Original · Unique Research · 2025-12-04
Historical-edition note: This preserves the complete 2025 panel report and all three transcript sections. The title’s “10,000 times” illustrates repeated delivery of expertise, not a documented sales count or income guarantee. Roles, 15 million registered users, publication circulation, launch dates, comparisons, model assessments, customer relationships and market-value observations are attributed to the original reporting and speakers as of that time, not independently audited current facts. ChatExcel’s reference to audit firms and “Big Four” practice does not identify or verify any particular Big Four member as a customer or imply that all four firms use it. The Zi Wei Dou Shu example is a speaker’s account of an astrology product: its claimed accuracy and the subsequent aspiration to scientific methodology are retained as that speaker’s statements, not evidence that astrology predicts real-world outcomes scientifically. Statements about 100% accuracy, effectively unlimited productivity and passive income describe demands, aspirations or speaker claims, not guaranteed results.
If one day the professional experience in your mind could work for you in the cloud 24 hours a day, would you still rush from place to place giving lectures, attend back-to-back meetings, and sell your time by the hour as you do now?
At a roundtable during the 2025 Beijing UniqueBloom, Li Ling, founder and CEO of StudyX, entrusted students' and teachers' learning challenges to AI; Bai Shuang, founder and CEO of Fuzflo Technology, sought to turn the tacit experience in experts' minds into tradable Agent assets; Zhang Minsong, head of AI products at Century Tianhong and a 30-year veteran of educational publishing, was building an AI "external brain" for teachers preparing lessons; Pang Dawei, founder and CEO of ChatExcel, helped professionals navigate the data world behind Excel and databases through conversation; and Zhou Ze'an, founder and CEO of Biyou Technology, targeted office workers' "PPT nightmare," seeking to return documents to their fundamental purpose of expressing results.
The moderator was Xue Qian, content partner at Unique Research. Her first request was decidedly "unexpert-like": "Please introduce yourselves in the least expert way possible. Who exactly are you helping, what problem are you solving, and what kind of professional capability are you leveraging?"
That moment distilled the theme of the conversation: in the AI era, an expert's value is quietly shifting from "I know more than you" to "How many people can I replicate this knowledge for, and in how many ways can I deliver it?"
The people onstage represented roughly five different forms of leverage. Some use it to influence students' learning curves, some to improve teachers' lesson-preparation efficiency, some to strengthen professionals' data and communication capabilities, and one participant simply decomposes "the expert" into an operational Agent system.
I. Five Forms of Professional Leverage, Seen Through How Experts Introduce Themselves
If you remove their titles and look only at how they introduce themselves in the "least expert" way, the results are especially interesting.
Li Ling's StudyX does something straightforward: it helps students learn better and teachers teach better.
For students, it offers a more efficient channel for acquiring knowledge. For teachers, it helps transform teaching-and-research content and their accumulated expertise into "results" that students genuinely absorb.
The leverage here is learning methodology and teaching-and-research capability.
Bai Shuang's Fuzflo does something that many experts claim to do well but have never done systematically: it turns the tacit experience in your mind—what you cannot clearly explain but constantly use—into an expert Agent asset that can be traded, replicated, and operated continuously.
You once charged for your time; now an "expert Agent" can work continuously on your behalf.
The leverage here is the expert's tacit knowledge and decision process, what we call Know-how.
Century Tianhong, where Zhang Minsong works, produced a source of childhood dread for many people: the educational workbook series Optimization Design.
For 30 years, the company's educational publishing business has fundamentally pursued one objective: serve teachers well. Its AI product enters through lesson preparation, work that requires both substantial labor and creativity from teachers and fundamentally involves acquiring, organizing, processing, and creating instructional content. Its current product, Xiaohong Teaching Assistant, combines models with an educational-resource repository to create an innovative lesson-preparation model of "personalized customization + dynamic generation."
The leverage is the personalization of subject content and the ability to integrate knowledge.
Pang Dawei's ChatExcel sounds the most "practical": it helps ordinary people process data in Excel, spreadsheets, and databases through conversation. Cleaning, pivot tables, reports, and dashboards once required learning formulas or asking a colleague for help. Now a data agent turns data from an "incomprehensible table" into "structured information that can be used directly for decisions."
The leverage here is the capability to understand, model, and analyze data.
Biyou Technology, led by Zhou Ze'an, reaches into one of office workers' "three major sources of mental exhaustion": PPT.
It built ChatPPT, which has 15 million registered users and also serves as the "document engine" behind many major technology companies.
What the company truly wants to do, however, is not simply "help you create several more attractive PPT slides." It asks whether the roadshow, retrospective, or report you need to present has been communicated accurately and effectively.
In other words, it seeks to shift "content production" from "Do I know how to use the tool?" to "Can I explain this clearly and communicate it?"
Behind these five paths lies the same question:
Can the value of professionals shift from "selling time" to "selling systems" and "selling outcomes"?
AI is the lever, but the height you can reach is determined by the part of the professional chain that you control and that cannot be replaced.
II. In High-Stakes Scenarios, Why Would Anyone Dare Entrust Work to AI?
This leads to an obvious problem: none of these scenarios is trivial, and after a single error a user may never use the product again.
If a decimal point is off by one place, a financial report becomes useless; if educational content is wrong, students actually stumble on their exams; if one PPT slide communicates the wrong idea, a project may collapse.
In these scenarios, where tolerance for error approaches zero, why should anyone trust AI?
1. ChatExcel: In the Data World, Being Off by One Decimal Place Means Death
Pang Dawei put it plainly: "In a scenario such as Excel, if the decimal point is off by one place, you basically will not use the product a second time."
The company therefore raises "accuracy" to an almost obsessive level across four dimensions. Rather than flaunting highly technical language, he described several very concrete methods:
The processing chain must be rigorous—what the Agent does at every step cannot be left to improvised chance;
Build proprietary parsing models dedicated to the "fine work" of processing datasets;
Make the processing workflow visible—a person can still scan 100 rows of data, but with 1 million rows, the rules are all you can inspect. Expose the chain of thought, or CoT, to users so they can "open the box" and "close the box";
Data models differ across industries: logistics, e-commerce, pharmaceuticals, and other sectors each require specialized knowledge models.
The company is essentially doing one thing:
It exposes "errors the model may make" as much as possible at a level users can understand and verify, rather than "handing you a black-box result and asking you to gamble on it."
2. ChatPPT: Document Scenarios Do Not "Require Less Trust"; Their Implicit Trust Standard Is Higher
Many people assume PPT does not need to be so serious, provided it looks attractive.
Zhou Ze'an reaches the opposite conclusion: "Excel requires objective trust, while PPT requires the speaker's internal trust."
In a roadshow deck, annual summary, or strategy presentation, if AI-generated content does not match what you genuinely think, you will feel awkward onstage even if the logic is smoother and the layout looks better.
The company therefore did four things:
Native conversational interaction: ChatPPT launched in March 2023, allowing users to "feed" their viewpoints, tone, and details into the product gradually through back-and-forth Chat;
Encourage users to provide their own original content: Feishu and DingTalk daily reports, old documents, and other materials can all be added. AI is responsible for "structuring, reorganizing, and expressing" them;
Build Traceability: every viewpoint and keyword can link back to the original text, and even the mathematical derivation process can be viewed;
Distinguish identity and scenario: a teacher's courseware, an internet professional's weekly report, and a salesperson's customer proposal are fundamentally different. The system identifies them proactively and follows different logic.
Behind this is a simple but frequently overlooked principle: the less serious a scenario appears, the more demanding its hidden trust requirements may be.
3. Century Tianhong: When Subject Matter Meets Hallucination, First Admit That "the Model Is Not Qualified"
Subject education may be one of the most demanding scenarios for AI.
You may accept entertainment content that is "a little ungrounded," but one wrong problem or concept affects the understanding of an entire group of students.
Zhang Minsong's assessment is very measured: "In their current native state, models are still completely unable to meet the trustworthiness requirements of subject education."
The company therefore chose not to place all its hopes on "a slightly smarter model," but to use engineering methods to confine models to a trustworthy track, attaching knowledge bases, tool libraries, and teaching-and-research teams to seek an adapted solution.
Lesson plans, courseware, and exercises may appear to be "AI generated," but a complete content system and review logic support them behind the scenes.
The company is doing something harder but more reliable: turning AI into an "amplifier" for the content team rather than its "replacement."
4. Fuzflo: The True Risk Lies in "Planning," Not "Execution"
Bai Shuang approached the issue through the two steps of an Agent—planning and execution—an especially interesting perspective.
During execution, errors arise from improper tool use or model over-reaction, which can be improved progressively through model evaluation, tool governance, and ecosystem development.
Planning, however, determines whether "this Agent truly performs at an expert level."
He offered a contrast: a chain of thought generated by a foundation model using capabilities such as DeepSeek R1 may look complete, but it is merely "the path the model considers reasonable." A genuine industry expert's internal chain is often different from the model's, perhaps operating on an entirely different dimension.
The company therefore abandoned canvas-based drag-and-drop interfaces and returned planning to natural language. It asks experts to describe in their familiar language "how I make judgments step by step." If they genuinely cannot write it, a "knowledge-extraction Agent" speaks with them and gradually draws out their inputs, outputs, examples, and thought process.
This ultimately becomes a series of reusable chains of thought that form the expert's Agent assets.
This is why, after an instructor of Zi Wei Dou Shu astrology organized his approach to "reading an astrological chart" into a Workflow on the platform, his Agent spread directly through the community by enthusiastic word of mouth.
Here, AI serves as an externalized incarnation of the expert, delivering productivity to the outside world according to the expert's standards—and that productivity is effectively unlimited.
5. StudyX: Two Standards of "Trust" for Students and Teachers
StudyX began by helping students solve problems and later expanded into teaching and research for teachers.
On the student side, it used a familiar set of methods: RAG, knowledge bases, and problem-solving paths tailored to different subjects and question types.
When it began serving teachers, however, the company discovered a new problem: "For teachers, 'professional trustworthiness' does not mean simply that 'the model's answer is correct.' It means, 'Is this a teaching method, explanation, and standard that I recognize?'"
The company therefore did something that looks "un-AI-like": it returned content to teachers for secondary calibration.
In teaching-and-research scenarios for specific subjects and regions, it seeks to make AI output not only technically correct, but consistent with the shared understanding of that group of professionals and that educational system.
This is an important turning point: shifting from a "technical perspective" to a "scenario perspective" reveals that "trustworthiness" has many layers. Some concern mathematical correctness, some concern consensus within a professional community, and some concern the speaker's internal conviction.
III. As Foundation Models and Giants Crowd In, How Can Expert Products Avoid Being "Built as a Side Feature"?
Everyone building AI applications faces two unavoidable pressures:
Foundation models are becoming stronger;
Large technology companies increasingly enjoy building applications "as a side feature."
Why, then, would small, specialized products built around expert scenarios avoid being swept away?
1. A "Sense of the Problem" Matters More Than a "Sense of Technology"
Zhou Ze'an offered a painful observation: "If you merely reuse technology without thinking deeply about user scenarios, you are indeed in danger."
Major companies can certainly build document generation, perhaps faster and more generally.
But when you carefully decompose user scenarios:
Someone running a roadshow wants to secure funding;
A salesperson wants the customer to agree;
A boss wants the team to genuinely understand the strategy;
You discover that the true difficulty is "working backward from the outcome" to determine how a PPT or document can function in complex interpersonal settings.
This is why the company expanded into AI roadshow pages, AI voice cloning, AI-generated answers to audience questions, and even teleprompters integrated with AR glasses.
PPT is no longer a "file," but the entry point to an entire "persuasion system."
Pang Dawei offered a similar judgment: after AI appeared, demand did not decrease; it increased.
Data is more complex, comes from more sources, and has messier structures. Teams with genuine sensitivity to scenarios see new opportunities in this "troublesome work."
2. Treat the "Heavy Work" as a Barrier
Producing educational publishing content is "heavy and exhausting." In the AI era, however, that "heaviness" becomes a barrier.
Zhang Minsong said that content companies have a difficult job, but once they possess a complete teaching-and-research system, content assets, and review mechanisms, AI can release that value faster and more broadly. The more patiently you commit to "heavy content," the more favorable foundation-model iteration becomes.
Read in reverse, his statement is also a warning: if you only add "light packaging," you occupy the part that major companies can most easily build as a side feature.
3. Redefine Your "Species": Not a Tool, but an Ecosystem Node
Bai Shuang positions his company as Uber rather than an automaker: "I do not build general-purpose Agents or vertical Agents. I build the platform that allows experts' Know-how to accumulate and serve end users in the form of Agents."
Behind this is a broader judgment: the AI era will bring ever more point tools. What is genuinely scarce is the ecosystem position of an "expert agent economy" that can connect experts, tools, demand-side users, training, IP, and investors into a complete chain.
Its proposed Value-as-a-service model means something simple: stop charging by interface, Token, or time, and let users pay only for valuable outcomes.
When your business model itself is organized around "value" rather than "features," you have already moved from being a tool toward occupying an ecosystem role.
4. Beyond "Efficiency," Protect "Growth" and "Identity"
From the perspective of learning scenarios, Li Ling added a critical point: general-purpose foundation models focus more on "efficiency" and "replacing people," but in a learning setting the real concern is whether the student has improved.
This means the product must take responsibility for the student's growth path, knowledge graph, and emotional state, not merely for whether "one problem was solved."
This echoes Bai Shuang's brand philosophy: "We want AI to replace productive labor so people can do more creative work."
If a brand is only "the name of a tool," it is unlikely to live long in today's world.
Brands with genuine vitality often embody the promise of a way of life. You use them not simply to save 2 hours, but to become a "better version of yourself."
IV. The Expert's Next Step: From "Person Who Lectures" to "System Running in the Cloud"
If I had to summarize the impression this conversation left on me in one sentence, it would be this:
In the past, experts were "people who lectured," "people who wrote proposals," and "people who corrected reports";
In the future, experts will be more like "authors of systems running in the cloud."
The paths taken by StudyX, Fuzflo, Century Tianhong, ChatExcel, and Biyou Technology have already drawn a clear sequence:
First, acknowledge that foundation models alone are insufficient.
Then identify what is truly distinctive about you: a subject system, industry know-how, tacit experience, content assets, or a method for decomposing problems.
Next, use AI to turn these elements into a learning path for students; a teaching-and-research external brain for teachers; a data agent and document-generation system for professionals; and Agent assets and sources of passive income for experts themselves.
Along this path, "de-expertization" does not mean eliminating experts. It means:
Explain in everyday language exactly whom you help and what problem you solve;
Gradually take apart, write down, and solidify what previously existed only in your mind, habits, and intuition;
Let AI and systems repeat the "reusable" parts for you, freeing you to make more creative decisions.
The real danger is not that AI will eliminate experts.
It is that experts remain trapped in the old model of "selling time by the hour," watching one generation of tools after another evolve from Word, Excel, and PPT to foundation models and Agents without ever seriously asking themselves:
If one day my experience no longer has to be delivered only through "my personal presence," in what form would I want it to remain in the world over the long term?
More Details from the Conversation
Part One: Guest Introductions and Interpretations of "De-Expertization"
Xue Qian: Hello, everyone. I am Xue Qian of Unique Research. Today's topic is closely related to everyone here. Many people in the audience may be experts with extensive professional knowledge. How can we use AI to leverage our own expertise? The five guests here today "hold five levers," so let us discuss carefully what each of them is building.
First, I would like each guest to introduce themselves. I have one small request: please do so in the "least expert" way possible. Who does your product or service help, what problem does it solve, and what professional capability are you actually leveraging? Let us begin with Mr. Li of StudyX.
Li Ling: Hello, everyone. I am Li Ling, founder of StudyX. To describe ourselves to everyone here in an "unexpert" way, we help students learn better and teachers teach better.
From the student's perspective, we seek to use AI to help students learn knowledge more effectively, or give them a better way to acquire knowledge. For teachers, we want to help them teach their professional knowledge to students more effectively; AI can even directly improve the quality with which their expertise is delivered to students.
Bai Shuang: Hello, everyone. I am Bai Shuang, and you may call me Shuangshuang. Our company is Fuzflo Technology, and our core product is Leapility.
Put simply, if those present are experts in their respective fields, you must possess some "tacit knowledge and experience" that cannot be found online. What we do is use products and business models to turn the tacit expert knowledge in your minds into tradable expert Agent assets. This can both generate passive income for you and free you from labor. That is what we are building.
Zhang Minsong: Hello, everyone. I am Zhang Minsong, head of AI at Century Tianhong. I consider myself a veteran entrepreneur in education. Our company has operated for 30 years and is a long-established educational publishing company. Our core product, Optimization Design, has now reached cumulative circulation of several hundred million copies. Many of you may remember being "tormented" by our products while you were students.
Although educational publishing is a specialized industry and students use the products, teachers are our primary users from the perspective of business architecture. We continually think about how to serve teachers well. Around 2023, we therefore launched an AI product focused primarily on solving problems related to teachers' lesson preparation.
Lesson preparation entails both a substantial workload and creativity for teachers. Fundamentally, it is a process in which teachers acquire, process, organize, and create resources. Teachers were previously constrained by their environment. With the advent of AI, however, capabilities such as acquiring personalized resources and integrating interdisciplinary knowledge align closely with our current lesson-preparation product. Our teaching-assistant product, Xiaohong Teaching Assistant, addresses the "four major components" of lesson preparation: generating lesson plans, courseware, exercises, and guided-learning plans. Thank you.
Pang Dawei: Hello, everyone. I am Pang Dawei from ChatExcel. As its name suggests, our product ChatExcel helps everyone address Excel problems in daily office work through conversation. In addition to supporting Excel, we also support database processing.
Our positioning is simple: help ordinary people rapidly work with all Excel and tabular data, completing data processing, cleaning, report analysis, chart generation, dashboards, and more. Across the data chain, we provide ordinary people with a complete data agent that helps everyone use data more effectively. Thank you.
Zhou Ze'an: Hello, everyone. I am Zhou Ze'an of Biyou Technology. Our product is actually very simple: it helps everyone write documents. We have several products, including the familiar ChatPPT, which now has 15 million registered users and underpins many major technology companies. Whether at Kingsoft or 360, many of the underlying APIs are provided by us. In document generation, we may be considered China's hidden champion.
For us, the problem we solve is simple: we want everyone to return to creating the document itself rather than focusing on the tool. With the previous generation of tools, whether WPS or Microsoft, we became immersed in "how to make it." We care more about this: if you are writing a PPT for a roadshow, we help you produce it quickly and reach investors according to your intentions; if you are writing a report, helping your customers understand you is more important. What we deliver is therefore more focused on the outcome of the document's content.
Xue Qian: Let me summarize again in the "least expert" way: StudyX addresses individual learning needs; Fuzflo helps everyone extract and scale their expert knowledge to create greater value; Century Tianhong can be understood as an "AI external brain" for teachers; ChatExcel makes Excel easier to use; and Biyou Technology addresses one of office workers' three major nightmares—PPT—and helps us create PPT presentations well.
Part Two: "Hallucinations" and Accuracy Control in High-Stakes Scenarios
Xue Qian: It sounds easy when explained, but each of you is working in relatively serious settings. We have very low tolerance for errors in product outcomes. After one use, a user may decide the result is wrong and never trust the product again. These are the exacting demands facing expert products.
I would like to ask each of you: faced with the unavoidable risk of hallucinations in foundation models, how do you address the issue through product design or technology while pursuing accurate professional knowledge? Let us begin with Mr. Pang of ChatExcel, where the risk associated with error tolerance is highest.
Pang Dawei: That is a good question. Because ChatExcel is defined as a data-processing product, it is intrinsically tied to 100% accuracy. Tolerance for error in data-processing products is almost zero.
I often give an example: in an Excel scenario, if a decimal point is off by one place, you basically stop using the product. In a data scenario, "stop using it" does not mean only this time; it means never using it again, because you will not trust the product to prepare your financial statements accurately.
We therefore address the issue across several dimensions:
First, the Agent's processing Flow must be rigorous;
Second, we build proprietary models, including parsing models dedicated to dataset processing;
Third, from the user-interaction perspective, we visualize the processing. A person can scan 100 rows of data, but cannot verify 1 million rows. What can be done? We expose the processing rules, the CoT or chain of thought, allowing users to see the rules and intervene by "opening and closing the box."
Fourth, match industry knowledge models. Data models and knowledge differ completely across logistics, e-commerce, and pharmaceuticals, requiring a multidimensional approach to preventing loss.
Xue Qian: Can you share the most complex case in which a user processed data with ChatExcel?
Pang Dawei: The most complex cases are actually financial. We have audit-firm users—the source describes an example as audit firms in “Big Four” practice—using the product for month-end financial-statement processing and reconciliation. Such scenarios are complex for three reasons: first, they require extremely high accuracy; second, their table structures are specialized, with income statements, profit-and-loss statements, and other structures differing; and third, the models and knowledge differ for clients in areas such as logistics and e-commerce.
Zhou Ze'an: People may naturally assume that PPT does not require trustworthiness—"speaking well is less important than writing a good PPT"—and think it is enough for AI to make it look better. That is not true.
PPT involves "implicit trust" and imposes higher requirements. A decimal point in Excel is a matter of objective trust, while PPT concerns the speaker's internal trust. If I am presenting Biyou Technology, for example, AI-generated information must align closely with what I internally believe but cannot express orally. Many competitors on the market generate PPT with AI. Imagine that the generated PPT represents you at a roadshow: if it contains only three words, can those three words represent your viewpoint?
We apply four layers of treatment:
Interaction: PPT naturally incorporates the speaker's understanding and cannot be completed with one sentence; it requires back-and-forth Chat. We launched ChatPPT on March 4, 2023, earlier than the Microsoft Copilot product. We firmly believe users need to provide continual feedback and supplemental information through dialogue.
Content sources: We encourage users to provide their own original content, such as Feishu or DingTalk daily reports, and generate from that content.
Traceability: We currently provide an industry-leading traceability capability. Every generated viewpoint and every keyword can be traced to its source. If, for example, you provide 100MB of articles and click a viewpoint to enter traceability view, the system tells you that the sentence came from paragraph 2 of article 5, and can even show the logic of mathematical calculations.
Professional versions, or Persona: We differentiate by industry. A teacher's courseware and a professional business analysis are entirely different, for example. The model proactively identifies your identity, generates from the corresponding perspective, and calls authoritative information.
Xue Qian: Thank you, Mr. Zhou. Mr. Zhang of Century Tianhong, how do you control correctness and accuracy?
Zhang Minsong: We work in subject education, where our rigor regarding data trustworthiness is essentially "zero tolerance." In my view, the trustworthiness of today's native models is entirely unable to meet the requirements of subject education.
There are only two modes for solving this problem: one operates within the model itself, and the other outside it. After 2 years of exploration, we chose the latter—using product and engineering capabilities to solve known model problems. A model itself cannot yet reliably generate a lesson plan, courseware, or exercises, so we attach RAG and knowledge bases to solve the problem. My point is that we should not focus exclusively on one-directional capability, but solve the issue comprehensively.
Bai Shuang: I will try to explain from first principles. We build Agents, and everyone knows that an Agent mainly consists of two steps: Planning and Execution.
The main difference between a human expert and an intern lies in professional planning. If you use the chain of thought produced by DeepSeek R1, it represents what the foundation model considers correct. In real enterprise or professional settings, however, an expert's chain of thought differs from AI's. We therefore seek to preserve the expert's mode of thinking and turn it into an Agent asset.
Canvas-based drag-and-drop products such as Coze and Dify exist on the market. I believe they inherently exclude experts with genuine business Know-how, because business experts often lack technical backgrounds and cannot use them.
For the planning step, we therefore abandoned the canvas and returned to natural language. Our product Leapility emphasizes allowing users to preserve chains of thought in natural language rather than through drag-and-drop.
During execution, the underlying model's capabilities matter greatly. We evaluated many Chinese and international models. Unfortunately, it is currently difficult to find a Chinese model capable of supporting systematic tool calls by an Agent system, or Tool use; they often Over-react excessively. International models such as GPT, Gemini, and Claude already support this well.
Tool variety also matters. If an expert plans a process that must call 5 tools and 2 are unavailable, the result cannot be executed. We therefore believe an Agent is a flexible system, but it also requires a "rigid system"—third-party tool integrations—for support. This is an ecosystem issue.
Xue Qian: I would like to follow up, Mr. Bai. Is there a real case of an expert using Fuzflo?
Bai Shuang: Although we shifted from a canvas to natural language, some experts still "do not know what they know" and cannot write on a blank document. We therefore built a "knowledge-extraction Agent" that simulates a conversation with you and examines your inputs, outputs, and chain of thought. Users need only adjust the resulting 80% foundation.
Here is an example: does anyone believe in mysticism? An expert in Zi Wei Dou Shu astrology appeared in our expert community. He turned his process for reading a Zi Wei astrological chart into a Workflow. Before we had formally launched, this Agent spread through the community and showed explosive momentum.
This expert met several criteria: first, he had the ability to get started; second, he could explain his methodology clearly; third, the task itself focused on delivering an outcome, the fortune-telling result, and had high accuracy. It fully enabled an expert Agent to work on his behalf.
The experts we target in the future will address tasks that the market actually needs, rely on scientific methodology rather than inspiration, and can deliver outcomes effectively in Agent form.
Li Ling: StudyX serves overseas students. In problem-solving scenarios, we also address trustworthiness through RAG, knowledge bases, and problem-solving Flows and technical implementations tailored to different subjects and question types.
We are now beginning to serve teachers and help improve teaching-and-research content. At this point, we do not rely solely on engineering, but combine it with traditional educational scenarios so that people, or experts, can review or recalibrate the content. We focus closely on the specific contexts and standards of teacher research in vertical subfields and want our solutions to align with their understanding; only then will they consider them trustworthy.
Part Three: Facing Pressure from Foundation-Model Evolution and Technology Giants
Xue Qian: Foundation models are advancing extremely rapidly, and major companies are also investing heavily. How should we address these dual risks or pressures and ensure differentiation?
Zhou Ze'an: There is certainly pressure, but if you merely reuse technology without thinking deeply about user scenarios, that is genuinely dangerous. No matter how technology iterates, you must always return to the question: what problem are you solving for which group of people?
Documents, for example, were previously tools; now AI can generate them. If you merely build an "AI version of a PPT generator," it has limited significance because major companies can do the same. Return to the scenario: users write PPT to move investors during a roadshow or help customers understand. This requires deep scenario exploration, such as our traceability function, which the foundation model itself does not solve.
Second, teams should use technology skillfully to extend scenarios. PPT naturally has low retention because it is work done in short, frequent, fast-turnaround bursts. We discovered, however, that the downstream scenario behind PPT is public speaking, so we extended into new scenarios:
AI roadshows: after generating a speech, publish it as a roadshow page with one click, clone your voice to deliver it automatically, and even let AI answer audience questions on your behalf and collect leads.
AI glasses integration: we work with RayNeo to project speeches into AR glasses with a teleprompter that turns pages automatically. This fully addresses the audience's pain point. Without AI, we could focus only on the process of "writing." Technological iteration therefore lets us do more rather than causing panic.
Pang Dawei: I strongly agree.
First, return to user needs. After AI appeared, demand scenarios continued to grow, leaving many opportunities in vertical categories.
Second, deepen technical work. We support more data sources, including PDF files, webpages, and databases, and turn all unstructured data into structured data. That undertaking has a high barrier, and no major company can complete everything.
Third, embrace the ecosystem. AI applications must be open, cooperating with major companies and third-party applications. I do not believe one future model will cover everyone; ecosystem diversity is a natural law.
Zhang Minsong: This is fundamentally a matter of understanding. Century Tianhong has operated for 30 years and is essentially a content company. Although people consider content production exhausting and heavy, that is precisely the barrier. AI is positive for us because it improves production efficiency. Do what you do best and embrace change.
Bai Shuang: The greatest competitive capability in the AI era is understanding. Understanding determines your ecosystem position.
First, I see knowledge as an iceberg. Above the water lies public knowledge that has been used to train foundation models; what we address is tacit knowledge below the surface.
Second, there is the question of general-purpose versus vertical. Investors ask whether I build a general-purpose Agent or a vertical Agent. I say neither. I do not build an "automaker," meaning a vertical Agent, and I do not build a general-purpose system. I want to build Uber, the platform. In other words, we provide a platform that lets experts' Know-how accumulate and serve end users through expert Agents.
Our business model is Value-as-a-service. Users pay only for valuable outcomes.
Beyond the product, we are building an ecosystem. Around it are parties providing training, building IP, and making investments. We hope to build an "expert agent economy" in which every party in the supply chain benefits.
Finally, there is the brand. Products in the AI era should become a "way of life." We advocate letting AI replace productive labor so people can do more creative things. Our brand name Leapility comes from "Leap your ability," giving users a sense of identity.
Li Ling: We have seen some education companies in the United States lose substantial market value because of foundation models, while others, such as duolingo, have seen their revenue and market value rise.
General-purpose foundation models focus on "efficiency" or "replacing people," but in learning scenarios we focus on whether "the learner can improve." We attend to students' growth paths, knowledge graphs, and even emotional motivation. General-purpose foundation models will not address these continually. It is similar to how WeChat solves most problems, yet vertical products such as DingTalk and Feishu remain necessary.
Xue Qian: Thank you all very much for your excellent contributions. Although I have more questions, our time is up. Thank you, everyone!