Original · Unique Research · 2025-12-08
Historical edition: This is the complete English translation of Unique Research's December 8, 2025 article, including its original analysis and the full published conversation. Business results, customer relationships, adoption estimates, product capabilities and forecasts are the speakers' or original author's claims at the time, not independently audited results or current offers. The approximately US$5 million ARR refers to Yu Beichuan's whole company, including its AI-marketing business, not necessarily Lessie AI alone. The source gives no currency for the 500,000 contract. Its mentions of “today,” “last night,” company ages and future time windows remain anchored to the historical conversation, not this republication.
Source-transcription note: The introduction lists five guests plus the moderator, while the moderator's opening says “four”; both are retained. Wang Yang is called a co-founder in the introduction and founder in his own introduction. The source's uncertain labels “No Banana,” “Nano Banana 2,” “DeepMind (conjecture)” and “Discord (Disgust)” are preserved as source wording, not silently resolved into verified model or company identifications. The stated 11:50 release and 11:53 integration are three minutes apart, although the speaker says “two minutes”; no timestamp has been silently changed. References to Claude Desktop as a programming tool and the “90%” adoption barrier are likewise the speaker's characterization, not a checked platform specification or survey.
Image context: The six original images are a panel photograph and speaker portraits, not data charts. Their stage identifies the event as UniqueBloom (非凡大赏), the “2025 AI Creators Summit and CHINA AI 100 & AI CREATORS 100 annual selection,” with the theme “Opening the Era of Intelligent Individuals.” The panel title and speaker roles are already retained below. The ordinary photographs are not reproduced; no image-only substantive data is omitted.
If all you have right now is a computer, an API Key, and a little unwillingness to give up, then in the AI era you already possess the minimum configuration for "one person running a company."
The question is: from here, will you live as a "user of tools" or as "someone who directs tools"?
The roundtable at the 2025 Beijing UniqueBloom event, themed "Tools and Platforms: An Ecosystem in Which AI Empowers Solo Entrepreneurs," was really about this question. The five guests onstage did not appear particularly "sexy":
One had worked in AI design for 11 years; one focused exclusively on a productivity Agent for PC; one had buried himself in Agent gateways and protocol infrastructure; one connected more than five hundred models to a single API; and one built a multimodal aggregation platform for developers targeting overseas markets.
Moderator Wang Lina is the founder of Zhiwai Culture (知外文化; transliteration), the agent for 1,400 content creators, and the person behind the nominations for this year's "AI Creator of the Year." She put the question bluntly: today we will not discuss "where the next trend will be." We will discuss only three things—how capabilities are distributed, how scenarios genuinely reach implementation, and how individuals can obtain a sustainable path for growth.
Yu Beichuan, founder of Lessie AI; Ma Liang, CEO of Guiji Jike (硅基极客; transliteration); Wang Yang, co-founder of Molink Technology; Wang Jinxing, co-founder of Chuangkit; and Li Yangbing, co-founder & CTO of WaveSpeedAI, may seem far removed from ordinary people. They work on foundational things such as gateways, protocols, templates, and model platforms. Yet the "foundation" they build determines whether an ordinary entrepreneur will be able to sustain a credible business in the future with one computer and a collection of Agent tools.
You will find that what truly concerns them is not "Does my product have many features?" but rather: how much effort must ordinary people expend to capture the dividends of this era?
Can the threshold be lowered further?
Can the division of labor between individuals and tools advance another step?
That is precisely why this conversation went far beyond the simple question of "which tool is easier to use" and reached a sharper one: as tools become more powerful, where exactly should an individual's value reside?
I. From Knowing How to Code to Daring to Say a Single Sentence
If you do not have a technical background, you have probably heard the same statement repeatedly in recent years: even without knowing how to code, you can build products and applications.
It sounds reassuring, but this conversation was the first time the claim was unpacked with sufficient specificity.
Yu Beichuan admitted that although he studied software engineering, he was not good at it and later became a product professional. Yet this person, who does not present himself as a technologist, is now building Lessie AI and giving large models the complex needs of more than ten thousand users to tackle. One of his core judgments is that large models themselves are not difficult; what is difficult is that the overwhelming majority of people do not know how to use them.
An ordinary user really behaves like this: they want to type only 10 to 15 Chinese characters and receive a perfect result. This is not laziness; it is human nature. You can hardly expect an entrepreneur to hold a long English Prompt and have a heart-to-heart conversation with a model every day.
Lessie's approach is to acknowledge this laziness and design around it. The product does not require users to explain all the background at the outset. Instead, it works like managing an intern: first give a general direction, then keep providing revision requests based on the result. The product records all these scattered remarks and accumulates them as Memory. Next time you say only one sentence, it already knows the style, tone, and boundaries you usually prefer.
The user's Input can keep decreasing, but the Context held by the product must keep increasing.
Behind this sentence is a renewed respect for nontechnical users: do not force people to become stronger; find a way for the tool itself to become smarter.
II. Turning Hard Terms Such as MCP, Agent, and Gateway into a Button
If Lessie solves the question of what to do when users cannot express themselves, Guiji Jike and Molink Technology focus on a deeper yet more decisive threshold—the difficulty of invoking external tools.
Many people heard the term MCP (Model Context Protocol) for the first time this year. It is a standard that allows large models to do more than chat by invoking tools: open software, search files, edit videos, run scripts, and more. It sounds futuristic, but the reality is that the overwhelming majority of people cannot even set up an MCP environment.
You need to install Python, configure Node, set up many other things, and then connect through some developer tool. The threshold is so high that it keeps 90% of people who have not yet truly used AI outside the door.
Ma Liang chose to quietly tear down this wall by building a PC Agent application that allows ordinary people to run thousands of MCP tools on their own computers simply by installing one program. For example, one sentence could say: remove every silent section from this video for me.
Behind that instruction lies a complex chain of calls, but for users it means pressing the little desktop assistant and continuing to drink coffee.
Wang Yang and Molink take a higher-level view: build an AI gateway that connects hundreds of domestic and overseas models through one unified outlet, with communication protocols, tool calling, and sandbox environments already configured, so developers and application teams can build Agent systems and businesses directly on this foundation. Put differently, they save everyone who wants to make an AI product the pain of repeatedly rebuilding the skeleton from scratch.
Above that sits Chuangkit. Wang Jinxing and his team spent 11 years working deeply in AI design and encountered countless pitfalls, only to discover that users can tolerate roughly 15 to 20 Chinese characters of input. "Make an invitation poster for my coffee shop opening" is already close to the limit of their patience.
They therefore use the underlying capabilities of large models and fine-tune models with their own structured data, but never expose the Prompt directly to users. Instead, every prompt is encapsulated inside the product:
First it identifies your intent, then automatically creates a structured prompt template—handling the image, style, and copy for you. The generated image can automatically land in an editable template, so the user only needs to change a few words and move two layers before using it. The complaint that generation takes only 3 seconds while editing takes 3 hours has been forcibly rewritten in their product.
WaveSpeedAI lowers the threshold in another direction: it connects about 500 open-source and commercial models worldwide to one system, allowing users to call all the models with one API Key. Small teams and individual developers no longer need to piece services together or integrate repeatedly. They can choose this single gateway and work with global platforms such as Hugging Face, Replicate, and Character.ai.
Ironically, the team has fewer than ten people, but its customer list already spans North America, Europe, Southeast Asia, and the Middle East.
When you view this group together, one commonality emerges: none asks every entrepreneur to become an engineer. Instead, all repeatedly ask the same question—how can we keep the complexity on our side and leave only the button to the user?
III. An Individual's Core Asset in the AI Era: Speed Is the Admission Ticket, Not the Endgame
When the discussion turned to core competitiveness, everyone's answers were surprisingly consistent, yet each had its own edge.
If he could choose only one word, Yu Beichuan would choose: speed.
Technology is running wild. What was science fiction yesterday is produced by a new model today. Once a new capability appears, whoever first installs it in a real scenario gains a little more of the time dividend. In this round, being half a beat slow sometimes means you can always see the opportunity but can never make it happen.
But Beichuan was also candid: speed is only the admission ticket, not a moat.
If speed is all you have, nothing remains after the window closes. The real difficulty is whether, at every moment when something new appears, you have the courage and muscle memory to try it immediately and assemble it immediately.
Ma Liang added a sharper point: more important is whether you dare to leave your comfort zone.
Doing a familiar task another hundred times feels safe. But they hope every knowledge worker will ask one question before opening a computer and beginning work:
Is there a more AI-native way to do this?
The question sounds simple but runs against human nature. It requires you to set aside the process you have already mastered and explore a new path on which you are temporarily less proficient and may even stumble. Over the long term, people who continually challenge themselves in this way will gradually separate from the mass of white-collar workers and become genuine carbon-silicon integrated professionals.
Wang Yang translates speed into two other qualities: intellectual stamina and emotional stamina.
In his view, this is the era in which one-person companies can emerge most easily. One person plus a collection of Agent tools, a model gateway, and a design tool may be able to coordinate work that previously required a small team.
What becomes amplified, then, is whether you can sustain high-density thinking for long periods and preserve the emotional stamina not to collapse through repeated trial and error.
Wang Jinxing is more like an observer, watching a circle of AI content creators every day.
Gemini 3 has been updated; Nano Banana Pro has been updated; new models keep appearing. Creators genuinely in the wave stay up at night rushing to test them, write tutorials, and create examples. Others scroll through short videos, sigh "That's impressive," and go to sleep.
That is the difference: are you watching the excitement, or are you in the wave?
Li Yangbing explained the meaning of speed through a concrete scenario. A new model was originally scheduled for release at midnight but appeared ten minutes early; his team completed integration at 11:53. The point was not to show off speed, but to seize that tiny first-mover advantage—amid widespread uncertainty, the first person to provide a usable version can easily become the default choice for a group of new needs.
He immediately added a warning: beyond speed lies an even tougher line—have you thought through the commercialization path from 0 to 1?
Many people begin by thinking about fundraising, forming a team, and discussing supply chains. Yet models are already powerful enough to help you quickly build a website, write the backend, and complete the first transaction. Why not have one person launch the minimum viable version first, get cash flow moving, and only then discuss whether to become bigger and faster?
This sounds thoroughly unromantic but exceptionally clear-eyed: in the AI era, the greatest waste is not missing a particular model, but already possessing many tools and still not daring to complete the shortest path to commercialization.
IV. Borrowed Wisdom: The Real Barrier Is How You Learn from Other People's Pain
When Wang Lina introduced the phrase "borrowed wisdom," the entrepreneurs onstage all laughed, because everyone understood that all people borrow; only the visibility of that borrowing differs.
Yu Beichuan quoted an old line from Zhang Yiming: all factors of production can be built; only cognition is a barrier. But where does cognition come from?
Not from meditating behind closed doors, but from gradually borrowing it through countless user interviews, peer sharing, and predecessors' mistakes.
After Lessie launched, the team repeatedly asked:
Why do users leave?
Why are some people willing to pay?
At exactly what moment does someone decide that this thing is valuable?
The surveys, complaints, and ratings they carefully recorded gradually became the intuition with which they judged the product's direction.
Ma Liang was even more direct: at the roundtable itself, the team offered benefits and invited knowledge workers to try Guiji Jike's PC Agent. Put another way, they turned users into their own product-manager group.
Every person's PC use cases form a highly fragmented long tail: some edit videos every day, some organize spreadsheets daily, and some need to process files in bulk. The team cannot possibly imagine every use case. Its only option is to let users bring in real scenarios and then gradually borrow the needs from them.
Wang Jinxing borrowed one user complaint: generation takes 3 seconds; editing takes 3 hours.
At first hearing, the statement sounds like criticism of AI. In essence, however, it reminds the product team that the problem is not that AI draws poorly, but that it leaves no room for subsequent collaboration.
Chuangkit turns generated results into structured templates that can be edited in layers, changing adjustments that once consumed half a day into fine-tuning completed in minutes. As a result, penetration of the AI feature increased substantially.
Li Yangbing borrows from a less obvious place—competitors' Discord communities.
They are filled with other companies' user complaints:
What is difficult to use, what frequently produces errors, and which scenarios remain unsupported.
He reads them one by one while observing how competitors position their users, package their products, and set prices. He tries not to repeat pitfalls others have encountered and gives priority to problems they have not yet solved.
But he also understands that after borrowing, the most important task is digestion. Ultimately, he must return to his original intention—serve businesses (B2B) and developers (B2D) only, without being easily tempted by the glamour of consumer (B2C) markets.
In the end, the emphasis in "borrowed wisdom, walking on thin ice" is not on borrowing, but on walking on thin ice.
You can learn from anyone, but ultimately you must return to two questions: which things are truly yours, and which choices remain after genuine deliberation?
V. Carbon-Silicon Integrated Professionals: What Truly Cannot Be Replaced?
The phrase "carbon-silicon integrated professional" sounds slightly cyberpunk.
DeepSeek explains it as a class of people who are skilled at directing AI while possessing irreplaceable human advantages in areas such as creativity, empathy, and ethics.
If the phrase remains only on a PPT slide, it is empty. These entrepreneurs, however, have already begun turning it into a living reality within their companies.
Yu Beichuan's answer is concise: what AI finds difficult to replace is decision-making and taste.
You can ask a model to organize data, write copy, and list proposals, but the final call—
Choose A or B?
What tone does this product represent?
Should this boundary be defended?
That responsibility still rests with people.
For creators, the word "taste" is even more decisive. The pacing of shots, visual style, and tone of copy you prefer collectively form your personal brand. Models can imitate them, but for now they still struggle to create a genuinely new aesthetic movement.
Ma Liang interprets carbon-silicon integration more pragmatically: any carbon-based organism that can fully harness the capabilities of the silicon-based world is carbon-silicon integrated.
He gives his team a small desktop assistant. Press it, and natural language can control the computer; image editing, file organization, and batch operations all go to an Agent.
The feeling is subtle: you remain yourself, but you operate the world with an invisible mechanical hand. Many people say they fear being replaced by AI without recognizing that those who first learn to direct AI are instead the hardest to replace.
Wang Yang recently promoted a radical reform in his company: originally, people wrote about 50% of the code; now more than 95% must be written by AI.
As a result, headcount fell 30%, while efficiency increased two to three times.
He discovered that people should no longer be divided into roles such as frontend, backend, and testing. They should stand at a higher level, abstract requirements, design systems, and align intent together with AI. Much of the time once spent on team communication becomes a process of each person collaborating with their own AI.
Wang Jinxing supplied confirmation from another direction.
Professionals looking at AI in their own fields often think it is not good enough; in unfamiliar fields, they often exclaim that it is impressive. This shows precisely that expertise will not disappear; its role is changing.
Future designers may no longer be craftspeople who complete a design alone from beginning to end, but providers of high-quality aesthetics across an ecosystem: defining corporate visual standards, producing high-quality templates, and providing paradigms for AI. The real work shifts from drawing images to shaping aesthetics and standards.
Li Yangbing simply puts skills second in interviews.
Knowing Java and Python certainly matters, but more important is whether you have an approach, whether you know how to use AI, and whether you are willing to let your way of working change completely.
Those willing to try and experiment will quickly turn AI into an exoskeleton. Those who only want to defend their old patch of ground will gradually be left far behind by the era's acceleration.
Conclusion: Tools Will Become Smarter, but People Must Decide Where to Apply Their Effort
If I had to summarize this conversation about tools and platforms in one sentence, I would say:
AI is quietly splitting solo entrepreneurship in two:
One half goes to tools; one half remains with people.
The half given to tools consists of everything that can be turned into a process, invoked, and copied: coding, editing, layout, generation, organization, calling external tools, connecting models... This work will only become cheaper and more automated.
The half left to people consists of several things that are difficult to replace:
How quickly can you understand the opportunity behind a new capability?
Do you dare take the shortest path and turn an idea into even a tiny business?
Do you have the habit of continually borrowing wisdom, and are you willing to invest enough experimentation and iteration in it?
Are you willing to preserve your taste, judgment, and sense of boundaries through long periods of uncertainty?
Tools and platforms build the ecosystem; each of us makes choices.
You can merely be a user of tools, or gradually become a species within the tool ecosystem:
Someone who knows how to select, combine, and put them to work for oneself.
When Lessie remembers your context, Guiji Jike operates your computer for you, Molink turns many models into one unified outlet, Chuangkit makes design a capability anyone can use, and WaveSpeed places global models behind one API, they are all doing the same thing—making the individual no longer small.
The next step is for you to answer one question:
In this new ecosystem, will you be someone who uses tools or someone skilled at directing them?
The gap between the two may become the greatest divide among solo entrepreneurs over the next ten years.
More Details from the Conversation
Guest Introductions
Wang Lina: Hello, everyone. I am your moderator, Wang Lina, and I am delighted to meet you. I am the founder of Zhiwai Culture (知外文化; transliteration) and the agent for 1,400 content creators. I also nominated some of today's guests for AI Creator of the Year. The four guests onstage today are entrepreneurs building ecosystem infrastructure. We will approach the discussion through three key points: capability distribution, scenario implementation, and sustainable paths for individual growth. Now, please allow each guest to introduce himself.
Yu Beichuan: Hello, everyone. I am Yu Beichuan, founder of Lessie AI. My previous experience is fairly simple: I mainly worked at ByteDance (Douyin) and later had a series of entrepreneurial opportunities. Our company currently has two main businesses. We began with AI marketing and later built the Lessie AI product. The company's annual ARR, or annual recurring revenue, is currently about US$5 million and continues to grow.
Ma Liang: I am Ma Liang, founder of Guiji Jike. Our company is a very early startup, established only four months ago, and our product has just launched. We make a PC Agent, or productivity Agent, that helps knowledge workers—people who ordinarily use computers for office work and study—operate their computers efficiently. It uses natural language to handle many tedious everyday computer tasks. The product has just entered public beta. Anyone who wants to try it can visit the AI Do experience area opposite the entrance, and we will also offer benefits shortly.
Wang Yang: Hello, everyone. I am Wang Yang, founder of Molink Technology. Molink mainly provides infrastructure capabilities for AI Agent systems, including communications protocols, model calls, tool calls, sandbox applications, and other foundational services required by AI Agent systems. Our goal is to allow developers and application teams to build Agent systems and AI applications easily on the platform. Our main live product is an AI gateway, through which users can call leading domestic and overseas models with one click for convenient development.
Wang Jinxing: Hello, everyone. I am Wang Jinxing, co-founder of Chuangkit. Let me explain what our team does: the company was founded 11 years ago and has remained deeply engaged in AI design. To date, our Chuangkit Design product has served more than 100 million consumer users; on the business side, it serves approximately 200,000 small and medium-sized businesses, and dozens of companies in a top-500 ranking also use it. Our primary business is providing easy-to-use AI design tools for consumers and a more efficient AI design production-management-distribution system for businesses.
Li Yangbing: Hello, everyone. I am Li Yangbing, co-founder of WaveSpeedAI. We operate a multimodal aggregation platform with roughly 500 open-source and commercial models. One API Key provides access to all of them. We currently serve customers primarily in North America, Europe, Southeast Asia, and the Middle East. We were established only in March this year. Our team is very small, with fewer than ten members. Because we have always focused on overseas markets, Hugging Face, Replicate, Character.ai, Freepik, and similar companies are among our customers.
How Can Entrepreneurs Without Technical Backgrounds Use AI Tools Well?
Wang Lina: First question: how can entrepreneurs without technical backgrounds make good use of your tools? What efforts have you made in product design and user education?
Yu Beichuan: How should people use tools well? I consider myself a nontechnical entrepreneur. Although I studied software engineering as an undergraduate, I was terrible at it, so I later moved into product work.
First, many products today, including ours, are actually designed for nontechnical people. The success of AI Coding products such as Lovable comes from enabling people who could not write code in the past to write it and enabling people who can code to write less.
I believe AI's greatest core appeal lies in the large language model (LLM). It converts traditional interaction into purely semantic interaction. You can speak Mandarin or natural language to it, and it can identify your intent and break down the task.
In product design, we hold one view: large models are intelligent, but the overwhelming majority of people do not know how to use them. Users are usually very lazy and hope to provide 10 to 15 Chinese characters and receive a perfect result. It is like hiring a new graduate and assigning a task without explaining its purpose or context.
What we do, therefore, is encourage the desire to express and establish Memory. Users may be unwilling to provide all the context at once, but they are willing to respond to results as if giving an intern a Comment, then we guide users to express themselves in response to results and use Memory to retain that context. It is like OpenAI building a browser: fundamentally, it wants to obtain all the context from your activity on the internet. The next time, you need say only 15 Chinese characters, and I know who you are and what your purpose is. This is our core design principle: require less Input from the user while holding more Context ourselves.
Ma Liang: For nontechnical people using AI, I think the main threshold appears when invoking external tools. You have probably heard of the MCP protocol, which became popular early this year. It is a standard that enables large models to invoke external tools. In reality, however, very few people can actually run MCP on their own computers. MCP remains fundamentally developer-oriented: users need to install Python and Node environments and use programming tools such as Claude Desktop or Cursor, among others.
That keeps 90% of Chinese users who have not yet used AI deeply outside the door.
Our current effort is to lower this threshold. For example, if I want a large model to edit a video and remove silent segments, an existing MCP tool can do it. We allow ordinary people to run thousands of MCP tools seamlessly simply by installing one program, without configuring a development environment, so the tools can handle everyday tasks on their computers.
Wang Yang: I may offer different advice. I believe AI entrepreneurs still need the courage to break through. If you use only mature, off-the-shelf tools, you may not be competitive in the market. We need to build products with the most advanced tools and capabilities.
I recommend that AI entrepreneurs, even if they do not understand technology, try Vibe Coding tools such as Claude Code or products like Bolt.new/Lovart, among others.
Once you begin, you will discover that technology is not so difficult. Even as a manager, you can speak to developers as an outsider: what is going on with this problem? Can you solve it for me? The AI tool may even try to solve it itself. I believe entrepreneurs in this field must take this step; it will help them greatly.
Wang Jinxing: Chuangkit's vision is to make design accessible and design the world through design. I strongly agree with Beichuan that a good product should not require users to learn how to use it.
For example, users making an invitation poster may have the patience to provide only a description of 15 to 20 Chinese characters. We therefore integrated models such as DeepSeek and had product managers encapsulate the Prompt. The first step is to identify the user's intent, then provide a highly structured prompt template containing the image description, copy, and other elements. Users need only make simple secondary edits to the generated result. When they click the button, AI provides four images.
As for using it well, I think the most difficult point is not daring to act. Many nonprofessionals believe they are incapable and do not even dare to try. As providers, we will make every effort to lower the threshold, but users should also embrace and use the tools, discovering the efficiency gains through the process.
Li Yangbing: I strongly agree. During the break, I helped Mr. Wu Wei (the source adds: possibly a guest at the event) examine a mini program that would not open. He had actually developed that mini program himself. Mr. Wu is a nontechnical person, but he is willing to experiment. These are the earliest people to benefit from AI.
AI, however, is moving from early AIGC enthusiasts toward productivity. As it proceeds further, most users may be less willing to experiment.
As a platform, what we need to do is first demonstrate outcomes through an example and reduce users' fear. We should make them feel, "I can copy this and do it too." The boundary between technical and nontechnical work is blurring. By lowering the threshold, we help users get started faster.
In the AI Era, What Is the Core Competitiveness of Entrepreneurs/Creators?
Wang Lina: This question is intended to stimulate discussion. Some time ago, I listened to a conversation with Kevin Kelly (KK), in which someone said that the most important competitiveness in the AI era consists of insight and talent, forward-looking cognition, and exceptional execution.
I also saw one example: a Weibo creator saw a high-quality discussion video I had shared, immediately used Tencent Yuanbao to extract the transcript, rewrote the article with AI, gained 6,000 followers that day, and monetized through Kuaituantuan's owned-audience commerce channels. I learned the method too and wrote a Zhihu article for the GEO, or generative engine optimization, era. An existing client saw it and immediately signed a contract worth 500,000 (currency not specified in the source). I therefore believe insight and execution are extremely important. What do you think?
Yu Beichuan: Calling one quality the most important tends to attract criticism, because success usually requires a comprehensive set of capabilities. But if I had to choose one, I would say speed.
If you are fast enough, you can capture the window. Technology changes too quickly. For example, No Banana (the source says this may refer to a particular model or code name) recently released a new version, or consider the releases of Sora and Claude 3.7 early in the year. When a new model appears and suddenly makes previously impossible things possible, the test is who can implement it fastest in a real scenario.
Over the long term, speed alone leaves nothing behind, but without speed you do not even have an admission ticket. Speed means not only rapid execution, but also rapid cognition and acceptance of new things.
Ma Liang: Beyond speed, I think the most important quality is the courage to leave your comfort zone.
When facing a task, do not habitually use the old approach. Remind yourself: is there an AI tool that could complete this better and more efficiently? Continually maintaining an awareness of possible AI alternatives is a crucial ability in this rapidly iterating era.
Wang Yang: Today's AI entrepreneurs have the opportunity to become one-person companies and accomplish what once required an entire team.
At this stage, entrepreneurs need intellectual stamina and emotional stamina. You will discover that you can complete the tasks of every role with yourself and a group of AI tools, eliminating a great deal of coordination and communication cost.
You need to keep yourself in a highly energized state every day and sustain that state. The real competition then lies in your intellectual and emotional stamina.
Wang Jinxing: I think there are two capabilities.
The first is imagination. AI's underlying capabilities evolve, but its unchanging requirement is to interact with the person before the screen. Many outstanding AI design creators are not necessarily the people who were best at Photoshop, but they are certainly those with the most ideas and the strongest desire to express themselves.
The second is execution. I follow many AI bloggers, and they must have had a difficult couple of days because Gemini 3, Nano Banana Pro, and other new models were released in quick succession. They stayed up all night rushing to test models and prepare articles. As observers, did we try them and play with them? Are you watching the excitement or participating in the AI wave? That is the difference.
Li Yangbing: No Banana (the source notes that this may refer to a new model) was released last night. The normal time was 12 o'clock, but DeepMind (the source marks this attribution as conjecture) released it at about 11:50. We at WaveSpeed went live at 11:53, effectively competing for those two minutes.
Beyond execution, I want to share a second point: have you thought through the commercialization path from 0 to 1?
Many entrepreneurs who want to build with AI begin by seeking financing, finding a team, and creating a supply chain. Models are now very powerful. Could one person first use Claude to write and launch a website in a day? Complete the process, earn the first revenue, and only then consider adding an Agent or a supply chain. The commercialization path must be simplified to the greatest possible degree so it can begin running first.
How Should Users, Customers, and Business Be Balanced? (On Borrowed Wisdom)
Wang Lina: I would like to share two statements: "Borrowed wisdom, walking on thin ice" and "Your competitiveness comes from what others cannot see and from your certainty."
I would like to ask everyone: have you recently borrowed wisdom from experts, users, or past experience and turned it into business results?
Yu Beichuan: I feel that I borrow every day. To quote Zhang Yiming: all factors of production can be built; only cognition is a barrier.
But where does cognition come from? It is all borrowed, or acquired later. ByteDance's organizational culture—being candid and clear, striving for excellence—was itself learned by Zhang Yiming from others, but it works extremely well.
After our product launched, we studied why users left and why they paid. The cognition we obtained from that research was also borrowed.
I believe only the ability to learn is one's own. Through rapid iteration and feedback, you transform external knowledge into your own cognition. Everything else—education, background, and money raised—is not truly decisive.
Ma Liang: Our product officially entered public beta today, and we also hope to borrow cognition from everyone here.
We hope knowledge workers will try our product and offer more suggestions. We have provided one benefit: scan the code to receive a three-month membership. Because everyone's work scenarios differ, the tools they need form a long tail. We hope to find real needs through actual use and promptly add the MCP tools people require.
Wang Yang: We have a very interesting customer who placed Claude Code behind a tool we provided, turned it into the foundational Agent capability for his vertical scenario, and then directly placed a platform on top to implement the business.
He used general-purpose Agent capabilities to build a nearly commercial Demo for a vertical scenario quickly. Something that originally required half a month or a month to develop was completed in two or three days. This method of using advanced tools for one's own purposes is a form of borrowing and a very useful lesson.
Wang Jinxing: We borrowed a pain point from users.
When large models first appeared, text-to-image generation was popular. But users complained in the community: generation takes 3 seconds, while editing takes 3 hours. AI smoothly produces the first 80 points, but reaching the publishable standard of 90 or 100 points is painfully difficult.
After receiving that feedback, we did one thing: using the hundreds of thousands of structured templates Chuangkit had accumulated, we fine-tuned the foundation model. The AI design content we now generate can be separated and edited layer by layer. Secondary editing has become very simple for users. This improvement substantially increased penetration of the AI feature.
That taught me that users do not care whether something is AI; they care whether it can solve a problem faster, more simply, and at lower cost.
Li Yangbing: Early in my entrepreneurial journey, my favorite weekly activity was browsing competitors' Discord communities (the source also prints “Disgust”).
Competitors' Discord communities contain extensive user feedback and complaints about unresolved problems. I learn fastest from competitors by examining their user positioning and product design.
But after borrowing, the key is how to absorb it and transform it into something of your own. You need your own original intention and resolve. For example, many users ask whether we will serve To C, but we remain firmly committed to serving businesses (B2B) and developers (B2D) only.
On Carbon-Silicon Integrated Professionals and Human Advantages
Wang Lina (Moderator): One final question. Some time ago at the Apsara Conference, I heard Chairman Qian of Ecovacs mention the term "carbon-silicon integrated professional."
DeepSeek explains it as people who are skilled at directing AI and possess irreplaceable human advantages in areas such as creativity, empathy, and ethics.
Within your current organizations, have you planned to recruit talent with strengths in human domains such as art, film and television, and music?
Yu Beichuan: From the perspective of a commercial startup, we consider what AI will replace and what will remain.
We have found that AI struggles to make good decisions and exercise taste.
Taste, or aesthetics and tone, is a core competitive capability. Kuaishou and Douyin, for example, have different aesthetics. The core competitiveness of artists is aesthetics.
We want the team to be as small as possible, retaining only people whom AI cannot replace. Such people use AI to handle dirty and exhausting work, then make the final decisions and uphold judgments of taste. AI finds it difficult to replace a CEO or senior decision-maker because those roles involve complex judgment and command of tone.
Ma Liang: The term carbon-silicon integration is very fashionable.
I think any carbon-based organism that can fully use the capabilities of the silicon-based world is carbon-silicon integrated. Our original purpose in building the product was to empower these white-collar workers.
We even built a small desktop assistant. I press it, issue a natural-language instruction directly, and it operates the computer for me. The process feels like mechanical ascension.
With the release of models such as Nano Banana 2 (described by the source as referring to the latest model generation) image editing has become extraordinarily easy. Carbon-based and silicon-based organisms are continuously merging and evolving.
Wang Yang: Carbon-silicon integration will enter our lives within the next six months to one year.
Our company recently undertook a transformation: people originally wrote 50% of the code; now we strongly require AI to write more than 95%.
This changed what we require of people. They no longer need specific functional divisions such as frontend and backend, but must abstract ideas and approaches at a higher level.
After completing the transformation, we reduced headcount by 30% while increasing efficiency by 2 to 3 times. The main reason was eliminating extensive person-to-person communication cost and replacing it with alignment between each person and their own AI. This poses greater challenges to every individual's capabilities.
Wang Jinxing: I broadly agree with Mr. Yu's view: what cannot be replaced is judgment.
Have you noticed that professionals doing work in their own fields often think AI is inadequate, while they find AI powerful in fields outside their expertise? I work in product and think AI writes excellent code because I cannot code, while programmers may consider AI code ordinary.
In the future, professionals such as designers will still exist but become symbiotic participants in the ecosystem. They will provide highly aesthetic templates and corporate VI standards, while ordinary marketing staff call those templates for efficient collaboration.
Creativity, judgment, and knowledge at the human level remain the core value of carbon-based organisms.
Li Yangbing: I recently discovered that our evaluation criteria in interviews have changed completely.
In the past we emphasized skills such as Java and Python; now we place greater weight on your approach or ideas. Do you know how to use AI? How would you do this?
If someone is very open-minded and willing to accept carbon-silicon transformation, that is the person we need. If someone is conservative and only wants to remain in the old territory, integration will be difficult.
This is not only an organizational transformation but also an individual one. In the future, everyone should be an all-rounder.
Wang Lina: Thank you all for your candid remarks. Today's roundtable concludes here. Thank you, everyone!