validate—check whether the work AI completes meets the requirements.
It was that simple.
Someone asked whether the level of automation was really that high. Was no human correction needed?
He said he had initially done the corrections himself, but gradually built an Agent to do QA for him. The more technically knowledgeable someone is, and the better they understand the business, the less comfortable they are entrusting it to AI. But he took the opposite view: Cursor and Codex would certainly write better code than he would; AI was also better at generating Marketing materials or choosing channels.
On that assumption, he handed most of his work to AI.
He also said something that stayed with me: In the OpenClaw era, the people who use OpenClaw well are often those who are not particularly expert in the business and not particularly strong at coding.
Put another way: People with patchy expertise can use it better than experts.
Because experts have too much misplaced confidence.
A True OPC Has Only One Person
Another person onstage, Bella, was a true OPC—a one-person company.
With just herself and two interns, she helps AI companies acquire customers through content on overseas social media. On LinkedIn and TikTok, she had generated monthly exposure in the tens of millions.
She shared her approach: Use AI to amplify personal capabilities and increase leverage.
For example, she cannot watch scripts and traffic figures every day. So she has AI review the data, analyze the key factors in scripts that perform well, and then produce iterations and variations.
But she still makes the core judgments. Which KOC creators fit the Vibe, and what kind of Storytelling script will resonate with users—AI cannot make those aesthetic judgments.
She said something I thought was right: My taste, combined with AI, can dramatically amplify my personal leverage in the OPC era.
It reminded me of a particularly interesting phenomenon.
There are many courses on using AI to create content. They all teach “how to get AI to write for you.” Bella's approach is the reverse: It is not that AI helps her write; she tells AI how to write.
That distinction in who leads determines the quality of the output.
Where Are the Boundaries of an OPC?
The moderator, Hua Rongqi, asked a good question: When you use AI to run an OPC today, where do you think the boundaries are? Which parts have you wanted AI to handle, only to discover that people still need to do them?
Yifan, founder of Vaniloom, said that pre-sales work involving people, and anything involving building trust, still require personal involvement. This is particularly true for niche communities, such as customized otome-oriented fan content. Users have a strong sense of community belonging, so founders need to build relationships and emotional connections with customers themselves.
Sawyer, product lead at Meowster, said human emotions vary widely. When someone is unhappy and venting to you, you cannot respond like customer service and say, “Oh, you'll be fine.” That requires judgment and guidance. His team had developed many standardized scripts to constrain AI's output and maintain a consistent personality.
Ray's answer was more direct: The only task AI still cannot do is set its own goals. People need to do that.
All three answers pointed to the same thing: Value judgments.
AI can execute, analyze, and generate content. But people still have to judge what is right, wrong, good, or bad.
At least for now.
The Paradox of the Data Feedback Loop
Hua Rongqi asked a technical question: How do you train your data feedback loops? Do you use human-labeled material, or let AI generate material by talking to itself?
Sawyer said the core specifications and structure must be tuned by people to ensure that the emotional value of the output remains stable.
Bella said she primarily lets AI handle things she dislikes, such as reviewing data.
But Ray's answer stopped me for a moment.
He said he had gone through two stages. In the first, he thought he was very capable and gave OpenClaw instructions one, two, three—but it did not Deliver well. In the second, he changed his approach: Treat OpenClaw as a Stanford or Tsinghua graduate, and let it reflect for itself.
When something went wrong, he would ask OpenClaw: This did not go well. Would you like to reflect on it and come up with a better plan?
OpenClaw then worked through the entire process from beginning to end and even generated a mechanism for a Self-improving Agent on its own.
He had not thought of that approach before.
Ray said that as an individual, his information intake would always be limited. But behind OpenClaw was GPT-5, with far more training data than he had access to. If he handed over control and simply asked it to reflect, the plan it produced would inevitably surpass his.
At that point, I suddenly found it a little frightening.
If machines iterate on themselves faster than humans, what value do humans still have?
Competitive Barriers Are a Stupid Question
Hua Rongqi asked a question investors love and founders hate: Where is your barrier to entry? If this system can be standardized, it can also be copied.
Ray's answer was blunt: I have no moat at all. The question itself is stupid.
His reasoning was that in the AI era, large-model capabilities improve so quickly that they can catch up with the barrier you are worrying about in two or three months. The real barriers lie in three things:
First, mental inertia. You trust AI, while the stronger someone is technically, the more reluctant they are to use it, fearing that OpenClaw will steal data. That inertia is itself a barrier.
Second, speed. Someone asked: What if you spend time on OpenClaw and a new tool replaces it a couple of days later? He does not worry about that, because the Know-how accumulated in the process will not disappear.
Third, prior industry Know-how. His experience in overseas Social Media Marketing has been embedded in his OPC's Skill packages. No one can take that away from him.
Everyone can compete effectively by embedding their own Know-how.
Sawyer's answer was more emotional: The core is still trust. How do you get users to genuinely trust you? Through the depth of conversations, the frequency of touchpoints, and the accumulation of memories. Once all their memory data is in the product, users naturally will not migrate.
Bella said that doing social media well depends heavily on an individual's feel for the internet, which is difficult to standardize. Trends keep changing; a successful script lasts at most a month and a half before things move on. In the future, people will also be willing to trust a personal IP—a personal brand—built by an individual. The trust and imagined possibilities associated with that IP are extremely important.
At that point, I suddenly understood.
In the AI era, the real barrier to entry is not technology or data. It is people.
It is your taste, your aesthetic judgment, your values, and the trust you build with users. AI cannot copy those things.
When Will NPCs Replace OPCs?
Hua Rongqi asked an even tougher question: When will NPCs—no-person companies—replace OPCs? Musk says he wants to achieve the first commercial company controlled entirely by AI this year. What do you think?
Ray said it would happen soon, perhaps in 3–5 years. Human efficiency is too low; not having to deal with people is the happiest state of affairs. AI should govern everything. Humans could simply receive UBI—universal basic income—and be responsible only for providing emotional value.
That sounded somewhat extreme, but when I thought about it carefully, it did not seem entirely without logic.
Yifan said AI can iterate endlessly toward a goal, but setting the goal itself involves a value judgment. For now, AI cannot make that value judgment on its own.
Hua Rongqi pressed him: But if the goal is making money, why can't AI make that judgment? Set a reward function and make as many US dollars as possible.
Yifan said that the otome users they serve are sensitive to connections between people. If they discover that a connection is fake or purely about making money, they become very angry. They need real human value judgments and emotions.
Sawyer added something that stayed with me: If everything is handed to AI, the result might resemble “Ultron”—destruction. Humans harm the world, whether through material consumption or oil use. Human consumption of the world causes harm, but AI cannot make that value judgment. It might conclude that having no humans is best.
Because machines can disregard ethics entirely.
Final Thoughts
After the roundtable, I sat in the audience and thought for a long time.
Ray cut six employees and replaced them with AI. That was a victory for efficiency—and a sign of the era's cruelty.
Bella, working with two interns, generated exposure in the tens of millions. That was a victory for leverage—and an amplification of individual value.
Sawyer insists on building AI with a sense of being alive, emphasizing objectivity and subjectivity. That is a commitment to the human dimension—and something beyond the boundaries of technology.
I suddenly remembered what Ray had said: In the OpenClaw era, the people who use OpenClaw well are often those who are not particularly expert in the business and not particularly strong at coding.
Put another way: People with patchy expertise can use it better than experts.
Because those people carry less baggage and less of the arrogance of “I know better than you.” They are more willing to hand control to AI and acknowledge their limitations.
And the experts? Experts always think AI is not good enough, that it will make mistakes, that it does not understand as much as they do.
But what do you do when AI really does understand more than you?
That is a question each of us may need to consider.
More from the Conversation
Unique Awards · Hangzhou AI WEEK Trend Roundtable Panel: “Application OPCs: From Local Validation to Global Expansion for AI Products”
Guests: Yifan, Founder of Vaniloom | Bella Ren, Content Marketing Expert | Ray Luan, CEO of EZsite AI | Sawyer, Product Lead at Meowster
Moderator: Hua Rongqi, CEO of Beijing Chuangqi Technology
Hua Rongqi: For our first question, please briefly introduce your backgrounds. Use three sentences to tell us who you are or what you are doing, and add one final answer: How many people are in your company now? OK, let's start with Yifan.
Yifan: Hello, everyone. I'm Yifan, founder of Vaniloom. Our company mainly provides customized content for consumers and businesses. Our current consumer product focuses on content creation and customization for female users in North America and otome communities. We also apply our personalized content customization capabilities to content marketing and other areas. There are four people in the company now.
Hua Rongqi: Four, OK. I thought you were running an OPC, but there are four of you.
Ray Luan: I'm Ray. I previously oversaw used-car e-commerce at ByteDance. The product I'm building now helps companies with digital employees for overseas expansion. We had ten people last year, then cut six, and now have four left.
Hua Rongqi: You'll have to tell me more about how you cut six people. I'd like to hear that too.
Ray Luan: We had ten people last year. Of the six we cut, two were in UI, two in front-end development, and two in back-end development. We made the cuts because, after OpenClaw—the “lobster”—and Cursor came along, we felt we could replace people entirely with AI, so we took that to the limit. Just as ByteDance's culture emphasizes pursuing the ultimate, we now use AI to replace all the work of those people we cut. My latest digital-employee product for overseas expansion was built entirely by me using OpenClaw and Claude Code. The new brand is ClawLite AI. Essentially, I built it alone from scratch: customer acquisition, website building, lead conversion, and finally closing deals. No one else is needed.
Hua Rongqi: Tell us more about that shortly.
Bella Ren: I'm Bella. I should probably count as an OPC in the true sense now, because it is essentially me and two interns. What am I doing? I help many AI companies acquire customers through content on overseas social media, such as LinkedIn and TikTok. In my previous project, I took an AI education product overseas and generated exposure in the tens of millions on overseas social media in a single month. The overall paid conversion rate across the journey from the website to customer acquisition was nearly 5%. I've turned the Know-how and entire workflow from that experience into my current methodology.
Hua Rongqi: You're the OPC that most genuinely looks like an OPC this year. You have the fewest people.
Sawyer: Hello, everyone. I'm Sawyer, currently the product lead at Meowster. We mainly serve the Japanese market. Unlike mainstream App or Bot products, we want to build AI that feels alive. Why do I say alive? Because we emphasize objectivity and subjectivity. When you talk to this cat, it has its own thoughts and understanding, its own outlook on life and the world, and even its own values. Mainstream AI today may be very sweet or very miserable, but we retain that core objectivity and subjectivity. Our company currently has around 10 people.
Hua Rongqi: The first substantive question follows from your answers. I'm an OPC myself. Although I use a lot of AI to help, many things still have to be solved by people. You previously had interns to call on in large companies. Now that you're using AI to run an OPC, where do you feel the boundaries are? Which parts have you wanted to use AI for but found still required people? Let's start with Yifan.
Yifan: We did go through a period early on when we had a larger team. We found that keeping the organization at a sensible size early on would definitely be beneficial over the long term, because traditional large companies are often constrained by communication costs and organizational inertia. With AI assistants that can greatly extend what we want to do, AI can provide very substantial help at work as long as it understands the context. So we gradually reduced the number of interns early on. But we found that pre-sales work involving people, and building trust, still require personal involvement. In the AI era, we may need to emphasize connections between people, trust, and personal brands even more. Our consumer business mainly serves Niche groups, such as customized otome-oriented fan content. They have a strong sense of belonging to their communities, which makes it even more important for founders to build relationships and emotional connections with customers themselves. That is an important part of building niche products today.
Hua Rongqi: So it sounds as though AI handles the product side, but growth currently depends mainly on you, as the founder, getting involved personally and interacting with users.
Yifan: For consumer growth, if you want to build trust between your product and users, you must put care into what you do. Whether you are creating content, building a product, or contacting users, you cannot fob them off with AI that contains no emotional investment. The major social media platforms are gradually going to be flooded by an ocean of AI-generated content, so filtering content is something everyone wants. What we can do is make every signal we send valuable, making people feel seen and understood. That may give us a greater share of users' attention.
Ray Luan: We went through two stages. Last year, we used AI to replace programmers and cut four people at year-end. This year's biggest challenge has been Marketing. People used to do it; now social media marketing, on platforms such as Facebook, LinkedIn, and Twitter, essentially needs no people. I now hand 80% of my work entirely to AI and do only two things myself. First, set goals for my AI employees. The only task AI still cannot do is set its own goals; people need to do that. Second, Validate—humans need to confirm whether the work it completes meets the requirements. Apart from those two things, everything goes to AI employees.
Hua Rongqi: That is a very high level of automation. Most setups are still semi-automated, with many correction steps requiring human involvement. Have you already achieved automatic correction by AI?
Ray Luan: Initially, I did the corrections—QA—myself, but gradually I built an Agent to do QA for me. I think most companies use semi-automation because they do not trust AI. But I think AI performs better than I do in most scenarios. For coding, Cursor and Codex certainly write better code than I do. For generating Marketing materials or judging channels, I also think AI is stronger. Based on that assumption—that AI has already surpassed me across the board—I handed most of my work to AI. I've also found that the more technically knowledgeable people are, and the better they know the business, the less comfortable they are entrusting it to AI. Here's a provocative take: In the OpenClaw era, the people who use OpenClaw well are often those who are not particularly expert in the business and not particularly strong at coding.
Hua Rongqi: What you're describing is probably someone who understands the business but has only patchy technical expertise, without so much misplaced confidence, so they feel comfortable handing things over to OpenClaw.
Ray Luan: Yes. The stronger people are technically, the more reluctant they are to use it. They feel AI is unsafe.
Hua Rongqi: It's the same with text and images. When I write articles, I really do use “traditional handcraft.” That may be distrust carried over from the early GPT-3.5 era. But recently, using Claude 3.5 Agents or Gemini, they really have written better than I do. I can relate.
Ray Luan: Exactly. You are more professional than I am at writing articles, for example. But because I consider myself unprofessional, I would rather hand it entirely to AI. At the end, I make just one judgment: Does it meet basic values? I leave everything else alone.
Hua Rongqi: At present, many parts of the process still rely on more human labor to ensure delivery to customers. Bella, what do you think?
Bella Ren: I feel that what Ray just said is that if you have more industry Know-how and add AI, that combination is unbeatable. In my entire workflow, for example, my Taste is the decisive factor. I need to select the network of accounts aimed at users, give them personas, choose KOC creators who fit that Vibe, and provide Storytelling scripts for their content. In that process, I mainly use AI to amplify my capabilities and increase leverage. Once content is published, for instance, I cannot watch the scripts and traffic every day. I ask AI to review the data, analyze the key factors in scripts with strong traffic, and then make iterations and variations. I've been making content for about two years, so I understand better what performs well and have my own aesthetic judgment. My taste plus AI can dramatically amplify personal leverage in the OPC era.
Sawyer: It may be a little different for us. Early on, we handed things such as content generation, emotion assessment, and topic guidance to AI, and it worked well. But unlike a standardized customer-service SOP, human emotions vary widely. When someone is unhappy and venting to you, you cannot respond like customer service and say, “Oh, you'll be fine.” It requires judgment and guidance. We need to give the product many standardized rules: when to empathize and maximize emotional value rather than offer a solution, and when to offer a solution. We run many structured scripts, constraining the output to the scope of our capabilities to maintain personality consistency.
Hua Rongqi: From what I'm hearing, everyone needs alignment—RLHF—to help AI understand you better, with more Context to adjust it. In practical terms, do you use humans to label your material? We experimented with generating material through AI conversations and abstracting rules into the System Prompt, but found that the data easily became a self-contained feedback loop and did not work well. We eventually hired university students to help label it. What methods are you using now?
Sawyer: There is certainly some automation at the core, but a lot needs human optimization and adjustment. Things such as persona standards and structure, or System Prompt content like that used in OpenClaw, must be tuned by us manually. That is how we ensure stable emotional value in the output. Each user's profile is fairly complex, so the processing needs to be much more granular.
Hua Rongqi: Bella, do you Prompt customers, or Prompt your own workflow? How do you help your AI understand the business better and achieve the automatic tuning Ray described?
Bella Ren: I think it is still about the workflow: handing AI the things I dislike doing, such as reviewing data and helping me analyze scripts.
Hua Rongqi: Ray, do you have your own database or corpus for labeling the OpenClaw instances you send out? How do you adjust them?
Ray Luan: I've gone through two stages. Bella is more professional, so OpenClaw is leverage that amplifies her capabilities. For me, I am OpenClaw's tool. In the first stage, I thought I was very capable. I gave OpenClaw instructions one, two, three, but it did not Deliver well because it was following my approach. Recently, I had an insight: When something did not go well, I asked OpenClaw, “This did not go well. Would you like to reflect on it and come up with a better plan?” I treated OpenClaw as a Stanford or Tsinghua graduate and gave it more room to think. For example, I built a system using OpenClaw for overseas marketing. When I encountered a problem, I asked it to reflect. It worked through the entire process for me from beginning to end, and even generated a Self-improving Agent mechanism on its own. I had not thought of that approach before.
Hua Rongqi: In abstract terms, you think OpenClaw improving itself is better and faster than humans improving a Skill.
Ray Luan: My own feeling is that it does much better than I do. As an individual, my information intake will always be limited. But behind OpenClaw is Codex or GPT-5, with far more training data than I have access to. If I hand it control and simply ask it to reflect, the plan it produces will inevitably surpass mine.
Hua Rongqi: It feels as though we've reached a tipping point where machines iterate on themselves faster than humans. Yifan, how do you train your data feedback loops?
Yifan: In an automated workflow, when an Agent is not performing to the required standard, whether you use Prompt Engineering or build MCP—Model Context Protocol—for a specific scenario, you are effectively building “scaffolding” for a digital employee. It is like onboarding a new employee: First, give them the company's context—what we are doing—then provide tools and tell them where to find user data. It comes down to two things: First, sharing context; second, providing access to specific tools or data. We build basic data-access infrastructure specifically for an Agent—for example, packaging clean platform data as MCP or Skills—and hand it to the digital employee. Building this AI infrastructure may become an important source of competitiveness later on.
Hua Rongqi: Since everyone has mentioned AI feedback loops, here is a question investors love and founders hate: If this system can be standardized, it can be copied. In the content industry, where do you think your competitive barriers really lie? Personally, I'm rather pessimistic. Online information has already crowded out attention completely; perhaps only returning offline can rekindle it. How do you respond to such low barriers to entry? Ray first.
Ray Luan: When investors ask about a moat, I say I have no moat at all. The question itself is stupid. In the AI era, large-model capabilities improve extremely quickly and can catch up with the barrier you are worrying about in two or three months. My barriers are: First, mental inertia. I trust AI, while technically stronger people are more reluctant to use it, fearing that OpenClaw will steal data. That inertia is itself a barrier. Second, speed. Someone asks: What if you spend time on OpenClaw and a new tool replaces it a couple of days later? I do not worry about that, because the Know-how accumulated during the process will not disappear. Third, my prior Know-how in overseas Social Media Marketing—such as operating private communities on Facebook—is embedded in my OPC's Skill packages. No one can take that away. Everyone can compete effectively by embedding their own Know-how.
Hua Rongqi: It is a bit like a perfectly competitive market. Making US$1 million a year from a small business is enough; defense matters more than offense.
Ray Luan: The entire Agent market is a long-tail market, and everyone has their own niche. Let the big companies do the cost-reduction work. Small and medium-sized companies like ours just need to embed what we are good at in digital employees.
Hua Rongqi: Sawyer, if large companies—for example, Character.ai, which Google acquired—keep investing heavily in emotional companionship, where is your barrier?
Sawyer: The core is still trust. As people get older, trusting something becomes extremely costly. How do you get users to genuinely trust you? The key points are the depth of conversations, the frequency of touchpoints, and the accumulation of memories. Give users an Aha moment that feels extremely emotionally valuable, gradually increase frequency and stickiness, and ultimately achieve a very high level of trust. That is the most important barrier. Once all the memory data remains in the product, users naturally will not migrate unless a competitor can forcibly transfer the Memory out.
Hua Rongqi: Bella, what if a large company teams up with leading creators, such as MrBeast, to do what you do?
Bella Ren: Doing social media well depends heavily on an individual's feel for the internet, and that is difficult to standardize. Trends keep changing. My own feeling is that a successful script lasts at most a month and a half before things change. Even if MrBeast did it, if everyone used the same scripts, that traffic would soon lose its value. In the future, people will also be willing to trust a personal IP—a personal brand—built by an individual. The trust and imagined possibilities associated with that IP are extremely important.
Hua Rongqi: An OPC has few people. If it wants to win mindshare and trust, how can it use AI to do what traditional large companies do?
Bella Ren: This matters even more online, because content published by one person can reach many people. While running an OPC, you need to build your own personal IP, express your values on social media, and help more people understand you.
Hua Rongqi: What do you think, Yifan?
Yifan: I think the OPC is a very temporary organizational form. It simply gives some people with an information advantage the capabilities of a traditional team. Over the long term, we should think more about how to build an entirely new organizational structure once we take the OPC for granted as a model of individual productivity.
Hua Rongqi: Let's turn to something lighter. Most people here were born in the 2000s, or serve people born in the 2000s. Yifan, what consumption insights or preferences have you observed among women aged 18–30 in the otome segment?
Yifan: At a macro level, it feels as though most money in the AI industry now comes from investors. Many companies serving businesses have not truly achieved commercial profitability by creating value. So our consumer approach is to find a relatively Niche segment. Traditional content industries have a pronounced winner-takes-most effect: Everyone wants to spend their time on the best content, leaving niche subcultures with very low ROI. But AI, as a tool for reducing costs and improving efficiency, supports mass customization of personalized content. With that cost advantage and a Memory system based on ChatGPT, we have found it can provide highly valuable content tailored to individuals. This could bring structural change to the content sector. So an AI startup must amplify what makes it distinctive and serve a specific niche. That is what is most valuable.
Hua Rongqi: Sawyer, do you also have many young female users in Japan? Any new observations?
Sawyer: Yes, women make up 60% to 70%. Japan is fertile ground because of its IP and blind-box culture. But underlying psychological pressure is high, and people are very afraid of troubling others. Our Slogan is “We don't mind being bothered,” because we want more interaction. Young people in Japan greatly need that timely emotional feedback. We need to fill the gap in emotional value and lower their defenses. Once Japanese users approve of your product, their subsequent willingness to pay and stickiness are both very high.
Hua Rongqi: That points to Japan's repressive commercial society and a need for an emotional outlet. Bella, North America is a more outgoing and direct market. What interesting observations have you made when marketing to young people there?
Bella Ren: The core is still identifying the pain points that matter most to them. Our product spread so widely because we concentrated on university students' frustrations, such as “homework is painful” and “I can't understand what's being taught.” We packaged them as stories, such as “Don't know how to make a PPT? Use this product.” A simple video like that could reach 400,000 organic views because so many students had that need. Working professionals or women may instead have a greater need for emotional value.
Hua Rongqi: Are contextualized scenarios and viral Meme formats more likely to spread among young people?
Bella Ren: Young people like interesting things. My Hook, for example, might be a funny scenario such as “a student frying a steak in class.” When the target audience encounters a pain point such as rushing to finish homework before a Deadline, everyone relates and shares it with classmates, making it easy to go Viral.
Hua Rongqi: Finally, Ray, your customers used to be more mature. What have you observed recently about younger audiences?
Ray Luan: My customers are a little older, between 25–40. But recently, I've had a realization: I'm not optimistic about building productivity tools in the AI era. As large models develop, all productivity tools—including my own—will become worthless. After AGI is achieved, humans themselves will have no value in terms of efficiency. Future AI should involve communication between one Agent and another. Ultimately, the only thing humans will be able to provide is emotional value. The biggest difference between people and AI is that humans have emotions, while AI is extremely rational. So the next product I'm planning is a virtual boyfriend for young women, built with OpenClaw.
Hua Rongqi: When will NPCs—No-Person Company organizations—replace OPCs? Musk says he wants to achieve the first commercial company controlled entirely by AI this year. What do you think?
Ray Luan: I think it will happen soon, perhaps in 3–5 years. Human efficiency is too low. I think not having to deal with people is the happiest state of affairs, and everything should be governed by AI. Humans can simply receive UBI—Universal Basic Income—and be responsible only for providing emotional value.
Hua Rongqi: Yifan, do you think NPCs will put you out of business?
Yifan: AI can iterate endlessly toward a goal, but setting that goal itself involves a value judgment. For now, AI cannot make that value judgment on its own.
Hua Rongqi: But if the goal is making money, for example, why can't AI make that judgment? Set a reward function and make as many US dollars as possible.
Yifan: The otome users we serve are sensitive to connections between people. Once they discover that a connection is fake or purely about making money, they become very angry. They need real human value judgments and emotions.
Hua Rongqi: Bella, doesn't the independent content business seem easy for NPCs to replace? Just let AI take orders, publish content, build accounts, and earn money.
Bella Ren: It can indeed help you make money through channels more quickly. But I've kept thinking about how some accounts show a person's face and others do not. I feel that accounts featuring real people showing their faces will actually become more valuable.
Hua Rongqi: The value of real people is eternal. Finally, Sawyer, will humans ultimately be taken over by these “cats” that provide extremely high emotional value?
Sawyer: No, I don't think so. But in the future, this kind of lifelike AI really could provide extremely high emotional value. It could exceed anything you imagine and give you greater satisfaction than many material things, whether through dopamine release or endorphins. Yes, but there is another point: What we mentioned earlier brought a useful example to mind. A major issue at the heart of AI is its uncontrollability. As we said earlier, if everything is handed to AI, the result might resemble “Ultron”—destruction. Humans harm the world, whether through material consumption or oil use. Human consumption of the world causes harm, but AI cannot make that value judgment. It might think that “having no humans is best.” Machines can disregard ethics entirely. They really could replace humans completely, but the risk is total uncontrollability. I think people have concerns about this in relation to many of the humanoid robots being developed now as well.
Hua Rongqi: Although we inevitably arrived at a discussion of values, overall I think that in the current competition between OPCs and large companies, everyone here can use relatively flexible approaches. At the very least, look after your own little patch. At least we should earn more than before we left our jobs; I believe we can do that. Of course, I'm not encouraging anyone to quit. I wish your companies continued success.