---
title: "Li Biao of Zhidun Technology: When Large Models Shatter the Execution Barrier, What Direction Should Founders Actually Take?"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-05-12T12:01:45+00:00"
canonical: "https://ffcap.cn/en/research/src-20260512-02html"
source: "https://uniqueresearch.substack.com/p/src-20260512-02html"
language: "en"
---

# Li Biao of Zhidun Technology: When Large Models Shatter the Execution Barrier, What Direction Should Founders Actually Take?

_Original · Unique Research · 2026-05-12_

_Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening profile essay, all nine thematic sections, the complete 28-question Q&A interview across six sections, and the closing founder profile. All company names, personal names, product names, technical terms, numbers, dates and direct quotations are preserved as source-stated. Claims about company metrics, funding, hardware costs, industry trends and personal experience are Li Biao's or the author's source-attributed statements, not independently verified findings. Company and personal names are transliterated where official English forms remain unverified. The direct quotations from Li Biao are preserved as translated in the source. Twelve source images are declared but not processed in this backfill._

Unique Awards · Guest Interview

A Tech Founder Decides to Stop Believing in Technology

In the AI Era, the Most Valuable Thing Is Not Being Able to Build,

But Knowing What Is Worth Building

There is a type of AI founder who is particularly容易 misunderstood by the outside world.

They seem to understand technology best, write code best, know the boundaries of model capabilities best, and are most likely to build a product. So people naturally assume that the biggest advantage of this type of founder is "technology."

But after talking with Li Biao, founder of Zhidun Technology, I instead feel that the most值得 writing about him is not technology, but his restraint toward technology.

He spent 6.5 years at Megvii Technology, experienced the previous AI wave, and also saw the huge gap between many AI companies' technical ideals and project delivery, and then commercial payment. After leaving Megvii in 2024, he did not immediately shut himself in to write code, but first went out to meet customers.

Beijing-Tianjin-Hebei, Pearl River Delta, bosses in different industries—factories, offices, warehouses, operations teams, production links—he went to see them all.

This sounds uncool.

But for a tech founder, this may be the most important lesson: don't rush to prove what you can build—first figure out what the market is actually in pain about.

One sentence from Li Biao left a deep impression on me:

"AI capabilities are already very strong. Everyone can use AI to write code. The problem is not whether you can build it—the problem is whether what you build actually solves a real problem."

The Easiest Mistake for Tech Founders Is Treating the Hammer as the Answer

People with a technical background often have a strong sense of control.

A complex system, as long as the architecture is clear, the data flow is clear, and the modules are broken down clearly, can always be built in the end. Over the past decade or so, China's internet and AI engineering systems have trained a group of very strong technical people. They are accustomed to solving problems with engineering capability, and they are accustomed to believing that as long as the technology is good enough, the market will naturally pay.

But entrepreneurship is not like that.

Li Biao calls this misconception "holding a hammer and looking for a nail."

The nails in the market may not be the kind your hammer can drive at all. More realistically, customers don't care what model you used, don't care how many Agents you have connected behind it, don't care how beautiful your architecture is.

Customers care about only one thing:

Can this thing help me solve the problem?

Zhidun Technology also took detours in its early days.

They once接触 industries including mobile, insurance, design, cultural and creative, e-commerce, cross-border, education and more. At that time the logic was very simple: try everything, learn whatever you don't know, and first be a "good student."

But later they discovered that many demands that looked like opportunities were actually time black holes.

Writing bids, supplementing materials, doing PoCs, making demos, repeatedly explaining product capabilities, accompanying customers through long processes—the team was consumed heavily. What stung more was that the proportion that could actually convert into contracts and revenue was not high.

This is not a technical problem.

This is a problem of a startup's ecological niche.

Large companies have enough people to cultivate the market, accompany customers to educate internally, and do long-term relationship maintenance. Small companies don't have this margin. The biggest advantage of a small company should be fast decision-making, close service, and fast iteration. If you spend your time in long persuasion and consumption, the advantage is gone.

So Li Biao later formed a judgment criterion:

Don't just look at the amount—look at two-way matching efficiency.

Does the customer know what they want? Is the willingness to pay clear? Is the decision chain short? Can the demand beaccumulated into reusable capability?

If the answer is no, even if it looks like there's money, be cautious.

This is critical for many AI startups. Because the biggest problem in the AI era is not too few opportunities, but that opportunities look too many. Every industry says it needs AI, every boss says they want to reduce costs and increase efficiency, every scenario can be packaged as an Agent.

But what is truly worth doing is never a demand that "sounds very AI"—it is those demands where the pain is strong enough, the payment is clear enough, and the delivery can be reused.

Customization Is Not Original Sin, But It Cannot Drag the Company into Body-Shop Outsourcing

Over the past year, many AI companies have been stuck at one point: should we do projects or products?

Doing products is too slow, with great cash flow pressure.

Doing projects lets you survive, but it easily drags the team into a delivery quagmire.

Li Biao does not simply deny customization. He believes customization is necessary in two stages.

First, when you have no money in the early days.

When a company is just founded and has no stable income, doing customization to get cash, support the team, and survive is very normal.

Second, when entering a new industry.

Doing a few projects in the same industry is faster than shutting yourself in to do research. What customers actually want, where the process gets stuck, how the budget flows, who makes the decision, which types of features get mentioned repeatedly—these things can only be learned on site.

But once customization becomes the company's main profit source, it is dangerous.

Because it is essentially a labor-intensive business. The more projects, the more people, the higher the management cost, and the higher the delivery risk. Among the previous generation of AI companies, many large projects looked like big amounts, but with long payment cycles, heavy advance funding, and high server and labor costs, in the long run they were not worthwhile.

So Zhidun Technology now judges whether to take on a custom project by looking at only one core question:

Can it serve productization?

If a project canaccumulated reusable capability, can become a template for the next customer, can be abstracted into a product component—it is worth doing.

If it is just a one-time delivery to a person, even if the money looks like a lot, be cautious.

This is actually the most easily overlooked dividing line in AI entrepreneurship.

A project opportunity is when someone throws a problem at you and asks you to solve it once.

A product opportunity is when, after you solve it once, you discover that there is a group of people behind with similar problems, and the solution can be reused, scaled, and continuously strengthened through the system.

Many companies don't die from lack of opportunity—they die from wanting to seize every opportunity.

What Enterprises Really Need Is Not an AI Tool, but an Intelligent Execution Layer

What do enterprises most easily buy when they buy AI now?

A demo.

A flashy-looking chat box.

A "smart ornament" that makes the boss nod during a demo, but that nobody opens when they return to their real workflow.

Li Biao says the last thing enterprises should buy is this kind of toy.

But he does not stop at the shallow conclusion of "don't buy toys." What he cares more about is how enterprises can increase the probability of buying something good.

Because what enterprises really need is not an AI tool, but a layer of intelligent execution that can be embedded into business processes.

Ordinary AI tools give general capabilities. Chatting, writing copy, looking up information, summarizing files—reaching an acceptable level.

But a truly valuable Agent has only one measurement standard:

Whether the result meets expectations, or even exceeds expectations.

This sounds simple, but is hard to do.

It requires the Agent to truly understand user needs, know the context and constraints; it requires valid data, accurate sources, complete coverage; it requires the system to link with the customer's existing processes, rather than standing alone to one side; it requires traceable results, deliverable forms that can be used directly, and a delivery rhythm within user expectations.

Conversely, why is an Agent valueless?

Also simple: the result does not meet expectations.

Incomplete data, wrong results, wrong timing, wrong form, what should be there is missing.

Li Biao himself also uses Agents heavily every day. Writing code, code review, running automated tasks. His requirement for Agents is simple:

The variance at delivery nodes must be small.

What can be done today must still be doable tomorrow. It cannot be smart today and stupid tomorrow; it cannot be stable today and drifting tomorrow. Stable, professional, exceeding expectations—this is the kind of Agent enterprises are willing to pay for long-term.

This is also the watershed between "selling tools" and "selling results."

Tools sell capability.

Results sell certainty.

AI Implementation Failure Is Often Not Failure, but a Failure to Get the Rhythm Right

Many people ask why enterprise AI implementation fails.

Li Biao's answer is interesting: don't rush to define it as failure.

The fit between AI and enterprise processes is not black and white, but a gradual spectrum.

An enterprise has 20 work links. Some AI-native enterprises may have all 20 links connected by AI, even extending to more scenarios. Some pragmatic enterprises first cover 3 to 5 links, with the rest still manually reviewed. And some government, medical, and serious industry sectors only pilot at individual links.

None of this can simply be called failure.

Different industry attributes, different risk tolerance, different data security requirements, different decision chains—of course the AI coverage rhythm differs.

Marketing, advertising, and e-commerce industries are more sensitive to new technology, with relatively low trial-and-error costs, so AI can run very fast.

Medical, government, serious industry, and construction sectors are naturally cautious. Because in many scenarios, a probabilistic error from AI can bring great consequences.

Li Biao gave an example of interior design.

Using AI to assist designers in quickly generating renderings—this direction is very suitable. Because diffusion models are already relatively mature in architectural, portrait, and spatial effect expression, and can quickly improve communication efficiency.

But if the customer further demands direct integration with CAD, floor plans, 3D modeling, full rendering, and even letting design institutes and project managers construct according to the drawings, the problem becomes completely different.

Construction dimensions cannot be wrong, and industrial software cannot rely on probability. Technology can do part of it, but the maturity and scenario requirements do not match.

So he has a very practical judgment formula:

Technology maturity vs. scenario requirement.

Technology at 80 points, scenario requirement at 50 points—this is where AI most easily forms a crushing advantage. For example, writing copy, making e-commerce images, generating materials, cross-border video, first-draft reports, knowledge analysis.

Technology at 80 points, scenario requirement at 98 points—this is dangerous. Medical, autonomous driving, serious industry, and construction all fall into this category.

What founders should do is not prove that AI can do everything.

But find those places where technology is already mature enough, scenario value is high enough, and risk is controllable, and break through first.

Why Might Small Teams Actually Be Stronger?

In the past, judging whether a company was strong easily looked at scale.

How many employees, how many R&D, how many sales, how much funding, how many offices.

But in the AI era, this judgment is failing.

Li Biao says scale no longer equals company capability. Scale only说明 that this company needs this many people.

What matters more is those people who have "a feel" for technology, capability, and user scenarios.

Because a large amount of execution work is being taken over by AI. Writing code, organizing materials, running data, generating copy, editing videos, doing research—more and more tasks can be completed by Agents.

At this point, human value shifts forward.

Can you define the problem?

Can you judge what is good?

Can you distinguish the difference between three-star, five-star, and ten-star?

Can you hear what the customer did not say?

Can you know an industry's real decision chain, budget rhythm, trust relationships, and risk boundaries?

AI has made "doing" easy, but made "judgment" more valuable.

Li Biao shared an example from life.

Earlier this year, he and his family traveled and lived in three Southeast Asian countries for nearly a month. From 10-square-meter small hotels to 100-plus-square-meter five-star hotels, they went to stay and experience them all. Not because they really needed that much space, but to personally understand the differences between different levels of service, hospitality, dining, and facilities.

The difference between a 100-yuan product and a 1,000-yuan product cannot be imagined.

You have to experience it on site to know where "good" actually is good.

This is taste.

And in the AI era, taste becomes part of productivity.

Because AI can execute, but it needs someone to define the standards. What is fast, what is stable, what is beautiful, what is professional, what is exceeding expectations, what users are truly willing to pay for.

People who can define these standards will become increasingly expensive.

The Core of Going Overseas Is Not That Overseas Is Better, but That Validation Efficiency Is Higher

Zhidun Technology began accelerating overseas SaaS scenarios from 2026, and its first product is PainHunt.

This product's positioning is interesting: rather than directly making an application, it first makes the first ring of the "AI product factory."

It daily scrapes real purchase signals, complaints, rants, and unmet needs from 22 mainstream overseas social media platforms and 80+ data sources. Then, through strong models and engineering capability, it purifies, categorizes, and analyzes these raw signals into structured high-value judgments.

For example, which track still has blue ocean? Which user group is willing to pay how much where? Which demands are already strong enough to become a new product?

This capability on one hand drives Zhidun's own product incubation, and on the other hand is opened to those who need to make upstream decisions.

Why focus on overseas?

Li Biao's answer is not "overseas is more advanced," nor "overseas is definitely better than domestic."

He says the core is that at this stage, overseas validation efficiency is higher.

Payment infrastructure is mature, willingness to pay is clearer, the path from product launch to getting the first real user is shorter, and the average order value is relatively higher. For startups, the most precious thing is feedback speed.

The same amount of money, turning over 10 times a year in one market versus 2 times a year in another—the compound interest is completely different.

The same product, iterating 3 to 4 times a month in one market versus only once a month in another—will eventuallywiden a huge gap.

But he is also clear that Chinese AI companies going overseas most easily overestimate their product production capability, and most easily underestimate the difficulty of marketing customer acquisition and brand building.

Chinese teams have advantages in making products, doing engineering, and doing complex delivery.

But the real difficulty of going overseas is getting the product into users' hands.

Which channel are users in? How is trust built? What are the payment habits? How do you write marketing copy? How do you understand local culture? How do you continuously get feedback? How do you build growth?

These things are not done by simply translating the website into English.

From Megvii to Zhidun: The Entrepreneurial Foundation of a Tech Person

The most moving part of this interview was actually not the business judgments, but Li Biao talking about why he left Megvii.

Joined in 2017, left in 2024.

6.5 years—the longest he has stayed at any company in his career. He is grateful to Megvii and acknowledges that the talent density, technical accumulation, and engineering capability there were all top-tier.

But by 2023 to 2024, he began to realize a problem:

Much of his work was gradually entering his comfort zone.

Not that he couldn't do it well, but that he could do it too well.

When something changes from "I can break through" to "I can complete," for him, that is a signal.

During the pandemic years, he saw many people in poor mental and life states, and began repeatedly asking himself: can I use technology and professional capability to help more people, make them better and happier?

So before ChatGPT came out, he had already conceived a "happiness coach" product. Based on NLP and chatbots, combined with positive psychology and Eastern/Western philosophical understandings of emotion, happiness, and joy, to give users a conversational entry point.

After ChatGPT came out, he immediately got the OpenAI API, pulled a small team to make a mini-program called "Happiness Coach," which soon had paying users.

He did not continue to scale it later because he realized he lacked experience on the C-end, while B-end scenarios were closer to his comfort zone.

This also explains the name "Zhidun" (智敦).

Zhi (智) is human wisdom and AI wisdom.

Dun (敦) is honest and solid, genuinely doing things for customers without taking shortcuts.

Li Biao says whether it is the company logo or his underlying motivation for doing things, there is always one word in it: love.

This sentence, placed in a business report, might seem a bit "un-businesslike."

But placed on someone who came out of the previous generation of AI engineering system and personally went into entrepreneurship, it becomes concrete.

Because truly long-term entrepreneurship, in the end, is definitely not just about making money.

Making money is a business result, but what sustains you through chaos, anxiety, repeated trial and error, and constant education by reality is usually something deeper.

His Day Already Resembles a Small AI Company

Li Biao says AI has penetrated into every aspect of his life and work.

Writing code—he always uses the strongest model available. Not the most cost-effective, but the strongest. He believes you cannot save money on cutting-edge code models.

Personal records and diary—he has his own private model and "second brain." Inspiration, voice, images, videos, retrospectives, to-dos, self-media topic ideas—all are structured and recorded locally. AI not only records, but also helps him summarize, refine, and sync to TickTick (滴答清单).

Research, market research, reading, learning—AI helps him organize materials, polish, and analyze data. He reads for a long time every day, and when his eyes are not enough, he uses his ears, mobilizing both listening and reading simultaneously.

Overseas social media operations—he uses PainHunt to scrape data, then uses Agents to judge which signals are worth posting, generates candidate tweets, then connects to Twitter traffic, follower and visit data, and continues analyzing the next strategy.

Video content is even more exaggerated.

For his recent WeChat Channels videos, one command line automatically completes the title, cover, CTA, music, and subtitles. He doesn't even need to open CapCut (剪映).

This is not one person using AI tools.

This is more like one person working with a group of Agents.

His team even has an Agent called "Lao Yi" (老一, "Big Boss"), the eldest brother of the Agents, also called by him the "almighty war god." It is connected to what he considers the strongest type of model currently available. Servers, code, documents, knowledge bases, key operations—many are entrusted to these Agents.

But he does not pursue having as many Agents as possible.

Because Agents also need management. Context space is their working memory. If one Agent writes code, records materials, does research, and analyzes operational data all at once, the context will be polluted and effectiveness will decline.

So their division of labor is:

Lao Yi is responsible for high-level judgment, strategic architecture, and complex problems.

Code Agent focuses on writing code, code review, and deployment.

Record Agent maintains the second brain.

Operations Agent runs social media data, generates tweets, and analyzes traffic.

Few but refined—each Agent uses a strong model, each has clear boundaries, and each has agreed-upon work habits.

This is actually the prototype of many companies in the future.

A company does not necessarily expand people first.

But expands the intelligent execution layer first.

What Will the Agent Companies That Break Through in the Next Three Years Look Like?

Li Biao's judgment for the next three years is clear.

AI Agents will be everywhere, effectively integrating the software ecosystem of the previous internet wave. People will more and more easily accomplish more things and get better results.

But not everyone will truly reap the dividends.

Rather two types of people:

One type is those who are good at using AI capabilities.

The other type is those who know what the standard of "good" is.

The Agent companies that truly break through in the future will also have several common characteristics.

First, deeply rooted with customers.

Not a one-time transaction, but growing together. The product can meet customer needs today, and can continuously match customer changes in the future.

Second, synchronized with the technology growth curve.

Don't make the kind of application that gets swallowed by the base model in a couple of days. The base model getting stronger should not make you afraid—it should make you excited. Because as technology gets stronger, your product can also get stronger.

Third, delivery cannot scale linearly with customer growth by adding people.

Compliance, cost control, self-service capability, customer success—all must be scaled through technology and AI. Otherwise you are just replacing old-era labor delivery with an AI wrapper.

Fourth, both able to use current technology well and design ahead for future capabilities.

"Eating what's in the bowl, watching what's in the pot, thinking about what's in the field."

This sentence is very suitable for AI entrepreneurship.

Today's customers need delivery, tomorrow's products need iteration, and the day after tomorrow's technology changes also need reserved positions in advance.

This Era Is Good, but Don't Waste It

The most valuable part of this interview is not that it gives a standard answer.

On the contrary, it lays out the real problems that are most easily overlooked in AI entrepreneurship:

Why do tech founders easily get high on themselves?

When are customization projects cash flow, and when are they traps?

What should enterprises actually buy when buying AI?

How do you judge the value of an Agent?

How do small teams build advantages?

What is actually hard about going overseas?

How do founders use AI to restructure their day?

These problems cannot be solved by any pretty slogan. They all require repeated calibration in real customers, real products, real delivery, and real cash flow.

Li Biao's final advice to fellow tech founders is simple:

Calculate your bottom line, then focus on doing what you are best at.

How long can you tolerate no income? Will basic family expenses be affected? How much time and resources can you invest? Where do you start? These must be calculated clearly first.

Don't sell your house to start a business. Don't gamble your family's quality of life on an uncertain story.

And don't think you must raise a lot of money to start. Many times, self-fundinginstead makes the 0-to-1 stage freer, faster, and with less friction.

This is not chicken soup.

This is the realistic advice of a tech person who left a big company and truly went into building a company.

The AI era has indeed given small teams unprecedented leverage.

But leverage is not free.

It requires you to judge faster, choose more accurately, cut more decisively, use tools better, and know more clearly what is worth doing.

Li Biao says: "This era is very beautiful. Don't waste it."

This sentence sounds light, but when it lands on a founder, it is actually very heavy.

Because the era gives you leverage, it will also eliminate you faster.

What truly determines whether you can stay is not whether you can use AI, but whether you can use AI to produce results that customers are truly willing to pay for, repurchase, and recommend.

Interview Q&A

I. Why Do Tech Founders Easily Go Astray?

Q1. Explain what you are doing in one "plain language" sentence?

From the first day of founding, we firmly believe that AI will bring more beauty to people in the future.

The work we are focused on recently, in one sentence: help people in the AI era see clearly what the world is actually complaining about, what it needs, and what it is willing to pay for.

Specifically two lines:

Domestically, we use AI Agent capability to help enterprise customers improve the efficiency of office work, research, and knowledge mining. Covering scenarios such as schools, scientific research, enterprise knowledge bases, office documents, and data analysis, we connect the massive scattered data inside customers and deliver multiple intelligent systems that can truly be used.

Overseas, starting from 2026 we are accelerating SaaS scenarios. The first product is PainHunt (painhunt.dev), which is the first ring of our "AI product factory." This factory has three layers:

First layer (signal acquisition): daily scraping of real purchase signals, complaints, rants, and unmet needs from 22 mainstream overseas social media platforms and 80+ data sources;

Second layer (signal recognition): using strong AI models + the engineering capability we accumulated in the big data era, to purify, categorize, and strategically analyze raw signals, obtaining structured high-value signals—such as "which track is still blue ocean," "which user group is willing to pay how much where";

Third layer (decision support): based on these signals, on one hand we use them to drive our next product, and on the other hand we open this capability to users who need to make "upstream decisions."

The underlying belief is our judgment on human-AI collaboration. We believe this will go through three stages:

Stage 1: humans do most of the work, AI watches beside;

Stage 2: humans and AI pair up, AI does part of the work;

Stage 3: AI completes most of the work, humans are only responsible for defining the system and designing systems with feedback loops.

We firmly believe Stage 3 is not far away. Our mission is to make more and more people happier through our heartfelt creation—not just making or saving money, but giving everyone more time to do truly valuable things.

Q2. What is the first mistake tech founders most easily make?

Over-believing in technology.

People with a technical background have a strong "sense of control" in their own field, but this sense of control easily becomes "holding a hammer and looking for a nail"—the nails in the market may not be the kind your hammer can drive at all.

After I founded Zhidun Technology, I did not write a single line of code in the first few months.

What I did was interviews—starting from needs, starting from problems. In Beijing-Tianjin-Hebei and Pearl River Delta combined, I talked with dozens of bosses in different industries. I stayed for a few days at a friend's factory in Guangdong, from the workflow of the office operations team to how the warehouse production links run—I saw it all. What I observed was: where is their real pain, what are they already spending money on, what is the problem the market is crying and shouting for you to help solve.

This thing is far more important than technology.

AI capability in 2026 is already very strong. Everyone can use AI to write code. The problem is not whether you can build it—the problem is whether what you build actually solves a real problem.

Another thing tech founders should be most alert to is being熟 in production but weak in distribution and marketing. Everyone is familiar with the R&D process, but after you build something, how do you let users know about you? How do you make them willing to try? This link is a real challenge for people with technical backgrounds.

Q3. What is the biggest pit you yourself stepped into?

The biggest pit was treating "taking on everything" as "being a good student" in the early days.

In our early days we talked with dozens of potential customers, covering mobile, insurance, design, cultural and creative, e-commerce, cross-border, education and multiple other industries. At that time the logic was very simple: learn whatever you don't know, and first be a good student.

Later we discovered a very cruel fact—many demands, although they looked like opportunities, had particularly long cycles and cumbersome bidding processes. Just writing bids, supplementing materials, doing PoCs, and making product demos to prove capability to potential customers consumed a lot of the team's time.

The most painful thing looking back: among these actions, the proportion that could actually convert into revenue and contracts was very low. A large number of actions were actually无效.

But this time cannot be said to be completely wasted—it was tuition that had to be paid. Otherwise you would not know:

What type of industry is worth doing long-term

What type of customer is worth collaborating with

What kind of person is痛快 to work with, and what kind is painful

After this experience, my way of judging customers and needs fundamentally changed—not looking at the amount, but at the efficiency of two-way matching. I will expand on this change in Q5.

Q4. When are customization projects necessary? When are they traps?

Customization should be viewed by stage.

When necessary, there are two situations:

No money in the early days—a newly founded company has no other good choice. Doing customization to get cash and survive is beyond reproach;

When entering a new industry—doing a few projects in the same industry lets you feel what industry customers actually want faster than any research. This is a learning process; rolling up your sleeves and doing it is far more effective than shutting yourself in to analyze. Many early internet companies started this way—for example, Xiaoe-tech (小鹅通) also did many scenarios before its early collaboration with Wu Xiaobo, and finally broke through at the knowledge payment point.

When they become traps, usually these two signals:

The company only wants to obtain profit through custom project delivery—this path has a low ceiling, because customization is essentially labor-intensive—10 projects require 30 to 40 people, more projects require more people, and management costs and project risks rise exponentially;

Long payment cycles + heavy advance funding—I served at Megvii for many years and can clearly see this kind of problem in the previous generation of AI enterprises—large projects, heavy advance funding, and in the long run actually losing money. The time, labor, capital, and servers put in are not worthwhile for a company that wants to grow big.

My current judgment standard is: whether to do customization depends on whether it can serve productization. If a custom project allows us to extract reusable capability, become a template for the next customer, andaccumulated into standardizable product components—do it. If it is just a one-time pile of people for delivery, no matter how much money, be cautious.

A project opportunity is when someone throws a problem at you and asks you to solve it once. A product opportunity is when, after you solve it once, you discover that there is a group of people behind with similar problems, and the solution can be reused, scaled, and continuously strengthened through the system.

Q5. Which indicators do you look at to judge whether a customer demand is worth doing?

I look from two dimensions—project level and demand level.

Project level (decides whether to take it):

The most important thing is not the payment amount, but two-way efficiency. I prioritize three types of customers:

Customers with clear willingness to pay who know what they want—these customers make decisions fast and have clear needs, so we don't need to consume a lot of time cultivating and educating the market;

Customers with fast payment habits—when customers actively pay, we actively serve well. First serve the customers who have already signed contracts well, and distant customers will naturally be attracted;

Customers with short decision chains—low transaction friction.

Conversely, several high-friction signals we instinctively say no to:

A lot of time spent on repeated persuasion, bidding, writing bids, and doing PoCs

Long decision chains requiring countless clarifications and negotiations to settle

Relationship-driven decisions (for example, must rely on drinking and dining to advance)

Why instinctively say no? Not because these things are troublesome, but because a small company's ecological niche does not allow it. We don't have as large a marketing and training budget as big enterprises to cultivate the market. But as a startup team we have our own advantages—short decision chains, close service, fast iteration, like a vigorous youth. Bringing this advantage to the extreme is more important than anything.

Demand level (decides how to do it):

If the project is already settled, the needs are basically all negotiated. If we encounter unsuitable or better-solution capabilities midway, we will discuss with the customer to put them into Phase 2 or Phase 3.

Essentially look at four matches: implementation scope / fee value / cycle cost / user expectation. If they match, do it—the core of doing is always meeting user needs.

II. What Enterprises Really Need Is Not an AI Tool, but an Intelligent Execution Layer

Q6. What is the biggest difference between a truly valuable Agent and an ordinary AI tool?

A valuable Agent is one that can truly meet, or even exceed, a user's result expectations in a specific scenario.

Ordinary AI tools—you have a pile of them on your phone and computer—give you general capability, chatting, looking up information, reaching a "passable" level of及格 line. On some tasks they don't even reach the及格 line.

A valuable Agent is different. Its core measurement standard is only one: whether the result meets expectations, or even exceeds expectations.

To achieve this, working backwards you must be able to do these things well:

Truly understand user needs—as well as the constraints, context, and real intent behind the problem they want to solve;

Valid data—whether the data is complete, whether the sources are accurate;

System linkage—whether the Agent can call the customer's existing systems;

Data traceability—whether results can be traced to sources;

Result presentation—whether the delivery form is something the user can directly use;

Delivery rhythm—whether the cycle is within user expectations.

Conversely, an Agent is valueless, always in one sentence: the result does not meet expectations. Breaking it down, the reasons are nothing more than these—incomplete data, wrong results, wrong timing, wrong form, what should be there is missing.

I myself use Agents every day to write code, do code review, and run a lot of automated tasks. My only requirement for them is: extremely small variance at delivery nodes. What can be done today can still be done tomorrow, without suddenly becoming weaker, slower, or stupider. Stable, professional, exceeding expectations—this is the core of a good Agent.

Q7. How do Agents help enterprises make money? How do they help enterprises save money?

Making money and saving money are sometimes two sides of the same math problem—helping you save money is itself helping you make money.

Money-making scenario (PainHunt example):

Our system captures a large amount of user purchase intentions and unmet demands on social media. We have run reports for friends doing foreign trade going overseas—combining PainHunt-scraped data + various information sources, telling them:

Where user demands are unmet

Where users are willing to spend how much

How to do the subsequent marketing strategy

Effective business opportunities are connected to effective supply—this becomes orders. We essentially help them increase order volume, which is making money.

Money-saving scenario (domestic AI Agent example):

Our domestic customers often spend a lot of time writing reports, doing academic research, and doing knowledge mining. Originally they had to consult a large number of reports and analyze complex relationships.

The system we deliver lets users feed in a large amount of data and documents, and AI automatically analyzes the person relationships, location relationships, and event relationships within them, through knowledge graphs letting users quickly understand the whole thing from a global perspective.

Originally a report took several days, now a first draft can come out in tens of minutes, and the quality is not bad. This is saving time—time is cost.

Our mission is to make people happier and happier. "Making money and saving money" is its expression at the commercial level. But life is not just about making and spending money—we hope to help users free their time from repetitive labor to do more valuable things.

Q8. What is the last thing enterprises should buy when buying AI?

The last thing to buy is a "toy"—a flashy demo, a smart ornament not connected to business processes.

But I want to jump out of "should or shouldn't buy" to discuss a deeper question—each of us in the market is both a consumer and a producer.

From the consumer perspective, we all hope what we buy can meet or even exceed our expectations. A phone is not just for making calls—having particularly thoughtful small capabilities inside makes you feel it's worth more than the price. AI is the same—what you buy must truly solve your problem.

From the producer perspective, of course we hope users buy as much as possible. But we also hope that after users buy it, they truly realize its value, rather than leaving it untouched—because without realized value there is no repurchase, no next transaction.

Looking at both perspectives together: the key is not the "don't" of "don't buy toys," but the "do" of "increase the probability of buying something good."

I myself buy a lot of AI tool subscriptions every month—domestic, foreign, various models. You can hardly say every subscription is fully used. Some quotas are not used up. Is this waste?

I would say: moderate "waste" is actually validation, is trial and error. Just like the old marketing joke: "There's always 30%–50% of the advertising budget wasted—you can't tell which part is working, but you can't stop advertising because you can't tell." Doing things is all like this—if you don't try more, how do you know what works for you.

The disaster recovery, supplier backup, and solution backup of many key systems—essentially all are spending money to buy safety, buy information, buy speed. These "quoted wastes" are worthwhile.

The whole market needs this atmosphere of rapid validation and rapid iteration to prosper. Repeatedlytorn over one detail, with both buyer and seller exhausted—this state is not good for the ecosystem.

Q9. What is the most common reason for enterprise AI implementation failure?

We need to define in this context—what counts as "failure"?

In my view this is not black and white. The fit between AI and enterprise processes is a gradual spectrum.

Suppose an enterprise has 20 work links:

Radicals ("front-row good students")—all 20 links connected by AI, even extending to more scenarios. AI-native enterprises like OpenAI and Anthropic, and new-type small teams approach this state. But if you look at their podcasts and interviews, you will find that even they themselves are not 100% fully AI—still evolving;

Pragmatists—cover 3–5 of the 20 links, with the rest still having manual review. Do you call this failure? They are actually evolving at a speed they can bear;

Conservatives—mainly piloting a few links. This is usually government, medical, and industries sensitive to data security, risk, and rigorous decision-making.

This is not failure—it is the difference in industry adaptability itself.

Marketing, advertising, and e-commerce industries are naturally sensitive to new technology—users are willing to try and dare to bear risk, so AI can cover quickly.

Medical, government, and serious industry sectors—naturally cautious, trying at one link first is a good start.

My overall judgment on this is optimistic:

When DeepSeek R1 first came out, on the second day of the Lunar New Year I received a bunch of customer calls asking "can you deploy a set for us." This shows the market is willing to try.

A few years ago saying you used an "information system" still felt fashionable, now it is commonplace. AI will also go this path.

Domestic industries urgently need AI to improve efficiency—it's just that the "capillaries" are not fully covered yet. People with ambition can all participate in this wave to solve problems and gain value. This is a very good problem of the era.

Q10. Which scenarios look very suitable for Agents, but you would instead advise customers not to do them yet?

Core judgment: when a scenario's "difficulty requirement" exceeds the current "maturity" of technical capability, don't do it yet.

Let me give a real experience of our own:

We once worked on a directionbiased interior design. The easy part: assisting designers to quickly generate renderings, letting customers see the final look of the house faster. This took effect very quickly, because diffusion models have been iterated for many years in architecture and portrait fields and are very mature.

The customer later raised a bigger demand: hoping we would directly integrate CAD, floor plans, 3D modeling, full rendering—one-stop service. We spent a lot of time researching CAD integration, writing plugins for CAD, and understanding 3D modeling capability.

But in the end we judged this to be a thankless task. Because:

Design institutes strictly验收 your drawings at many links;

Project managers strictly construct according to the dimension units on your drawings during site surveys;

In the industrial software field, once a probabilistic model has errors, it can lead to catastrophic consequences—the entire construction process goes wrong.

This is not that technology cannot do it, but that technology maturity does not match the scenario requirement.

I use a formula to judge: "technology maturity vs. scenario requirement"—

Technology 80 points + scenario requirement 50 points = crushing advantage, this is where AI can run fastest (writing copy, making e-commerce images, generating content, making materials, cross-border video);

Technology 80 points + scenario requirement 98 points = disaster (medical, autonomous driving, serious industry, construction).

As a founder I am always calculating this match—using mature technology that forms absolute advantage to cover scenarios that are less demanding on maturity.

So we will say to customers: those scenarios where the technical boundary has not arrived, with large investment and poor effect, put them aside first. Let's together find those high-value, high-density, high-labor-cost links, break through those first, and let users see the benefit first. Technology is striding forward every day. The explosion of code model capability has made many problems that could not be solved before now solvable. This is a dynamic process.

Q11. What basic capabilities must an Agent that can truly operate in an enterprise have?

Still returning to that core: meeting or even exceeding user result expectations. Working backwards requires these basic capabilities:

Intent understanding—the model must be strong enough to understand user intent. But this does not mean you must use the strongest, most expensive model;

Data binding—the Agent must be organically bound to the user's customized, self-accumulated experience, information, and data. Otherwise it is an empty shell—this is the soul of an enterprise Agent;

Balance of cost and expectation—this is the most easily overlooked. In the early days we ran the DeepSeek 671B full version. At that time, for users to let AI run, just buying hardware (H100/H800) required 3–5 million. Which team is willing to spend millions upfront to try a scenario that has not been validated? So at that time weinstead delivered more scenarios under tens-of-billions models—hardware investment of 100K to millions, plus software, overall controllable;

Tool use, calling, analysis, result presentation—these basic capabilities that Agents on the market should have must all be complete.

ROI is the ultimate indicator—not just using the strongest model is good. Model selection, hardware configuration, software layer, data access—every step must be calculated.

Q12. What is the core capability Zhidun Technology wants toaccumulated?

Four pieces:

First, understanding of industry and technology evolution.

I joined Megvii in 2017 and left in 2024, diving into one industry for 6.5 years in the AI 1.0 era. Before that I had already worked in the internet industry for many years. From the base model to the value in the customer's hands—we have done a lot of work on this 0-to-1 chain, with deepaccumulated in engineering capability.

Second, judgment of customer scenarios.

Since I founded Zhidun, just in interviews and research I have talked with at least 50 to 60 company bosses of different scales—from millions, to tens of millions, to hundreds of millions scale—what problems each faces. This "sense of smell" for customer scenarios took a lot of time to grow.

Third, out-of-the-box engineering toolkit.

We have already accumulated a pile of engineering optimization capabilities—RAG knowledge bases, query algorithms, performance optimization, large-data-volume scenario implementation capabilities, etc. These toolkits let us meet user needs in a relatively fast time, not starting from 0.

Fourth, and most critical—AI-native production capability.

This is what I particularly want to emphasize.

Most of our internal code writing, code review, and deployment are already AI-native—a large number of Agents are writing and modifying code in our workflow. Our boundary use and integration of these tools are at the industry frontier.

In the future we will follow the "AI product factory" model: continuously turning research findings, delivered scenarios, and market signals scraped and analyzed by PainHunt into individual products delivered to the market.

The foundation of all this is—I can close my eyes and see clearly what a product's core architecture should look like. Just like an architectural architect who can close their eyes and see the rebar and concrete structure inside the house that users cannot see. This isaccumulated from over a decade of leading engineering teams, being an architect, and writing code—it is my scarcest asset.

PainHunt.dev will become the first ring of this factory—both a product and the underlying data foundation. Letting more and more people enjoy the joy and beauty of technology—this is our direction.

III. From Projects to Products: How Can AI Startups Avoid Being Dragged to Death by Delivery?

Q13. How do you distinguish between "project opportunities" and "product opportunities"? Which customer needs can make money short-term but you choose not to take?

Yes, and we have already turned down quite a few.

Last year I stayed in Shenzhen for more than half a year. What I particularly appreciate about Shenzhen is—fast pace, no dawdling. Sit down, have a cup of tea, chat, and there is immediately a result on whether it can be done. This environment forces you to develop fast judgment capability.

My judgment standard is simple and direct:

Several types of needs I will actively say no to:

Calculations actually lose money—there is cash on the surface, but the time, labor, and risk put in add up to negative;

The customer themselves has not thought through what they want—requires us to spend a lot of time persuading, educating, and doing product论证;

Scenario mismatch—exceeds our capability advantage area.

If what the customer proposes is something we cannot do or are not good at, we will actively give them alternative solutions, or recommend reputable peers in our circle. Not taking the order, but preserving the relationship—this is long-termism.

Several types of customers I will prioritize collaborating with:

Customer budget is in place, with high initiative and enthusiasm;

Wants to become better themselves, does not need us to persuade them what to do;

Rhythm match—they want to go fast, and we also want to go fast.

Being able to make money short-term does not equal being worth doing. We don't want to spend too much time on "persuading others"—this kind of thing is very consuming. We prefer to spend our energy on customers who "already have consensus."

Q14. What is the hardest balance for AI startups?

Two things: direction and energy allocation.

Direction is choosing which of the many unmetsegmented scenarios to深耕, or serving customers in a specific industry. This is a role positioning problem—choosing the wrong direction, and no matter how hard you try later, the effect is discounted.

Energy allocation is a more micro but equally critical balance:

How deep to root at the user site, and how deep to polish your own product?

What is the ratio of energy between marketing and market, and product R&D?

How much focus to put domestically, and how much overseas?

There is no "one-size-fits-all" answer. Every stage must be re-evaluated and re-calibrated.

In terms of product capability—most scenarios now no longer have big pure technical problems. The reason is that the match between technology and user scenarios, we have actually already explained clearly in Q10—the match between technical capability and scenario maturity is a dynamic judgment.

What is really hard is rhythm. It is whether the founder has the capability to continuously calibrate direction and continuously reallocate energy.

Q15. Which capabilities does Zhidun Technology most want to standardize?

Zhang Yiming once said "a company is a product"—I would say now a company is an Agent.

We hope every link of this "Agent" can be highly intelligentized. Looking from two levels:

First level: internal production layer

We have done many large systems before—big data, cloud computing, e-commerce, AI all involved. From market demand to the customer's hands, we have a clear feel for every link:

Market demand → direction selection

Research → product design → R&D → testing

Operations → release → disaster recovery → capacity growth

Market → customer success

We hope every link can be standardized and quantifiable—each link equipped with a pile of Agents to meet demand.

To give a specific example: when you write code with Claude Code, you can treat it as the entry point of a company—break down tasks, spin up multiple sub-Agents to do things. Real-life companies can actually also be architected this way: Feishu, DingTalk, Enterprise WeChat, Confluence, Wiki, Jira are all knowledge management tools, and a lot of company information is already built on them.

What people need to do has changed:

Not executing every link

But defining the whole system, defining the standards of each link, defining the positive and negative feedback loops—and then handing all this to AI to amplify and execute

This has a premise: when you define the input and output standards of each link well—just like defining functions when writing code—it is a stable structure. As long as you don't suddenly replace the underlying reasoning model with a weaker one, it can be very stable in the long term.

After defining well, it can scale horizontally—through computing power and servers expanding the multiple effect, bringing delivery speed to the extreme.

What is the key to all this? It is someone who can define "what is good"—what a three-star hotel is like, what a five-star hotel is like, what a ten-star hotel is like. Someone defines the standards, AI executes—this is the new division of labor.

Second level: customer delivery layer

Returning to the goal—meeting user expectations. Original software development was a waterfall passed layer by layer. My current judgment is that this will become much faster.

Even the "customization dilemma" we discussed earlier—with AI Agents, customization caninstead be delivered in batches and at scale. Originally customization piled 100 people, now it piles 100 AI Agents. But you must understand this field, be able to reserve interfaces well, and abstract capabilities—using software engineering's encapsulation, abstraction, polymorphism, and enterprise architecture patterns—it can be friendly for batch customization.

The scoring, validation, and correction of each link—this is the essence of what we call the "AI product factory"—how to build a system that can self-check and stably output.

What do users finally feel? "The demands I give you are being done faster and faster, and the feedback rhythm is getting faster and faster." Companies still using the old model will have their market seized by fast, stable, and good companies—without a doubt.

This is the capability we want to accumulate—from internal production to external delivery, systematically integrating all links into workflows and customer delivery, letting both users and the team enjoy productivity dividends.

Q16. What is the difference between "selling tools" and "selling results"?

Our core will still run the software track.

Software is what we are best at—can serve more users at scale and in batches. This kind of thing is like "blessing"—what you do can empower more people and help more people, that is a beautiful thing. Returning to real business, it is larger revenue. The two complement each other.

Service model we will do, but usually "heavy AI-assisted" service. We hope most services are automated—a large number of bots are already helping us carry daily work.

Result model we also recognize. We use PainHunt to run vertical reports for specific users, and after chatting with a group of users we immediately received 8–9 demands in different directions. We cannot do all of them. Perhaps at some point in the future, we turn data + insights + algorithm strategy into an online service—users directly pay for results.

Like when we were at the e-commerce industrial park in Guangzhou, a boss proposed "I help you increase sales, you do revenue sharing with me"—paying by results. This model is also not out of the question.

But current energy is still focused on software—doing what we are best at first solidly, thoroughly, and to the best.

Q17. "Why shouldn't I just use big-company models, why must I use yours?"—How do you answer?

What we deliver is simply not what big companies deliver.

First, we will not do what big companies themselves can do well. This is basic ecological niche judgment.

Second, big-company models are just one link in our delivery chain. This question can be asked in reverse—"why not just buy Nvidia hardware and train the model yourself?" Obviously the answer is: ecological division of labor. In a market everyone has different roles—big companiesbiased向 infrastructure, webiased向 scenario integration.

What we are good at is: combining user needs with the AI tools and productivity capabilities flying everywhere, giving users a move-in-ready experience.

To make an analogy—users won't buy bricks, cement, and paint themselves—you would find a renovation company. You can also directly buy an already-built house. Everyone has different roles, and the market can prosper.

Third, when we chat with customers about needs, we use all our past experience, knowledge, pitfalls stepped into, and tuition paid—this is the greatest value in saving customers time. Customers don't need totorn over underlying details, only need to focus on the results they want. Essentially it is still a division of labor problem. Big companies, startups, customers—each role has its own position, and when matched well everyone can get what they need.

IV. Small Teams, AI Tools and Overseas Markets: Why "Fewer People" Might Actually Be an Advantage?

Q18. What is your new understanding of "small-team entrepreneurship"?

Scale indeed no longer equals company capability. Scale only proves that this company needs this many people.

What matters more? Those people who have "a feel" for technology, capability, and user scenarios. The value of this kind of person will grow larger and larger.

I understand it as a matrix—on the left is your capability boundary, on the right is the boundary of market demand. People who can dig deep on either side, with insights and their own "taste," will become scarce.

Why? Because a large amount of execution work has already been liberated by AI—

For example, writing code—you only need to think clearly:

Where the data comes from and where it flows

What the result standard is

How to judge good or not, beautiful or not, fast or not, stable or not

As long as you can define it, the problem is half solved, or even more than half—the rest is handed to the strongest AI.

The market side is the same. There is a lot of "unsaid" in the questions customers raise—

Can you hear what they truly want?

What deeper dimensions are behind this problem in their field (e-commerce, insurance, medical, cross-border)?

What can the industry do now, and what can we do?

The user's rhythm, expectations, budget, trust relationships, decision chain—do you have a grasp?

People who are deep on one side, or versatile enough to define clearly, are the most valuable—they can distinguish "three-star standard vs. five-star standard vs. ten-star standard."

What does this kind of person usually look like? People from big companies who have seen complex systems, built large-scale systems, and crawled through pitfalls—these experiences areinstead more valuable now. AI has made "doing" easy, and "judgment" is the barrier.

I want to share something not directly related to work—I encourage the team (including myself) to personally experience different things more.

Earlier this year my family and I traveled and lived in three Southeast Asian countries for nearly a month. From 10-square-meter small hotels to 100-plus-square-meter five-star grand hotels—we actually couldn't use that much space—but we just wanted to personally experience different levels of service, hospitality, dining, and facilities.

Only by being on site, on the scene, and personally experiencing will you truly know: three-star is like this, five-star is like this, and ten-star is how different. You will know where a 100-yuan product and a 1,000-yuan product are truly different.

This "taste"—is the source of product power. Human perception of many things comes from things previously experienced personally. These unique feelings and past accumulation—in the AI era haveinstead become the scarcest assets.

Q19. Which position or capability should most be amplified by AI?

Start with what you are best at.

The more you understand something, the easier it is to use AI to do it well—if you don't even understand what something is yourself, it is hard to direct AI to do it well.

Tech startup teams should be particularly alert to one thing—easily getting high on technical details and ignoring market feedback and growth. After accelerating R&D with AI到位, the freed-up time must be投 back to market, growth, and customers.

Q20. What do you think about AI companies going overseas? Why focus on overseas?

Core in one sentence: at this stage, overseas "validation efficiency" is higher.

Specific manifestations:

Mature payment infrastructure—willingness to pay and payment channels are smoother;

Short product launch path—the distance from product to the first real user is closer;

Relatively higher average order value—for the same product value, there is more pricing space;

Some scenarios run faster overseas—this is the core.

The core is feedback speed.

I prefer to do things efficiently—procrastination and dawdling are things I cannot accept. This is wasting your own life and also wasting others' lives.

For example: the same 100,000 yuan of funds,

In Market A can turn over 10 times a year—the compound interest generated is huge

In Market B can only turn over 2 times a year—the same funds, different results

Iteration count is the same:

Market A can iterate 3–4 times a month

Market B may only iterate once a month

This is the founder's biggest differentiation.

But going overseas also has a lot to overcome—cultural customs, consumption habits, user experience, local trust—all require rapid learning and rapid validation. No market is absolutely good or bad, it depends on team resource matching. We are阶段性 choosing to put part of our energy overseas to run. There will also be a lot of opportunities domestically in the future—we will continue to look.

Q21. For Chinese AI startups going overseas, what is most easily overestimated? What is most easily underestimated?

Most easily overestimated: their own product production capability.

Most easily underestimated: the difficulty of marketing customer acquisition + brand building.

Over the past year, as AI capability has evolved, the cost and efficiency of making products themselves have both improved—but thisinstead sets off the relative difficulty of "distribution."

The real challenge is:

How do you truly get your product into users' hands?

How do you quickly get feedback and continuously evolve?

How do you adjust according to local culture, customs, and consumption rhythm?

Which bank card do users use to pay? Which bank's payment efficiency is highest?

What kind of marketing copy, what kind of channel, can truly reach the target user? How do you truly build growth?

In product and technology building, Chinese teams have advantages—we have already navigated through more complex problems than this scenario. It isinstead "delivery to users' hands" and "customer acquisition"—that are our real challenge in going overseas. This is our own real feeling.

V. Finally, Please Leave a Few Judgments

Q22. In the next three years, what common characteristics will the AI Agent companies that truly break through have?

Two core characteristics:

First, deeply rooted with customers.

Not a one-time transaction, but growing together. Your product can well meet customer needs now, and can continuously match in the future. Customers evolve with you, and the product gets stronger and stronger.

Second, synchronized with the technology growth curve.

Don't make the kind of product that "gets replaced by OpenAI, Anthropic, Zhipu, or Doubao models in a couple of days."

We have seen in the past few years—some AI application companies raised a lot of money but ultimately had poor results, the reason being that the rapid evolution of base models compressed the space of upper-layer applications very thin.

Two perspectives on technological development:

One is fear—lying in bed at night afraid this technology will subvert you;

One is longing—longing for spring to come and technology to become stronger.

The second attitude is truly "being friends with technology." Technology gets stronger → integrate into your own product → product gets stronger. This is a顺势 posture.

Other characteristics:

Compliance, cost control, user self-service capability—cannot let delivery scale linearly with customer base by adding people, must expand through technology and AI;

Both able to use current technology well and design ahead for future capabilities—eating what's in the bowl, watching what's in the pot, thinking about what's in the field.

Q23. In one sentence: what is the most easily misunderstood part of enterprise AI implementation?

The most easily misunderstood are two extremes—one believing AI can do everything, one believing AI can do nothing. The correct approach: return to your own scenarios, your own business, your own heavy-cost, tedious, labor-intensive, repetitive things, pick out 2–3 of them and first use the tools on the market, see how to realize the benefit, then slowly enjoy the dividends brought by technological change.

Q24. In one sentence: what is your judgment on AI Agents in the next three years?

AI Agents will be everywhere, effectively integrating the software ecosystem of the previous internet wave, and people can more and more easily accomplish more things and get better results. And those groups that are "good at using AI capability + know what the standard of 'good' is" will reap the biggest dividends of the era.

VI. On "Entrepreneurial Journey, a Day in the AI Era" — Founder Profile

Q25. From Megvii Technology to Zhidun Technology, what kind of decision was leaving in 2024?

This was a decision酝酿 for a long time, not a spur-of-the-moment impulse.

I am grateful to Megvii from the heart.

Joined in 2017, left in 2024, stayed 6.5 years—this is the longest I have stayed at any company in my career. In the previous AI wave, Megvii's talent density, technical accumulation, and engineering capability were all top-tier, and working with a group of very strong people, I learned a lot.

But by 2023–2024, several things were交织 in my heart.

First, the signal of the comfort zone.

Most of the work was in my comfort zone, and the leverage effect it could play gradually decreased. I am not a particularly "able to stay idle" person, and I have extremely high requirements for the use of my time—I have worked at many enterprises before, and Megvii is the only one I rooted at long-term. When something slowly changes from "I can break through" to "I can complete," for me this is an内在 signal.

Second, the introspection brought by the pandemic years.

During those pandemic years I saw many people around me in poor states—mental, emotional, and life-level all included. A friend once evaluated me like this: "Seeing you is seeing the word altruism," "How come your energy never runs out."

I began repeatedly asking myself one question: can I use my technology, my professional capability, to help more people—make them better and happier?

Third, finding the prototype of a product entry point.

Before ChatGPT came out, I was already conceiving something—based on traditional NLP + chatbots, making a "happiness coach" or "psychological coach."

I spent a lot of time studying positive psychology, took Tsinghua University's positive psychology course, and studied the definitions of "human emotion, human happiness, human joy" in Eastern and Western philosophy. I wanted to build a system that absorbs these classic ideas and provides a conversational entry point so users can get help from it.

After ChatGPT came out, I immediately got OpenAI API access, pulled a small team and quickly made a mini-program—called "Happiness Coach." It soon had paying users. At that time I was doing it in my spare time, but I could already feel the prototype of this thing.

I did not scale it later because I realized I did not have much experience on the C-end, and at the same time, while employed it was not easy to fully compete with new players. After leaving to start a business, I first returned to my comfort zone—B-end scenarios. Most of my previous career was B-end delivery, with greater certainty.

About the name "Zhidun."

Zhi = human wisdom + AI wisdom;

Dun = honest and solid, genuinely doing things for customers without taking shortcuts.

If you look at our logo—whether before or now—there is always love in it.

This is not decoration, it is the founder's deepest reminder to himself. Everything I do is driven by "love." I hope to transmit more energy and more beauty through AI and technology. From near to far, letting more and more people become better and better.

About the founder's core.

The ancient Chinese said "when poor, attend to your own virtue in solitude; when successful, bring benefit to all under heaven." The Four Sentences of Hengqu (横渠四句) are also always in my core.

I am from Hunan, and Huxiang culture has deeply influenced me—family, country, and all under heaven. But the premise for doing this is—you must first let yourself, your team, and your company truly grow, with greater strength, to better help this society.

So after leaving Megvii, I did not take too long a vacation, and quickly started researching the market, talking with bosses, and doing things.

I look forward to our small flame, our small company, being able to slowly grow stronger—with greater energy and greater strength, to help more people. This era is very beautiful. Don't waste it.

Q26. Describe your "ordinary day." From waking up to bedtime, what role do AI tools play in your life/work?

I can put it this way—AI has penetrated into every aspect of my life.

Briefly聊 a few scenarios.

Writing code—always use the strongest model.

Not the most cost-effective, but the strongest currently experienceable. You cannot save money on cutting-edge code models—on this point I am very firm.

Personal records and diary—private model + my "second brain."

I have kept a diary for many years.

My approach now is: running a private model on my own device that can handle image, video, and voice understanding. When I have an idea I chat with the bot—it records everything I say, and automatically helps me do daily retrospectives, thinking, summarizing, and insights.

It also connects with my TickTick (滴答清单), combining my to-dos for the day, content I need to send externally, and my self-media topic ideas—automatically identifying what I have judgments on and what is worth talking about, syncing back to TickTick through API.

All materials are finally saved locally as structured documents—images, information, fragments, all systematized. This is my "second brain."

The more data accumulated, the stronger technology becomes in the future, and the more deeply AI can understand you, forming greater compound interest.

Research, market research, reading, learning.

I invest a lot of time in this area every day. AI helps me do research, polish, and analyze data.

I read for a long time every day—sometimes my eyes are not enough. So I also fully use my ears—mobilizing both eyes and ears simultaneously. This should be the same for all modern knowledge workers.

Reading, writing, thinking, creating, producing, delivering—these links are already inseparable from AI.

Overseas social media operations—data-driven + Agent closed loop.

I did not used to post much overseas content. Recently I started systematic operations—the approach is:

PainHunt's daily scraped data, corresponding to dedicated skills tools that run the latest situation;

These tools automatically judge which signals are worth posting and which are worth promoting;

Run several candidate tweets for me, most of which I directly post;

Twitter's traffic, follower, and visit data, connected to my Agent;

The Agent automatically analyzes this data and gives me the next optimization strategy.

Video content—one command line handles the whole process.

This thing may be hard for many people to believe. For my recent WeChat Channels videos—one CLI automatically completes: writing the title, adding the cover, adding CTA, adding music, adding subtitles. I don't even need to open CapCut (剪映).

Although I have bought annual memberships for these editing software, I basically don't use them now. Voice input replaces typing, and video editing uses command-line tools—a lot of things that originally required tool operations can now be done by AI.

These tools can all become products in the future. Some are open-source projects, and if we use them a lot we directly PR feedback back; some are capabilities we have accumulated ourselves, and will become products beyond PainHunt in the future.

Generally speaking: everything that can be done with AI, I already do with AI. AI is not a tool, it is a collaborator.

Q27. If there is someone with the same background as you (senior in technology, family responsibilities, considering entrepreneurship) wanting to do AI Agent entrepreneurship, what is the one piece of advice you most want to give?

Calculate your bottom line, then focus on doing what you are best at.

You don't need to calculateespecially finely, but you must be clear:

How long can you tolerate no income? This is the cycle.

During the no-income period, will basic family expenses be affected? Will living conditions be downgraded? This is the bottom line.

How much time and how many resources can you invest in this? This is the boundary.

Where do you plan to start? This is the direction.

Calculate these things clearly, and the bottom-line risk is locked.

After locking, you will find that the remaining things are actually easy to solve.

Why do I think "bottom-line risk" is the first principle?

Because I have seen too many people make two types of mistakes in entrepreneurship:

One type is not calculating the bottom line clearly—previously talking about "selling your house to start a business" "selling your house to trade stocks"—this kind of risk exposure is too large. If it takes 1.5 years in the middle to see the first decent cash flow, your family will suffer with you—this should not happen.

The other type is misestimating funding needs—always feeling "I must raise a lot of money to start."

My own real feeling is—doing things with self-raised funds, many times isinstead better than taking big money.

Having more money has its benefits, but more money also brings many side effects: more friction, slower decision-making, more extra things to do, greater reporting pressure. During the product validation and iteration period, these are all reverse leverage.

The decision speed, response speed, and rhythm freedom of self-raised funds—are one of the biggest advantages in the 0-to-1 stage of entrepreneurship.

About "focus on doing what you are best at."

Compared with those big capital players, you indeed cannot play high and hard—hiring the best people, using the biggest budget for marketing and market. But this is just like fighting a war—you must choose the battlefield you are best at, the enemy you are best at, and the time and space you are best at to fight.

You cannot say you are a small team and go fight against their thousands of troops.

You must find your own advantage.

My summary is only 14 characters:

Figure out the bottom-line risk, figure out your advantage,

Use limited resources to maximize your value and leverage.

Everyone is good at starting from different places. Some people already have orders before coming out; some people need to crawl through before they can get the team and funding done. Nothing is set in stone, but bottom-line awareness must come first.

Q28. There is an Agent in your team called "Lao Yi" (老一). Talk about the culture behind this and how you work with AI daily.

Lao Yi is the "eldest brother" of my team's Agents, and I also call him the Almighty War God.

He is connected to the Opus 4.7 1M model—currently one of the strongest I have experienced.

Besides Lao Yi, I also have several capable AI partners—everyone has different divisions of labor and performs their own duties. Our team's positioning of them is very clear—they are not tools, they are team members.

What is Lao Yi doing now?

Simply put: all servers, code, documents, knowledge bases, and key operations—I have basically entrusted them all to them. I can now basically handle everything by chatting with them through my phone.

Specific example: I recently used the Claude Code system to build a collaboration system. From server selection, deployment, validation, launch, to operations—basically all completed by Lao Yi and his sub-agent system.

But I don't let Lao Yi do everything. I position him as a "high-level strategic partner"—what I discuss with him are relatively architecture-level and judgment-level problems. For specific execution—I use other Agents, through IM conversations, or through Claude Code to分流.

Why分流?

Because context management is a real problem.

Each Agent's context space is its "working memory." If one Agent writes code, records materials, does research, and analyzes data all at once—its context will be polluted by various topics, and effectiveness will decline.

So my division of labor principle is:

Lao Yi: high-level judgment, strategic architecture, complex problems

Code Agent: focuses on writing code / code review / deployment

Record Agent: helps me maintain my "second brain"—diary, retrospectives, material library

Operations Agent: runs social media data, generates tweets, analyzes traffic

---

Original publication: https://uniqueresearch.substack.com/p/src-20260512-02html
On-site reading page: https://ffcap.cn/en/research/src-20260512-02html