---
title: "Stop Calling Yourself a Boss—A CEO Is Just a \"Workhorse\"?"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-04-28T10:01:28+00:00"
canonical: "https://ffcap.cn/en/research/src-20260428-01html"
source: "https://uniqueresearch.substack.com/p/src-20260428-01html"
language: "en"
---

# Stop Calling Yourself a Boss—A CEO Is Just a "Workhorse"?

_Original · Unique Research · 2026-04-28_

_Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening essay and full panel, including all named speaking turns and their continuation paragraphs. Revenue, conversion, user, efficiency, customer, financing and market figures are source or speaker claims, not independently audited findings. The "牛马" (workhorse / beast of burden) and "老登" (old-timer / boomer) terms are colloquial Chinese slang; their English renderings preserve the speaker's tone without softening it. Statements about model capability levels (P3/P5/P8/P9) use the source's internal grading analogy, not an industry-standard certification. Company, personal and work titles are transliterated where official English forms remain unverified. The source is dated April 28, 2026._

Unique Awards

"Stop Calling Yourself a Boss—A CEO Is Just a 'Workhorse.'"

When software is no longer designed for humans, people not only cease to be the users—they become the "bottleneck" holding everything back.

"

Last year AI was still at P3 level; today it has reached P5.

If you still hire people to manage it, and next year the model reaches P8, can your company still afford to hire P9 people?

"Stop calling yourself a boss—A CEO is just a workhorse." This was a blunt truth dropped by Song Jian, founder of NoDesk AI, at a forum. When his post-2000s employees show up at 2 p.m. every day, he, as the boss, has instead become the person who "kowtows" most often in the company.

Why? Because in his company, all the work is done by AI Agents, while human employees are only responsible for giving AI "unlimited Tokens."

Even more cutting was the prediction from Jingling, CEO of CometAPI: "Last year AI was still at P3 level; today it has reached P5. If you still hire people to manage it, and next year the model reaches P8, can your company still afford to hire P9 people?"

At this roundtable, four entrepreneurs on the front lines of AI (Jingling, CEO of CometAPI; Song Jian, founder of NoDesk AI; Xu Jia, VP of Product at Pollyreach.ai; and Alex Li, founder of AllyClaw) tore open a truth that sends chills down every worker and boss's spine:

Over the past decade-plus, the underlying logic of software design never changed—everything was built for human users. But today, that script has been completely rewritten. The newest software is abandoning humans. People not only cease to be the users; they become the "bottleneck" holding everything back.

The Interface Is Disappearing, and Humans Have Become the Biggest "Bottleneck"

We always assume humans are mastering the tools, but in real business scenarios, human reaction speed and processing capacity have long been holding things back.

"Humans have become the bottleneck," said Alex Li, founder of AllyClaw, bluntly. In the past, when independent e-commerce sites did email marketing, human operators could only do simple group blasts by region and gender, at most achieving a crude "thousand people, thousand faces."

Now? AI doesn't need people to tag anything at all. It can clearly tell which pain point a consumer cares about, and can even generate emails in real time based on purchase intent for one-on-one communication. "No human management is needed; things can be fully automated." Alex shared a striking figure: previously, sending 1,000 emails yielded only one paying customer; now it might yield five—a direct fourfold increase in conversion rate.

In overseas phone-call scenarios, the situation is similar. "When the other side's human customer service can't keep up, they become the bottleneck of time," said Xu Jia, VP of Product at Pollyreach.ai. Especially when handling utilities (water, electricity, gas) or when overseas students cross language barriers, AI has innate concurrency capabilities and operates across time zones. As long as humans exist, phone communication will exist—and in these frustrating waits and exchanges, AI is clearly much more useful than people.

When the interface shifts from a human-visible UI to machine-to-machine API calls, how are we supposed to position ourselves?

Power Devolution: Stop Seeing Yourself as a Boss

If all the execution work is done by machines, what are the people in the company doing?

Song Jian, founder of NoDesk AI, answered with extreme candor, even a touch of pathos: "A CEO is just a 'workhorse,' just a role of an employee. Don't have any trace of 'old-timer' flavor or boss posture."

This is absolutely not a joke. When the abilities to generate images, videos, and write code are thoroughly "democratized" by large models, skill barriers are broken. "Skills are democratized—why should anyone work for your ideal?" Song Jian's words are worth every manager's deep thought. He admitted he is now the person who "kowtows" most diligently inside the company. Facing post-2000s colleagues who sleep at 3 a.m. and come to work at 2 p.m., his strategy is: "Give them unlimited Tokens within the team, and give them a fair, objective, and free environment."

Jingling of CometAPI deeply agrees. As a company providing Token compute power, their team defaults to "unlimited Tokens." They promote an "AI subordinate mindset"—every frontline worker is essentially a manager; it's just that they used to manage people, and now they manage various AIs and Agents.

This means organizational form is undergoing dramatic change. Alex Li of AllyClaw said that in the past he still had to tell his team what features to build; now he can get hands-on himself and build a prototype in about an hour. In this new phase, he doesn't need purely passive-response R&D engineers. He needs people who can stand on their own, have engineering thinking, and unexpectedly pull out a new tool saying, "You guys should try this together."

Human value has shifted from "possessing specific skills" to "possessing orchestration and decision-making capabilities."

The Ideal 1:9 Ratio: Humans Only Do the Most Critical 10%

In the future division of labor, how far do we actually need to go?

Alex shared a highly illuminating phenomenon: they opened read-only database access directly to AI, and found that users no longer look at reports on the interface—they directly let AI connect to the database and ask questions. "80%–90% of the questions are long-tail needs we previously didn't know about and hadn't solved."

This reveals an ultimate state: humans may only need to use 10% of their energy. In complex e-commerce operations, humans only need to give the highest-level instruction: "one—approve, two—reject, three—fine-tune." AI understands hundreds of metrics, finds the Root Cause, and humans only handle the final call.

"Humans make decisions; AI executes," summarized Xu Jia. We need to define what we want, what we don't do, when to do it, and to what degree. Currently, human-AI collaboration is only at about a 6 out of 10. On one hand, engineering technology still needs refinement—how to put a "saddle" on AI so it doesn't run out of control. On the other hand, there is psychological distrust from humans.

We don't dare let go completely, which is normal. But the trend cannot be stopped.

Surrounded by Hidden Bombs, or Making Peace with Yourself?

Toward the end, the conversation inevitably turned to fear. This fear doesn't come from the sci-fi narrative of "AI destroying the world," but from the most real体感 in daily work.

The most direct fear is loss of control. Alex admitted that his code-review speed can no longer keep up with the speed at which his colleagues write code with AI. When you cede decision-making power to AI, what if it takes it upon itself to send a "50% discount" email to a customer? Who bears that loss?

Another fear is concern about the degradation of one's own abilities. Xu Jia put it vividly: "If Claude Code goes down for an hour, I feel terrible and can't get anything done. I worry that if large models disappear, will my English proficiency and thinking ability decline?" It's easy to go from frugality to luxury, but hard to go back.

Jingling posed a soul-searching question: "Last year large models were P3, today they're P5, and we hire P6 and P7 people to manage them. Next year models reach P8—can we even afford P9 people?" When model capability surpasses our ability to discern, do we still have judgment?

The answer may lie in Song Jian's words. He shared his journey from grinding himself to death to finally letting go: "Even if you eat all the books, you can't keep up with new version releases. Finding balance—neither degenerating into a chimpanzee nor failing to make peace with yourself—is quite important."

Dinosaurs lived for 100 million years; humans have only had 5,000. Nobody can say for certain what the future holds for silicon-based and carbon-based life. In this new world where software no longer needs to be designed for humans, what we need is to relax a little, keep learning, and then bravely take our hands off the keyboard and tell our AI assistant:

"Go do the work. Call me if something comes up."

More Conversation Details

Hong Kong · Global Unique Awards Trends Roundtable Panel

Theme: Product Revolution: When Software Is No Longer Designed for Humans

Guests: CometAPI CEO — Jingling; NoDesk AI Founder & CEO — Song Jian; Pollyreach.ai VP of Product — Xu Jia; AllyClaw Founder — Alex Li

Moderator: Unique Capital Partner — Wang Chaochao

Wang Chaochao: What does it look like when software is no longer designed for humans? Over the past decade-plus, the underlying logic of software design never changed—screens, clicks, forms, workflows—everything was designed for human users. But AI Agents in this era have completely rewritten that script. Software is shifting from a tool controlled by humans to an autonomously acting intelligent agent—from API aggregation to e-commerce Agents, to intelligent customer service, to voice automated marketing. So our guests are personally experiencing this transformation in their respective fields. Today we won't talk about vague concepts; we'll discuss four practical questions: pain points, division of labor, collaboration, and challenges. Please introduce yourselves and what your companies are doing.

Jingling: Hello everyone, we are CometAPI, aggregating all known large-model APIs globally. This includes the latest large model OP4.7 just released the day before yesterday—we'll be launching it soon—as well as Doubao, ByteDance's latest video model, and GPT, Gemini, and others. Users can call almost all the APIs they need through us in one stop. Our platform mainly serves international clients and the global market, and we also serve many well-known AI applications on the a16z list. Behind all of these are our programmers; we do a lot of operations and optimization work to ensure production stability. That's roughly it.

Song Jian: Hello everyone, we are NoDesk AI. We specialize in Agent applications for the e-commerce industry. We were founded relatively recently, in March 2025, so it's been just over a year now. But in this past year, we've witnessed the continuous evolution and maturation of large models, along with the maturing of Agent capabilities. So we firmly believe that e-commerce moving toward unmanned operation is just around the corner. With the arrival of unmanned e-commerce, it means all the human positions and colleagues remaining in human companies—their capabilities, profiles, and how they collaborate with Agent capabilities, profiles, and needs—will all need to undergo transformative change. I estimate this is a problem that our generation of new startups, and current white-label and major brands, or brands that will grow in the future, will all face. At least I feel that what our company wants to do and is doing aligns very well with today's theme.

Xu Jia: Hello everyone, I'm Xu Jia, product lead for Pollyreach.ai, an innovative business of QuickCEP. QuickCEP provides intelligent customer service for our overseas brands. The new Pollyreach.ai is a telephony infrastructure we built for AI Agents. We provide global voice-calling capabilities, giving users a real phone number and letting AI make and receive calls for you, with voice dialogue capabilities that can help you handle things in life or business. Our positioning is a C-end personal assistant that can help with scenarios like restaurant reservations in Japan, AI receptionists, batch interviews, and business call answering or customer acquisition.

Alex Li: Hello everyone, I'm Alex. Our AllyClaw is also an AI product focused on e-commerce. I think our biggest difference is that we want to make AI products that help merchants make money. For example, we have an AI data analyst and an AI email marketing master that can actually help e-commerce brands earn about 10% more revenue. Why can we do this? One important point is that we may be one of the few products in independent-site e-commerce like Shopify that has the complete behavioral path and context for every user. We can know which Google keyword a consumer searched for when they came to the store, or which Meta ad they saw. These very important intent data provide the contextual foundation for us to achieve "thousand people, ten faces" marketing with AI. We can make AI deeply understand consumer intent, thereby delivering AI solutions that truly help merchants convert, with quantifiable results. That's our product's current situation.

Wang Chaochao: Three of our four guests are helping merchants make money in e-commerce scenarios, and one is in APIs, but overall they're all B2B. Our first question: in your respective product and service scenarios, which links no longer require human work, or no longer require human intervention—instead, humans have become the bottleneck? Please share a real case from your scenario, starting with Alex.

Alex Li: Building on what I just said, let me introduce how a very old product is being transformed by AI—email marketing. Email has been a primary communication method for over a decade. Of the US$1 trillion GMV in global independent-site e-commerce, email generates roughly US$100 billion in revenue—a very large number. But in the past, the approach was to group users by region and gender, send different product news to different groups, or send coupons to users who abandoned their carts. In summary, at most you could achieve "thousand people, ten faces." Humans also had to constantly maintain various tags, tag consumers, create groups, and design email templates. Today none of this is needed anymore. Why? Because AI can very easily figure out what the consumer's intent is—even which pain point you care about when buying a certain product, AI knows very clearly. It can generate emails nearly in real time based on the consumer's purchase intent, communicating with you one-on-one, making you feel like a sales consultant is communicating based on your needs. What's the result? First, humans have become the bottleneck, because no human management is needed and things can be fully automated. Second, after you send the email, open rates, click rates, whether consumers placed orders—this forms a data loop. You just need to tell AI the goal is to convert users, and it will continuously analyze for you, finding the user's pain point: is it a price problem, unclear product requirements, a certain function or material not communicated clearly, or does the user lack trust and need qualification proof to break the ice? It will understand the need and do personalized communication for you. The result is that no people are needed, and conversion rates may increase three to four times. Previously, sending 1,000 emails got one consumer to buy; today, sending 1,000 emails might get five consumers to buy—a direct fourfold increase.

Wang Chaochao: Alex described another scenario. A lot of innovation in the AI application layer now talks about "intent alignment"—making products understand the operator's intent. But what Alex is saying is that e-commerce service staff don't need to directly align; instead, they align directly with the end consumer's intent. This has elevated things to another level.

Xu Jia: Regarding phone calls—in China, where infrastructure is relatively complete, people don't feel it strongly, and may even think that communication between Agents doesn't need phones, that it can be solved through APIs. But the reality is that overseas phone calls have a high presence. In places like the US and Japan, phone communication is still very frequent. Even as our B-end intelligent voice customer service grows, as long as people exist, phones will definitely exist. Let me give some specific business scenarios: for example, handling utilities like water, electricity, and gas overseas—you must call a certain customer service number, and in extreme cases you wait in line for an hour and a half, and you must get through to get things done. At that point, the other side's human customer service can't keep up and becomes the bottleneck of time. Then there are language barriers. For example, international students going abroad may have psychological barriers or not know technical terms and procedures, can't describe the problem clearly, and can't understand the other party's reply—things may not get done, and they might even give up on sending a package. Or when traveling on business with time differences, how do you answer important calls? Addressing these human limitations and telephony infrastructure barriers, we believe AI has inherently better language capabilities, batch concurrency, and the ability to operate across time zones, helping you handle these things and further liberating human life.

Wang Chaochao: So in the phone-call scenario, humans are not as good as AI. I should note that email marketing and phone calls both have a big characteristic: they focus on overseas. Because in mainland China or Hong Kong, the overall infrastructure, marketing methods, and interaction methods are quite different from overseas. Mr. Song.

Song Jian: I may have a slightly different perspective. Although what our company is doing assumes that e-commerce's future trend will move toward unmanned operation, there's actually a premise: factories have gradually shifted from large-scale human factories to smart factories—what hundreds of thousands of workers did is ultimately done by a large number of machines, which is also the "Super Lighthouse Factory" promoted by the Chinese government. I also firmly believe that e-commerce will inevitably shift from customer acquisition, traffic driving, marketing, product listing/delisting, shopping guide customer service, and private-domain—links that require people—gradually, as AI capabilities improve, from large-scale labor-intensive to fewer people, or even truly unmanned. But what I want to express more is: "unmanned" doesn't mean the company has no living people; it means that all on-the-job positions and colleagues—their capability tech stacks and cognitive boundaries—may undergo structural change. For existing companies, whether they're white-label merchants, brand merchants, or suppliers like us who want to use Agent capabilities to serve clients. What everyone should discuss today is not "what positions will be replaced by AI," but rather the reverse: "how should the colleagues still in these positions be respected and seen?" Take the current exhibition venue as an example—we see many booths outside, and every booth has staff, and these staff are living people. They might be ordinary employees, executives, or founders. Although AI keeps advancing, if I hadn't personally flown from Hangzhou to Hong Kong to attend this conference, if everyone hadn't taken the subway and bus to get here, I wouldn't be able to clearly articulate what my company's philosophy is, and you wouldn't be able to feel whether this company is reliable through my on-site expression. So my current feeling is that AI and humans are not opposites; there is no replacement or being-replaced relationship between AI and humans. Even if e-commerce truly moves toward unmanned operation, there will still be experts in every position. And this expert simply used to operate SaaS software, but now will evolve to build Agent products, or call the model capabilities behind Agents. I feel that since software is no longer designed for humans (software is for Agents), Agents still need to be built by the people here. That's my view.

Wang Chaochao: In some links, humans are not as good as machines, but humans must always be preserved.

Jingling: We're in APIs, and I've also done e-commerce for many years. Among the e-commerce clients we serve now, unmanned operation is indeed being推进. We are APIs ourselves, dealing with machines—you see there are customer service Agents, marketing Agents, behind which they call the capabilities of major models through our platform. But I also strongly agree with what Mr. Song just said—it's not completely unmanned, it should be called fewer people. The talent structure is adjusting, unlike before when customer service was large-scale with many tiers (e-commerce customer service director, below them team leads reviewing conversion rates daily). Gradually AI may代理, first solving after-sales issues, then soon advancing to pre-sales. This is the continuous evolution of models, but ultimately people are still needed—people are needed to continuously iterate scripts, and some high-end B2B links and key account Sales will definitely still need people to solve. Including the marketing link, companies doing AI marketing build various Workflow processes, unlike before when they needed large numbers of designers and editors to cut videos and images—now with tools like Canva and Sora, one person can produce excellent material. But ultimately this person is still needed, and the requirements are higher than before: they need engineering capabilities, and before AI appeared they were already very professional in that field, possibly at Alibaba P7 or P8 level. At the same time they need AI-first thinking, willing to embrace AI iteration. AI is definitely still evolving; we must recognize that today's AI may not be as good as you, but in one or two months AI will get stronger and stronger. Look, the latest OP4.7 was released the day before yesterday, and yesterday there was a product like Claude Code. Once released, it will replace a large number of applications, and even basic Agent applications themselves will be replaced by more advanced Agents. I think it should be fewer people, but the engineering capability requirements for people will get higher and higher—that's the future trend.

Wang Chaochao: Humans are not as good as machines at some jobs, but in the human-machine food chain, humans are still at the very top. To ensure humans have a sense of presence and value within it, some people must not become bottlenecks. Now the second question—let's be sharper. Humans no longer need to use software, but instead tell AI what the goal is, shifting from operators to goal-setters, letting humans stand at the top of the food chain. How is this transformation progressing in your teams? What changes have there been in your team structure and talent needs? Starting with Alex.

Alex Li: The biggest change should be using myself as an example. In the past I might tell someone to build this feature; now I directly get hands-on myself and can roughly build a functional prototype in about an hour. In this situation, from my perspective, what kind of people are needed? People who can take responsibility, independently make business decisions, become the business Owner and be accountable for results. Ideally they have capabilities from understanding products to Marketing, preferably with a computer science background or engineering thinking. This kind of person often has a characteristic I particularly like: they'll unexpectedly tell me, "This thing happened twice repeatedly, so I made it into a tool, and now it's on GitHub—you guys can use it together." That's the person I want. What kind of person would I fire? This kind of person has mostly been eliminated already—the classic "pure R&D engineer" role who only knows how to passively respond to requirements, implement requirements into code, while complaining "I think AI has many problems and might not be able to handle it." The more people say this kind of thing, the sooner they'll leave my team. This kind of person no longer really exists in our team.

Wang Chaochao: In the past, bosses liked employees who shared methodologies and experience; now they directly squeeze skills. But how much can skills be squeezed? Subsequent requirements for people may get higher and higher.

Xu Jia: The changes in our team can be summarized as: boundaries are becoming blurred, and everyone is accountable for roles and results. Our internal R&D team no longer talks about front-end engineers or back-end engineers, and everyone no longer manually writes code like before—everyone, whether product or tech, uses AI. Everyone's responsibilities become owning a module or a certain function, and more time is spent aligning on direction and strategy, and on post-launch performance feedback. We rarely talk about execution details anymore. So when we plan to hire new people now, we define them by a needed "role type," and we won't hire multiple people. As long as one person can be responsible for it, the subsequent work we'll find ways to complete using various AI tools and building workflows. That's a major change in how we collaborate now.

Wang Chaochao: To briefly summarize: no distinction between front-end and back-end; second, aligning toward goals, strategy, and direction—this used to be what bosses or Leaders did, now it's delegated to people who can stand at the top of the human-machine food chain. The boss may be relatively more relaxed.

Song Jian: Our changes may be similar to most companies founded now (e.g., in 2025, 2026). Because we were founded in March 2025. The biggest change is myself. Our Founders all came from large-model companies, and we're considered relatively early in China to have awareness of "how to use models to make Agents, use models to build teams, use models to落地 business scenarios." I'm fully aware that in the AI era, a CEO cannot be the dumbest, most brain-dead person. In plain terms: stop calling yourself a boss—A CEO is just a "workhorse," just an employee role. Don't have any trace of "old-timer" flavor or boss posture. AI's capabilities have become stronger—image generation, video, code writing, all aspects have become stronger—which means everyone's skills have been "democratized." Skills are democratized—why should anyone work for your ideal? If you're high and mighty, nitpicking from a God's-eye view, you'll get beaten by society, beaten by your colleagues. I'm now the person who "kowtows" most diligently internally. Almost everything I do uses the "begging" approach: "I beg you, can you do this thing, spend a little more time thinking, you can go chat with OP4.7." I only hope young people of the new era can teach me more. I have some post-2000s colleagues—none come before 2 p.m., and none sleep before 3 a.m. The problem is I'm older and have a regular schedule, and the consequence is I haven't seen some colleagues for over a month. If you ask how to change in the AI era, it ties back to this theme: software is no longer designed for humans, it's for Agents, but for the "human colleagues" who remain, you must give them more respect. It's okay if they learn a bit slower—think about it this way: three years from now, hiring a human colleague will be very difficult. Cherish the human colleagues still willing to work with you right now. Give them unlimited Tokens within the team, give them a fair, objective, and free environment, let humans and AI jointly create more of the future—this is the problem this generation of innovation must face.

Wang Chaochao: So respect the brilliance of human nature, and also respect the weaknesses of human nature. With unlimited Tokens, many things behind can be solved.

Song Jian: One more thing—stop seeing yourself as a boss. Bosses are cheap now; capabilities are democratized, and if an employee is displeased, they'll fire you tomorrow.

Jingling: Speaking of unlimited Tokens—we're in Token compute power, so our team defaults to unlimited Tokens, we can't even use them all. Internally we have an "AI Native" culture with three modules: First is AI-first—everyone who joins must first have AI-first thinking, always wondering if AI can solve the current link's problem. Second is "AI subordinate mindset"—every frontline worker is also a manager. They used to manage people; now they manage various AIs and even Agents. Engineering capability is essentially also a kind of orchestration and management capability. How do you train Claude Code? Prevent it from slacking, install various capabilities on it. Third is "AI expert mindset"—everyone must become an expert in their own field and share within the team. We're now pushing everyone to use Claude Code. We believe that from Chat UI, to Dify workflows, to now general AI Agents, it's no longer limited to vertical fields—as long as used properly, it can handle almost any field. Our体感 has changed a lot; software is no longer for humans to look at. Just two weeks ago, we launched a CLI (command-line interface), entirely for machines to read. Many users no longer visit our website—they're all using AI Agents running CLI to call our platform and query data. This is an obvious trend of the future era—everything gradually becomes truly for machines.

Wang Chaochao: To briefly summarize: from the customer interaction side it has already become invisible, directly back-end through Agents calling APIs, and specific services are already running on the CPU. Another point is that bosses need to squeeze the team's skills to deliver results. The four guests are at different stages but all moving toward "team democratization" and "Native-ization." Let's discuss the third question: building an ideal model—in your imagination, how should humans and AI divide labor? Who does what? What score have you reached now? Where's the gap—is it inadequate technology, organizational non-cooperation, or customers not paying? Starting with Alex, referring to the products or services you provide.

Alex Li: A large part of our functionality is data analysis. Data used to be for e-commerce operators or bosses to view—watching reports, looking at ROI. This year we directly opened up read-only database access, letting AI directly read our database, which was very uncommon before. Why did we do this? We found that more and more users no longer look at reports on the interface; they directly let AI connect to the database and ask questions. Looking at these questions in our backend, we found that 80%–90% are long-tail needs we previously didn't know about and hadn't solved. But through AI writing its own SQL queries, the answers were found. This shows that the user's interaction form has changed. Software can only ever satisfy partially overlapping needs; long-tail customized needs can now be realized through AI. For the future, I think humans may only need to use 10% of their energy. For example, OpenClaw crawls whole-site data weekly, helps analyze material funnels and audience ad budget issues, and draws conclusions. Humans only need to tell it: "one—approve, two—reject, three—fine-tune." The hundreds of complex metrics of an e-commerce store—AI will understand, correlate, find the Root Cause, and give conclusions; humans only do judgment and confirmation.

Wang Chaochao: Humans only care about KPIs; other execution Agents can handle it. Human-AI collaboration at 1:9—that's an ideal ratio for the future. Mr. Xu. How should humans and AI divide labor in the future? What level is it at now, and where's the gap?

Xu Jia: The ideal human-AI collaboration should be: humans make decisions, AI executes. Specifically, how to make decisions? It's defining what we want, what we don't do, when to do it, and to what degree. What score have we reached now? Whether internal R&D or product, it's about a 6 out of 10. Where's the gap? I think there are two points: one is people—psychologically they're not yet willing to let go. Training AI has a time cost, and at first it won't perform perfectly; not everyone is so actively embracing it. The other is technology. It's not that the model isn't smart, but rather specific "Know-How" or vertical fields (like the engineering details in AI phone calls)—how to put a good "saddle" on the model so it performs well without going out of control? This requires continuous engineering effort. As trust increases, the gap will close. There's a US statistic showing that 60% of people used AI within six months, but the proportion that actually lets AI complete tasks like bookings end-to-end is extremely small. The general public currently doesn't have sufficient trust in AI autonomously executing tasks.

Wang Chaochao: In the future, humans set goals and machines execute. Now it's around 6 out of 10; the core reasons are psychological defensiveness and engineering capability still needing improvement.

Song Jian: On the fifth day of the Lunar New Year in 2025 (the business license bureau only opens on the eighth day), we thought of our company name: "Hangzhou No Desk Artificial Intelligence Technology Co., Ltd." The partnership is called "No Chairs," the investment entity is called "No Contact," and our branch in Yuhang is called "No Code." This shows the government is quite tolerant of young people. On the evening of the fifth day, we came up with a vision: "humans and AI jointly operate." Regarding human-AI collaboration, I only have one insight: first, respect human nature. AI has no consciousness today and doesn't need consideration for emotions, but at least for the next three to five years, Human-in-the-loop will definitely exist. The ideal situation is definitely that machines do machine-like work, and don't make humans act like machines. E-commerce moving toward unmanned operation doesn't mean experts disappear; it means liberating people to do more valuable thinking. I also agree with what people are saying now—don't be too anxious. Models iterate too fast; you just signed an expensive model, cut the cake, and that evening a better "Happy Pony" or new version is released. So make peace with yourself. The day before yesterday I met a CIO of a major brand who shared seven points, the core being: equip the whole company with Claude Code, introduce advanced models, establish an AI service desk, devolve productivity (let business write code), do permission control for data architecture, and finally must have AI-First thinking, dare to驾驭 AI employees, and force yourself to improve.

Wang Chaochao: As you can see, the AI Native team's approach is devolution and upgrading.

Jingling: Building on what Mr. Song just said, overseas companies indeed have it quite hard, encountering many network blockades and issues with whether usage is smooth. Our company's internal human-machine integration score, I think, is still relatively low. Although we're AI Native, many places haven't been fully AI-ified end-to-end. There are many problems: for example, letting a layperson use AI to make an SEO plan—you look at it and think it's impressive, but in the eyes of a ten-year-experience expert it's full of holes. Another example: the APIs we call are strong operations products; handing them fully to AI—if it goes down, the company is gone, the responsibility is too great. Also, humans have one most irreplaceable part: "production relations." AI only solves productive forces; this world exists through social relations. The reason we're sitting here communicating today, with many Sales outside—this part of trust and emotional exchange, AI cannot completely replace. In the long run, humans still need to solve very critical problems.

Wang Chaochao: That's three questions done. Let's move to the final question. Let's talk about something that keeps you up at night: as software shifts from being used by humans to acting autonomously, what's the biggest risk? Summarize in one sentence, starting with Alex.

Alex Li: I've felt this strongly recently: in the past, software development required reviewing feature priorities, but today my review speed can no longer keep up with the speed at which my colleagues write code with AI. I'm quite helpless and can only focus on key points. The risk this brings is: when you cede this part of decision-making power, it comes with a price tag. For example, one day, AI sends an email to a customer for conversion and itself offers a "50% discount" and sends it out, and consumers come to place orders. Who bears the loss of this discount? The biggest risk may be being unable to keep up with its execution speed, and my judgment becoming the bottleneck, thereby potentially bearing irrecoverable risks.

Wang Chaochao: AI is too powerful; with layer-by-layer devolution, AI may have buried bombs in hidden corners.

Xu Jia: What I worry about most is the degradation of ability. I'm now extremely dependent on AI—if Claude Code goes down for an hour, I feel terrible and can't get anything done. We can no longer go back to the days of hand-drawing prototypes and hand-writing code. I worry that if large models disappear, will my English proficiency and thinking ability decline? Another worry: if compute and electricity prices rise, I seem to have no way out.

Wang Chaochao: After living the good life for too long, going back to the hard life is quite difficult.

Song Jian: My biggest worries are three: first, like the previous two guests said—the more you trust it and delegate, the more explicit and hidden bombs are there; second, fear of network disconnection—once the network is cut, you can only say "I decide to use my own brain"; third, I chatted with my partner last week—from before Chinese New Year to Qingming Festival, every day from 10 a.m. grinding until 2 a.m., I've already reached the human limit. He asked me: models iterate so fast, why must you be so diligent, why can't you let yourself go? Even if you eat all the books, you can't keep up with new version releases. So finding balance—neither degenerating into a chimpanzee nor failing to make peace with yourself—is actually quite important.

Wang Chaochao: To summarize: relax. Mr. Li.

Jingling: We also have obvious体感 in business—when you've used a good expensive model, you don't want to use a cheap one. Also, I'm thinking about a question: last year large models were P3, today they're P5, and we hire P6 and P7 people to manage them. Next year models reach P8—can we even afford P9 people? When model capability surpasses your ability to discern, you'll think it's amazing and you won't have much judgment yourself. But technological progress doesn't shift based on human will. Why should humans be the last species on Earth? Dinosaurs lived for 100 million years; we've only had 5,000. The future may be silicon-based-dominated, who knows—we can only take it one step at a time, solving these pain points as models and ourselves iterate.

Wang Chaochao: Carbon-based organisms have limits; although technology can "extend life" for humans, human-machine collaboration still has a long way to go. Borrowing the core viewpoints of Mr. Song and Mr. Li: relax, what must come will come. Don't be too anxious; just keep learning. Thank you to the four guests. This roundtable concludes here.

---

Original publication: https://uniqueresearch.substack.com/p/src-20260428-01html
On-site reading page: https://ffcap.cn/en/research/src-20260428-01html