Original · Unique Research · 2025-12-11
Historical edition: This preserves Unique Research’s December 2025 roundtable account and selected Q&A. Figures, business results, competitive claims, and forecasts—including the six-month replacement prediction—are the source’s or speakers’ claims at that time, not independently verified current findings. The headline’s 95% refers in the discussion to Zhang Fan’s description of an MIT report about proof-of-concept projects; it is not a verified failure rate for all AI applications. The narrative uses a US$200-per-month app example, while Q5 says only a US$100 app; both source formulations are retained. “Shenzhen 300” preserves the source’s index wording, without independently resolving the index identity. Wilderness Stars and Yuanli Intelligence are descriptive English renderings of the Chinese company names; Mulan.pro is the product name supplied by the source.
If a conversation opens with, "If this question is bad, blame Gemini 3 Pro; if it is good, my prompt must have been well written," then what follows will not be a forum so serious that everyone yawns.
In Singapore this time, Unique Research CEO Wu Wei brought together three entrepreneurs with completely different styles.
Li Jinglin, founder of CometAPI, which is building a high-speed middle-layer highway between models and applications;
Liu Rushan, CEO of Wilderness Stars, which bundles editors and post-production teams into AI for deeply synthesized video;
And Zhang Fan, who left his role as COO of Zhipu and now works on digital labor + enterprise AI agents at Yuanli Intelligence.
From the outset, the three thoroughly discussed one question: in the AI era, the real opportunities are simply not where everyone is holding up a phone and chasing trends.
I. Where Are the Opportunities? The Places Everyone Else Dismisses
To summarize Zhang Fan's view in one sentence: the scarcest thing today is not models, but people willing to seriously turn models into productivity.
He offered two sobering figures.
On the supply side, OpenAI alone already produces tokens at the trillion scale. Over the next few years, the combined effects of model improvements, expanded chip capacity, and declining costs could quite plausibly multiply token capacity another thousandfold.
On the demand side, Zhang cites MIT research as saying that only about 5% of proof-of-concept (POC) projects are truly implemented, while 95% remain at the demo, presentation-material, and emotional-value stages.
Zhang Fan said that, based on more than two years at Zhipu and an enterprise-service portfolio worth several hundred million (currency unspecified in the source), that figure is entirely fair.
In other words, the model "power plant" is already large enough. What is genuinely lacking are the people who bring its wiring into factories, shopping centers, assembly lines, customer-service centers, and finance departments.
He therefore reached a distinctly contrarian judgment: nearly all resources in the primary and secondary markets are being poured into production—larger clusters, stronger chips, and more powerful models—but what has not been addressed seriously is demand: how to make enterprises willing to pay over the long term.
This is the enormous gap he describes, and the fundamental reason he left a leading Chinese foundation-model company to build B2B digital labor, a field many people disdain.
It sounds like moving against the wind, but extend the timeline and you will find that in every technological revolution, those who truly make money are often not the people standing directly beneath the spotlight.
II. The Middle Layer: The People Saving Developers Time and Money Are Carrying the Load
Li Jinglin builds something many people overlook despite its critical importance: a model-aggregation middle layer.
He describes the pain points he encountered when previously building AI applications in highly specific and realistic terms.
If you want to connect a dozen models for testing, each provider requires you to apply for an API key, complete KYC, add funds, reconcile bills;
Every provider uses different interface parameters and writes its documentation in a different style;
Before you have even decided which model to use, preliminary debugging has already exhausted your patience.
What CometAPI does is therefore simple and direct: it provides one unified OpenAI-compatible interface, connects virtually every model you can imagine, and lets developers switch and test with one click.
You can try Claude today, compare Gemini tomorrow, and examine the performance of Chinese models on video and audio the day after that, all with a single configuration line.
More practically, once your GMV rises, the middle layer can secure larger discounts than official channels because it has far greater bargaining power with model providers than individual small developers.
This may sound like merely saving developers time and money, but the truly interesting point is that from the middle layer you see bubbles and trends earlier than anyone else.
He said their users span Russia, the United States, Japan, South Korea, Southeast Asia, and Africa. Every major foundation-model upgrade washes out another group of speculative developers.
Some site-network factories rush to register domains such as SomethingAI.com, create one prompt per page, connect to an API with one click, and harvest quick money through search-engine optimization (SEO).
The moment models iterate or a new toy launches, their traffic is cut in half.
The products themselves accumulate nothing; they have only trends and paid acquisition.
Li Jinglin gives application developers a harsh warning: if all you do is patch and mend models, your real competitor is not another app but the next version of the model itself.
Today you use a workflow to fill gaps in what it cannot do. When the next generation learns those things itself, you disappear.
From the middle-layer perspective, he therefore offers one key signal about opportunity:
Applications that last must be embedded in working relationships and organizational arrangements, becoming an irreplaceable part of an enterprise's daily work like the previous generation of SaaS, rather than toys swept away by the next model upgrade.
III. Deep Synthesis: Turning Two Weeks of Video Editing into One Spoken Request and Letting AI Finish the Job
If Li Jinglin connects the wiring for developers, Liu Rushan packs a series of complex video-production lines into one all-in-one product. She began with a simple question: how many people in the room already use AI to create video? Almost everyone raised a hand. She smiled and said, "Then you are all potential customers."
She then asked whether they found AI video production painful. Everyone nodded vigorously.
Why? Because the pain is very real. To make a presentable video, you may go through this process: generate visuals in one model, adjust characters in another, find a third model for voiceover, use yet another tool for sound effects, decide how to format subtitles, and repeatedly convert formats while uploading and downloading materials. Moving from the first version to something deliverable may take several days or even two weeks. Wilderness Stars reduces all this to stating a requirement: "I want a one-minute promotional film for a company mobilization meeting." The system automatically selects models, arranges the workflow, generates visuals, adds voiceover, sound effects, and subtitles, edits the footage, and ultimately provides a commercial-grade video ready for production or advertising. If you dislike a detail, every node can be adjusted with high controllability. Two weeks become two hours; manual work becomes a natural-language-driven Deep agent.
More interestingly, Wilderness Stars did not begin with consumers, but with the heaviest enterprise delivery. It first embedded itself in publicly listed companies to build entire AI video matrices, then produced advertisements for major US brands and genuinely delivered individual television commercials and short videos. The R&D team spent six months doing agency work itself, personally experiencing creators' pain points, and only then returned to product building and iterated by eating its own dog food. This made it bolder about rejecting consensus in technology choices: when everyone rushed into GUI-based one-sentence video generation, it did not follow. The reason was blunt: that was a toy, not a design that made AI work more conveniently, and it could not deliver commercial-grade video. For AI to truly work, it needs a linear, controllable workflow; to be controllable and adjustable, it cannot remain a total black box.
That is why, as canvas-based workflow-node orchestration has only just begun to become popular in the industry, they had already validated the route at the beginning of the year and put it through a full round of practical use. Liu Rushan's logic is simple, blunt, and sincere: do not follow the crowd; take your own path. Start from genuine pain points and choose the right direction half a year early, and places others mistake for moats may already have become minefields.
IV. AI Agents, Digital Labor, and the Electricity Paradigm: What AI Really Changes Is Who Pays the Wages
When the discussion turns to AI agents and B2B, Zhang Fan's perspective carries the unmistakable calm of someone who came from the model side.
He divides agent evolution into several generations: from the earliest single-purpose tools, to workflows with simple planning, to self-planning and self-learning MemAgent systems, and then upward toward the L1~L5 AGI capability levels described by OpenAI.
His dissatisfaction with workflows is clear: they are quick to adopt but limited in expressive power. You cannot write 10,000 Flow definitions to cover the complexity of real-world tasks.
Giving the next generation more freedom through foundation models creates a different pitfall: controllability. It assumes every brain is Newton's and differs only in experience, while reality is far messier.
Yuanli Intelligence wants to go one step further: give every specific environment a dedicated optimal agent and write industry know-how and enterprise preferences directly into the weights rather than merely stuffing them into context.
This introduces a larger question: is the AI revolution more like the internet or more like electricity?
If it resembles the internet, it implies new connections, markets, entrepreneurs born in the 1990s, and traffic playbooks.
Zhang Fan offers a completely different analogy: the internet addressed information connectivity, while AI addresses productivity. Electricity did not transform society because electricity itself made money, but because it was embedded in elevators, assembly lines, subways, and factories, redefining urban form, divisions of labor, and management theory.
Applied to today's AI, he stresses that the real issue is not whether we can build larger models, but when financial reports from the S&P 500 and Shenzhen 300 will show a clear AI contribution.
Behind this lies a traditional B2B pain point: in the software era, if you sold CRM, customers could never clearly calculate how much direct value it created. Whether they paid depended entirely on budgets, relationships, and instinct.
But define AI as digital labor and the situation changes completely: global enterprise labor expenditure this year is US$64 trillion.
That money is, at its core, what enterprises pay for human labor.
If you tell a company with 500 customer-service employees that you can complete the work of 500 people at the cost of 450 and guarantee the result, it becomes an extremely difficult proposition to reject during an economic downturn.
That is why he emphasizes that real B2B in the future will not sell another pile of features, but a measurable result.
Outcome-based payment can also go astray. It should not mean taking a share of the money a customer earns, which would drag you into endless bargaining.
The healthiest approach instead resembles electricity: if you use electricity to earn 10,000 trading stocks, the power plant does not take a share; if your monthly electricity bill is 100, the power plant does not pay that bill for you. The source does not specify a currency for either amount. Electricity charges only a reasonable basic fee. The key is that if it is genuinely useful, you continue using it.
Treating AI as electricity rather than as a traffic tool is the path Yuanli Intelligence is attempting.
V. One-Person Companies and the Illusion of Moats: Do Not Mistake Minefields for Moats
One story in this discussion belongs in the required curriculum for AI entrepreneurs.
Zhang Fan described a friend's experience. The friend wanted to build an AI product with a moat, so he drew four defensive walls: first, build a powerful editor; second, place a large collection of prompt templates inside it; third, buy 1 million PPT templates; and finally, partner with the largest copyright library and control the copyrights too.
It sounded like a detailed and rigorous strategy. Then Midjourney launched and crushed the entire logic: people did not need an editor because HTML could express web pages more powerfully; prompt templates became instructions that the model could infer itself, such as "you are a middle-school student, so solve the problem this way"; and image generation redefined copyright questions.
By the time he finally reacted and tried to build another wall, new models consumed the entire process again. What he had considered a moat turned overnight into a ring of minefields.
Zhang Fan therefore has reservations about enthusiasm for one-person companies: if you spend two months building an app that earns US$200 per month, it may sound exhilarating, but consider calmly that perhaps 95% of its value comes from the model and only 5% from you.
Someone else then needs to expend only 5% of your effort to replicate the same app, and 100 people can do so.
Over time, you encounter reality: if a first-mover advantage cannot quickly become a barrier, it is only a beautiful screenshot of a growth curve.
He closes with an analogy: today's foundation models are like an ocean. You think the water is not deep or stable enough, so you keep raising a lighthouse in the same place. But every six months the water level rises 100 meters, and even the tallest lighthouse will be submerged. What you should build is a boat that floats—as the water rises, you rise with it.
For most entrepreneurs, what does the boat mean? Deep insight into a specific industry, user relationships you control, your operating system, and the courage to pursue a non-consensus route. Models will not replace these things anytime soon.
VI. Whom Will Digital labor Replace? Be Rational, but Also a Little Romantic
When Wu Wei brought the discussion back to labor and asked whether Yuanli Intelligence's digital labor would replace human labor, Zhang Fan answered bluntly: it certainly will.
But he explained it on two levels.
Rationally, he compared humans 5000 years ago with humans today: people 5000 years ago had roughly the same brain capacity as we do, yet one person today can deliver hundreds or even thousands of times the productivity of someone millennia ago.
The increase did not happen because everyone acquired a brain with an IQ of 200. It happened because we invented divisions of labor, organizations, and tools, stacking individual intelligence into organizational intelligence.
If you view all human society as a giant MoE, a mixture-of-experts model, some people research only algorithms, others build only products, some specialize in supply chains, and others work deeply in a single niche. Each person is an expert node, and the whole system's intelligence far exceeds any individual point.
Future AI will likely work the same way: every individual will have a cluster of AI agents, and enterprises will have complete digital-labor teams. The real difference will not be which general foundation model you use, but how you embed it in your division-of-labor system. On the emotional level, his framing becomes very internet-native:
Today people jokingly call themselves corporate cattle and horses. If we can truly assign all the genuinely beast-of-burden work to machines and preserve the tasks more suited to people—creation, judgment, empathy, and storytelling—that may be one small piece of romance most worth anticipating in the AI era.
Of course, the road will not be smooth. For an individual, the worst condition is refusing to face the reality that a role may be replaced while also failing to design a new division of labor proactively.
The best state is to ask yourself honestly: in this future system of people+agent+digital labor, where should I stand? What can I contribute that models cannot do well in the short term?
VII. Do Not Rush to Find the Next Trend; First Choose Your Coordinate System
This session was titled New Opportunities in the AI Era, but its real value did not lie in offering several startup ideas ready for immediate use. It helped replace the coordinate system through which you think.
From Li Jinglin, you realize that you should not stare only at the fireworks in the application layer. The real money often hides in the dullest and most foundational places—the ones that save developers from losing a few more hairs.
From Liu Rushan, you see that a real product is not assembled from several glamorous models. It is refined through six months, a year, and repeated rounds of real delivery that compress two weeks of user pain into two hours.
From Zhang Fan, you are forced to confront a reality: AI resembles electricity more than the internet. It is not helping you create a few more traffic gateways; it is quietly rewriting enterprises' cost and profit structures.
If I had to summarize in one sentence the lingering force this conversation left with me, I would put it this way:
In the AI era, real opportunities never grow where spotlights shine brightest. They grow in the gaps others disdain, overlook, or consider too slow, too difficult, or too B2B. Whether you can seize them depends not on how fast you run, but on whether you dare to place yourself where the rising water will not repeatedly submerge you.
Over the next few years, we will probably see countless new model launches, news about hundreds of millions or trillions of parameters, and screenshot after screenshot of one-person companies earning thousands of dollars a day.
But you may need to ask yourself only three questions repeatedly:
Am I merely patching the model?
Does what I am building rise and fall with the waves, or can it rise with the water level?
In this new system of people+agent+digital labor, which node do I actually want to become?
If you can think these three things through, the new opportunities of the AI era may already have quietly moved to your side.
Selected Roundtable Q&A
Q1: Could each guest briefly introduce yourself, your company, and your company's position in AI?
Li Jinglin: We build CometAPI, a one-stop API aggregation platform for foundation models aimed specifically at overseas markets. OpenRouter is a similar overseas product. Our aggregation is broader: beyond language models, it includes video, audio, music, Embedding, and vectors, covering virtually every mainstream model in the world. We serve AI application developers worldwide, helping them use different models at lower cost, including Anthropic Claude, GPT, and excellent Chinese foundation models. Our business is growing rapidly, roughly doubling every month.
Liu Rushan: We focus on implementing AI video at the application layer. We aim to solve pain points in AI video creation by bringing all models together. Users only need to provide a prompt; we automatically recommend suitable models, saving them from switching and debugging models, and let them edit and compose directly on a canvas without manual work in CapCut. Based on our vision of the future, we believe the videos people watch will eventually be generated individually for them. Current products cannot support that scenario, so we must build a product that lets AI work. Mulan.pro not only aggregates every mainstream model all-in-one, but also completes editing directly in a linear fashion. In June, we had already implemented editing inside a canvas workflow with natural language driving AI to build the workflow automatically. After all this iteration, we find that we remain the only product in the world to achieve it.
Zhang Fan: I have long worked on the commercialization of foundation models and commercial products and previously served as COO of Zhipu. Yuanli Intelligence focuses on closing the enormous gap between models and applications. We believe existing agent approaches have limitations, from single-point solutions through workflows and self-planning approaches such as MetaGPT. We want to build the next generation of agent: an optimal agent tailored to every environment, with its capabilities reflected in every weight across 1 trillion parameters. We seek a unified approach that truly turns AI from foundational intelligence into productivity embedded in the physical world.
Q2: Li Jinglin, why did you choose to build a middle layer such as CometAPI between the application and model layers?
Li Jinglin, CometAPI CEO: I discovered the problem while building with AI myself. As a developer, connecting to different models creates many pain points. You need to manage every provider's API key and bills and discuss cooperation with each. The largest pain point is reading and understanding each provider's documentation during debugging because their interfaces and parameters all differ. The early development workload is very large because you must select models. CometAPI solves these problems by providing a unified OpenAI-compatible interface, enabling one-stop, one-click testing and invocation of mainstream models worldwide. Aggregation also lets us provide larger model discounts than official channels. In the AI era, the time window matters. We can rapidly add the latest models—for example, Gemini 3 became available 20 minutes after its official release—saving developers both effort and time in making money.
Q3: Zhang Fan, you previously worked at Zhipu, a leading domestic model company. Why did you leave and turn toward the business-to-business (B2B) direction that everyone now dislikes?
Zhang Fan: I saw the enormous gap between models and applications. Today nearly all resources are invested on the production side: stronger chips, larger clusters, and stronger models. We now produce 1 trillion tokens, and output may increase 1000-fold over the next five years. On the demand side, however, an MIT report says that only 5% of today's POC projects are implemented and 95% create no value. Most applications still provide emotional value and remain concentrated in peripheral work such as contract review. If AI cannot be directed into core operations, it cannot consume that 1000-fold increase in token output.
We believe the AI revolution resembles electricity more than the internet. The internet created connections and a new marketplace, but AI addresses productivity. Electricity's revolution came from being embedded in the physical world. The core logic is that AI must appear in the financial reports of S&P 500 and Shenzhen 300 companies and genuinely improve fundamentals; only then is the progress substantial. We should not view the problem from a model company's perspective, but from the enterprise perspective, embedding AI in the physical world to create greater change. Traditional SFT, manually labeled data, and workflow construction cannot solve this problem; we see a possible dawn in applying reinforcement learning to commercial environments.
Q4: Liu Rushan, competition in AI video synthesis is intense, and applications from large companies may replace or reconstruct your work. What differentiates you?
Liu Rushan: If we must discuss competitive strength, I believe the greatest point is people and the fact that we do not follow the crowd but have our own path. As the company's founder, I am in a completely different, wholly non-consensus state, as are our US CEO, my daughter, and our technical lead Yile.
Focus on results and efficiency: we can shorten the time needed to produce one video from two weeks to two hours and raise an ordinary person's production level to that of a Hollywood blockbuster.
Technology-route choice: we refused from the beginning to compromise by choosing a GUI route because we knew it was a transitional product. We believe the generation process for AI should be linear. Rejecting the crowd was difficult, but we endured.
Forward-looking product design: we established the canvas-workflow route in January, produced a demo in April, publicly demonstrated it in June, and continued iterating, while large companies did not react until November. Many companies merely follow trends, but we have thought very clearly about every next step in technological development.
Go deeply into enterprise markets and solve pain points: we are highly grounded. Before R&D, we spent more than six full months personally delivering AI video projects for publicly listed companies and well-known US companies. Through direct delivery, our R&D team learned where creators' pain points lay and then found the right direction from those pains. This process would be impossible at a large company.
Q5: Zhang Fan, you said a startup's first-mover advantage must become a barrier and its model content can be neither too high nor too low. What is your view of today's enthusiasm for one-person companies?
Zhang Fan: I believe the current market is misleading people about one-person companies. If one person can spend two months building a US$100 app, then 95% of the product's value comes from the model. That means you found only one point, and someone else needs to spend only 5% of your effort to build an identical product. Your market will be consumed rapidly. A startup's first-mover advantage must become a barrier; today's one-person consumer-facing companies find it very difficult to build barriers.
In addition, I believe the AI revolution resembles electricity rather than the internet. The internet improved connection efficiency and created a new marketplace. AI addresses productivity. Electricity changed society by becoming embedded in the physical world—for example, elevators changed cities and assembly lines changed work patterns. AI must be embedded in the fundamental financial reports of S&P 500 companies to produce genuine social change. If a startup is too close to the model, a model upgrade destroys it; if it is too far from the model, it has nothing to do with the era. You must therefore create a Z axis, keep your model content neither too high nor too low, and find an overlay with the model industry's main track.
Q6: Will digital labor replace human labor? Where will replaced human labor go?
Zhang Fan: It certainly will, without question.
From a rational perspective, human brain capacity has changed little over the past 5000 years, while productive capacity increased 1000-fold. This improvement did not occur because everyone gained a brain with an IQ of 200, but because divisions of labor and tools turned individual intelligence into organizational intelligence. Society today resembles a giant MoE, a mixture-of-experts model. In the future, models will consume a large amount of work, just as today we do not debate whether automobiles will appear.
From an emotional perspective, we should let machines eliminate all the genuine beast-of-burden work. We should pursue work better suited to people, allowing them to focus more closely and do happier things.
Timing prediction: I believe AI may need only six months to genuinely complete work for us and achieve replacement at scale. This may happen faster than people expect.
Q7: What do the guests believe will be the primary business models and profit channels for companies in the future?
Zhang Fan: It has never been that B2B is difficult; only software B2B is difficult. Enterprises are always willing to pay for products that create visible value. If we transform AI into a labor market, it resembles advertising. If your service lets a customer make more money—for example, invest two units and earn seven—the market is not difficult to charge.
The key is turning the service into a visible, measurable result. We should treat ourselves as electricity. Customers need not share the money they earn using electricity with us, but as long as it is useful, they will consume it continuously. We should charge only a minimum electricity fee, push usage costs down, and deliver better direct business results.
Li Jinglin: Overseas markets have relatively mature business models and users are willing to pay. AI addresses productivity, much as the Industrial Revolution raised production efficiency without changing working relationships and organizational arrangements. I believe a major future direction for AI applications is irreplaceability: products that remain difficult to eliminate even as models continue to iterate. Established SaaS products, for example, permeate every aspect of our work and life and focus more on productivity, work coordination, and working relationships and organizational arrangements. AI models only complete your capabilities; they cannot replace your entire perspective.
Liu Rushan: Our business models include the following:
Consumer top-ups: offering US$20 and US$50 top-up options for consumers.
Template fees: many people will upload templates that eventually become paid models serving long-tail demand.
API licensing: we provide the API interface to other customers and help them use the product. They sell it under their own brands without indicating that it is ours.
Enterprise delivery: directly delivering products to enterprises.
We personally prefer the template business because it more closely resembles an open-source community and, from one perspective, can better withstand competition from large companies.