3-Hour Interview with Manus Founder Hong Xiao: Become a Key Variable
Summary
Hong Xiao’s core entrepreneurial method is not to predict the endgame linearly, but to position himself early at an affordable cost and try to become a variable capable of changing other players’ choices. In 2019, he judged that personal WeChat外挂 would be shut down but did not know when, so he prepared WeCom SCRM roughly 6 months in advance; in 2025, before model capabilities were fully mature, he built an Agent that could write code, call APIs and operate browsers. Xiao stressed: “The world is not a linear extrapolation. You have to make yourself the important variable in the game.”
The most valuable lesson from his first startup was that the business model, financing scale and product experience cannot serve a larger capital narrative; they must return to real demand. The WeChat editor abandoned ad matchmaking and sold SaaS directly; once paid plans appeared at login, revenue grew “many times over, perhaps by an order of magnitude.” A competitor may have raised more money on a higher-ceiling story, but also took on a harder target to deliver. Xiao’s discipline is that “capital should be a tool,” while investors’ day-to-day opinions “should carry roughly the same weight as those of a KOL”; serious disagreements belong in formal meetings or at the board level.
Monica’s cold start did not depend on lucking into organic growth; one acquisition bought the scarcest window of time, user sample and controllable initial volume. After the 2023 Spring Festival, the team acquired ChatGPT for Google, using its thousands to tens of thousands of daily additions to feed, interview and calibrate PMF for the new product, while keeping the 2 brands independent and forcing the team to relearn overseas growth. Once Monica had overseas payments and pricing in place, it reached the initial low target of “RMB500,000 a month” within 2 weeks; later, Monica and ChatGPT for Google each reached several million users, but the former had several million members while the latter had several million active users, so neither should automatically be described as MAU.
The overseas decision rested on a rough but effective calculation: willingness to pay for software overseas was roughly 5x domestic levels, the dollar exchange rate roughly 7x, or “5×7=35”; even if language and experience reduced the team to 10% effectiveness, it “seemed possible to earn 3x more.” Xiao acknowledged that no one on the team had lived overseas long-term and that his English may have peaked in high school, but believed that in a window of opportunity, “courage” mattered more than spending 6 months first building an overseas organization. Fragmented overseas markets, language and cultural differences raise costs, but also create multiple “small hilltops” and greater strategic depth.
Model companies control when capability leaps arrive; application companies turn those leaps into product value ordinary users can feel, and the 2 are not zero-sum. Monica chose to integrate nearly every model the team had heard of, across multiple endpoints. The cost is surrendering control over when capabilities appear; the advantage is being able to switch vendors quickly and choose among models. Xiao therefore sees models trending toward commoditization over the long run, while brand, workflow, user understanding and go-to-market are more durable accumulations. “The model comes first, but the shell has to evolve too.”
The new Agent’s core is not piling more features onto a Chatbot, but having the model plan, write code, call APIs and browse, then execute repeatedly inside a cloud VM. The product will shift from synchronous “A asks, B answers” to asynchronous collaboration: it first provides a Planner and steps, reports continuously during execution, and lets users correct course midway. In testing, it could identify the animal shown at a specified second in a YouTube video and scrape Elon Musk’s posts for semantic analysis. Xiao judged that models had “just crossed” the usability line, but at the time only Claude 3.5 Sonnet showed strong long-horizon planning and step-by-step agentic capability.
DeepSeek’s biggest contribution to application founders was not simply cheaper models, but proof that technical innovation can translate directly into user experience and distribution through free access, visible reasoning and connectivity. Xiao believes Perplexity strengthened trust through webpage citations, while DeepSeek showed the public its thinking process; OpenAI’s o1 missed the chance to make Reasoner capabilities tangible because of its paywall and simplified presentation. DeepSeek put first-tier model quality and a new experience directly in users’ hands, generating distribution without KOL spending; it also made model commoditization and open-source catch-up more realistic, easing application companies’ anxiety over API prices and model providers entering their markets.
Organizational strategy must switch with one’s position: operate conservatively without a tailwind, but once leadership is confirmed, hire more aggressively, capture users and widen the window. Xiao admits Monica was not aggressive enough in 2023 and that in 2024 he once extrapolated difficulties linearly; by 2025 he was again stressing that “AI will not let any Spring Festival pass quietly” and that “when big tech can understand your innovation, that is when it becomes dangerous.” His endgame is not a specific product, but to “think in the age of the era,” keep one’s own pace in AI’s 2nd or 3rd year, compete globally and “never short” the nonlinear leap in model capabilities.
Deep dive
1. The Second Startup Began with a Product Founder Already Proven in the Market
Butterfly Effect was founded in June 2022. The first interview took place in autumn 2024, when the company had just completed a financing round that Xiao classified as Series A. Born in 1992 and 32 at the time, he describes his MBTI as INFP.
After graduating from Huazhong University of Science and Technology in 2015, Xiao went straight into entrepreneurship. He first built a WeChat Official Account editor that eventually served several million editors, then incubated enterprise-WeChat SCRM in 2019. In 2020, he sold both products together to a unicorn.
His second startup entered through generative-AI applications and built Monica first. Zhang Xiaojun positioned him as a case study of “not starting from a foundation model, but building applications on top of one,” and focused on how that path could acquire users, generate revenue and build durable moats.
2. Staying in Wuhan Created Both an Information Edge and a Resource Deficit
Staying in Wuhan was not initially a grand strategy. Xiao had not yet graduated when the company was registered, and the team kept iterating in a familiar environment. Later, when it moved into WeChat plugins and then decided to go overseas, fast go-to-market mattered more than relocating and relearning overseas operations.
Wuhan kept the team away from the trend cycles in Beijing and Shanghai, but its information sources became Twitter, YouTube, overseas indie developers and technology bloggers. Xiao said the team was “physically in Wuhan,” while its learning channels looked more like those of overseas founders or indie developers in Chiang Mai and Vietnam.
The trade-offs were clear. Wuhan lacked not only top algorithm engineers but also mid-level managers, and the financing distance was much larger. In 2020, when the team was looking for roughly RMB10M–20M, 2 competitors separately raised $30M and RMB200M.
3. Cross-Disciplinary Ability Defined Xiao Earlier Than His Grades Did
Xiao spent years in HUST’s technology club, the Lianchuang team, while also running the university’s WeChat Official Account. He was not the strongest coder, but could combine a little technology with content operations, building features such as grade and electricity-bill lookup. That combination became a product advantage.
When Xiao was writing a calculator for addition, subtraction, multiplication and division, the technical co-founder who would later join him was already building tower-defense games with designers. Xiao learned early to “do what you are good at,” moved toward product, and continued bringing that classmate in to lead technology.
His blog, The Other-Dimensional Software World, trained another basic skill: explaining why software is good requires studying the product, competitors and category, then expressing the differences clearly. Xiao believes the writing experience directly helped his later product-management career.
His concrete advice to students is plain: “Invite the outstanding classmates around you out for a meal more often.” Back then, treating a technology-club classmate to northeastern Chinese food could build a relationship; after graduation, recruiting the same person might cost “Michelin meals every day.”
4. Confidence in 2015 Was Quickly Shattered by a Year Without Users
After registering the company, the team tried anonymous campus social networking, a secondhand marketplace and several apps. Neither users nor revenue materialized. Both founders had previously accomplished things at school, so the string of post-graduation failures delivered heavy negative feedback.
By August 2016, the money was almost gone and the team was taking outsourcing work to survive. They agreed to move to Beijing for jobs after the National Day holiday. Xiao’s biggest fear was not failing to make a little money, but that the opportunity cost was “especially high”—remaining trapped in a Wuhan apartment could leave them behind as the technology industry’s train departed.
The final attempt was a PingWest content-product hackathon. The team turned a familiar WeChat Official Account pain point into a Chrome plugin. They expected only a prize of several tens of thousands of yuan, but unexpectedly attracted Liu Yuan and ZhenFund.
5. One Hackathon Turned “Preparing to Get Jobs” into a Million-Yuan Financing
The day after the hackathon, the team was invited to meet ZhenFund and its investment committee. The investor directly handed them a SAFE requiring only a signature, worth roughly RMB1M. Xiao described it this way: “At this time yesterday, we were still thinking about going to work; at this time today, we already had a million-yuan investment agreement.”
There was randomness in the investment decision. Years later, Liu still remembered Xiao wearing a Pink Floyd T-shirt and assumed he was an artsy young man. Xiao said he merely listened to the band and was not a fanatic; he later gave Liu a thick book of the band’s album covers.
The team viewed RMB1M as money they could “never spend all of,” and encouraged itself with a crude market calculation: 10 million Official Account operators, 20% penetration and RMB10 per person per month implied RMB20M in monthly revenue. The real priority was taking the first step, not proving the ceiling.
6. 2 Years of Users Without Monetization Was the First Expensive Tuition
The RMB1M lasted roughly 1–2 years. By the end of 2018, the editor had several hundred thousand users but almost no revenue. More users brought operating, maintenance and iteration costs, so the team had to raise again to solve cash flow.
The story told to investors in the next round was “commercialization next.” Looking back, Xiao says that from the end of 2016 to the end of 2018, nearly all the team’s energy went into growth and product updates, with little systematic study of how to make money. It was a clear misallocation of time.
He admits the team had an early product-person’s “purity”: monetization was instinctively understood as harming the experience. In hindsight, that resistance was mostly immaturity. Good principles have domains of validity, while student founders operating without constraints may take longer for reality to break the habit.
7. Selling Software Directly Beat the More Sexy Advertising-Platform Story
The Official Account tool had 2 possible paths: aggregate traffic owners and advertisers and take a matchmaking commission, or charge users directly through a SaaS subscription. The first sounded like it had a higher ceiling, but the team chose the second.
Xiao’s mechanism analysis was that programmatic ad matching should be handled by Tencent. The core of non-programmatic commercial deals is business development, and there was no need to first build software to acquire those customers. Competitors were effectively doing A first and then betting on B, where they had no advantage; even whether A had to be done was questionable.
In the end, the competitor pursuing the advertising-platform route disappeared, while the company seriously selling software survived. Internally, the team used “doing one’s proper work” to align itself: they were building software, users would buy software, and the revenue would in turn motivate the product to improve.
8. A Login Pop-Up Turned a Product Manager into a Salesperson
The highest-leverage commercialization change was showing a paid plan immediately after a user registered and logged in. Traditional product thinking would provide a smooth free trial first and charge only when a feature was blocked; the Silicon Valley sales methods the team studied demanded asking for the purchase during onboarding.
The change increased revenue “many times over,” which Xiao estimates may have reached an order of magnitude. In product terms, he later explained that users are most patient at registration, and an aggregated page can show several sources of value at once, creating a fuller sense of value than blocking a single feature.
Earlier monetization also gave the team earlier positive feedback and built know-how. If the first experiment worked, the founder gained confidence to test other models. Xiao believes founders must learn not only product, but also deliberately allocate time to commercialization.
9. Revenue Turned Growth from Mysticism into Measurable Unit Economics
Unlike a network-effects platform, a tool product’s growth can be decomposed into acquisition cost, service cost and user value. If the unit economics work, spending can theoretically continue to scale without waiting for organic distribution to create a miracle.
Once the team improved monetization and internal data systems, it could measure how much revenue 1,000 purchased users generated and whether that covered acquisition and service costs. Charging, which appeared likely to hurt growth, instead created budget boundaries and repeatability.
This experience became a Day One principle for Monica: build product and monetization together, rather than accumulating users first and asking 2 years later where the revenue would come from.
10. VC Is Expensive; the Real Cost Appears When the Company Succeeds
Xiao defines capital as “one of N kinds of tools.” Advertising is a growth tool, and financing is another. Whether to use capital depends on what specific task the money can complete; fundraising itself is not company progress.
He warns founders that VC is not cheap money. If a project fails without an ethical breach, repayment is usually unnecessary. But if the company succeeds, the equity value a founder gives up may be far greater than loan interest. That asymmetry is precisely why venture capital exists.
The more dangerous cost is not financial return, but money forcing the organization to deliver on the wrong grand story. Xiao insists: “Even if what you ultimately build is smaller,” doing the right small thing is still better than using ample resources to do the wrong big thing.
11. Investors Are Not the Boss, and a Forwarded Article Is Not a Strategic Order
During his first startup, if an investor forwarded an article or mentioned a concept, Xiao would “digest it for half a day,” trying to guess the underlying intent. Later, he realized the investor might simply be sharing something casually, while a Founder’s overreaction could send many colleagues toward requirements that did not exist.
The internal standard was that “an investor’s business opinions should carry roughly the same weight as those of, say, a KOL.” If an investor truly insisted on a view, the team would hold a serious meeting, record the positions and, if necessary, take it to the board rather than leaving both sides guessing indefinitely.
Zhang Xiaojun asked whether investors should be treated as the Boss. Xiao answered clearly: “They should not.” A Founder’s decision can mobilize huge sums and redirect an entire team; the higher the leverage, the more important it is to distinguish daily information from formal governance.
12. A Small Business with Tens of Millions in Revenue Became Its Own Venture Fund
By 2019, the Official Account editor was generating close to RMB10M in revenue with good profit, but the number of operators was growing more slowly and the ceiling was becoming visible. Xiao became frustrated: “After all that work, it seems it’s just a business—a small business.”
The turning point came from recalculating the economics. The product was mature, and maintaining the experience and competitiveness required only a small product, R&D and operations team. The rest could explore new directions. Internally they asked: “If we can invest RMB10M in ourselves every year, what should we do?”
The team then tried mini-programs, photo albums, image editing, social-account data monitoring and other directions. The profitable product could continuously supply cash to new businesses, allowing the team to explore fresh opportunities.
13. The WeCom Opportunity Came from Understanding Both Customer and Platform
At the end of 2019, WeCom invited large numbers of development teams to learn about its new version. WeCom would connect with personal WeChat and open an area in the chat interface for third-party development. The team decided that day to start SCRM.
Most teams saw API restrictions: WeCom allowed far fewer marketing messages than personal-WeChat外挂, seemingly at odds with customers’ desire for “more ads, the better.” Xiao thought one step further—WeChat could not tolerate unlimited mass messaging indefinitely, or the ecosystem would inevitably collapse.
The conclusion came in 3 steps: personal-WeChat外挂 would probably be governed; as it was blocked, legitimate enterprise demand would be redirected to WeCom; waiting until governance occurred to build would mean falling at least 1 month behind. Therefore, “we not only had to do it—we had to do it immediately.”
14. When an Event Is Inevitable but Its Timing Is Unknown, Prepare as Early as Possible
Xiao summarizes the tactic this way: when facing something that is “almost certain to happen but whose timing is unknown,” prepare as early as possible at an affordable cost. The optimal state would be controlling the timing yourself, but platform entrepreneurs generally can only lie in wait.
The team initiated the project at the end of November 2019 and launched in December. Its documents specified the key trigger: once the platform began governing外挂, it would “spare no effort” to tell the market that a compliant, fully functional alternative already existed.
The profitable legacy product gave the team the right to wait. Xiao once said internally that they might be “the richest team in the sector”: a new To B company raising RMB5M might not dare burn RMB1M a month, while their own profits allowed them to commit hundreds of thousands of yuan or more decisively at the critical moment.
15. 6 Months Without Positive Feedback Made the Wrong Revenue Easiest to Mistake for Strategy
The product had almost no users during its first 6 months, and enthusiasm was quickly depleted. Xiao compares it to a system outage: when feedback is clear, everyone proactively fixes the problem; when there is no feedback, managers assign tasks every day and organizational combat power falls.
In spring 2020, a large company proposed a private deployment and custom development, and the commercial process advanced for a time. The sales team finally saw positive feedback from a large order, but the R&D work would slow the standard SaaS product, so the team ultimately apologized and withdrew.
Resisting the temptation was also possible because the old business was profitable: the custom-development revenue was not necessary for organizational survival. Without cash flow, Xiao believes the team could easily have been pulled into an order unrelated to strategy.
16. The Night the外挂 Was Shut Down, Preparation Delivered 100,000-Level Distribution
In May 2020, the team learned at 10 p.m. that personal-WeChat外挂 was being governed. By 2 a.m., it had published an article whose central message was: “It’s fine if the外挂 is blocked; switch to a compliant product on WeCom.”
For a vertical To B SaaS product, the article exceeded 100,000 reads overnight, showing that many merchants were simultaneously looking for alternatives. The team then bought placements heavily in Official Accounts followed by private-domain operators, repeatedly establishing category awareness.
They tracked “first-mention rate without prompting”—when users thought of SCRM, which name did they say first? Xiao saw the campaign as a clean realization of timing: when the model or platform’s capabilities cannot be controlled, the window must be fully captured when the trigger arrives.
17. Beyond Product, Growth and Monetization, Capital Suddenly Became a New Battlefield
SCRM growth was briefly the fastest in the category, coinciding with the 2020 financing frenzy for To B SaaS. But the team did not complete a large financing, and competitors’ resources quickly became multiples of its own.
Xiao described the frustration: after finally learning to build a team, make a product, drive growth, monetize and anticipate the market, “another new dimension called capital” appeared. He prefers battles with ample resources and did not romanticize the disadvantage of being undercapitalized.
VC financing, minority investments from mid-sized and large companies, and acquisition offers from industrial players then appeared simultaneously. Someone could prove that every option was “great,” making information overload and FOMO more painful than committing to any single view.
18. Selling the Company Was First About the Team, Then About Recognizing a Bubble
The fundamental reason for choosing an acquisition was the team. A group of people had been building a company since graduation for 5 years and needed a milestone and tangible return. 5 more years later, families and aging parents would change each person’s ability to bear prolonged low-pay risk.
The business warning came from a VC’s blessing: “We hope you become China’s Salesforce.” Xiao believed the statement was sincere, but asked why SaaS built on WeCom could carry such high expectations when WeCom itself had not become Salesforce.
He could imagine reaching roughly RMB1B in revenue, but expectations in the hundreds of billions of yuan might distort the organization. Zhang Xiaojun summarized the choice as “ending the war while things were still good.” Xiao agreed that timing remained decisive.
19. The Acquisition Did Not Change His Home, but It Corrected His View of Success
The sale did not make Xiao completely financially free. At the time of the interview, he still lived in the apartment rented during his first startup. His common-sense rule is that “it is easy to move from frugality to luxury, hard to move back,” and even small changes such as becoming accustomed to business class can raise the psychological cost of the next startup.
He credits the acquirer with letting him remain in the To B industry. If he had sold and left, he might have attributed all the success to personal ability. Living through the sharp deterioration in the capital environment from 2021 to 2022 showed him what the industry looked like after the exuberance cooled.
The downturn compressed early overconfidence back into caution. Wrong decisions made during a boom do not reveal themselves immediately; they collect the bill after the bubble bursts.
20. The First Startup’s First and Second Conclusions Were Both Timing
Xiao observed that seniors who started companies in 2013–2014 generally captured the mobile-internet dividend. He graduated only 1–2 years later, in 2015, when the major opportunities available to him had already narrowed. “Looking back, timing was still important.”
Product, growth, monetization, competition and financing are trainable skills. Their most important use, however, is judging when to market, raise, build and sell. Timing itself cannot be copied, but judgment can accumulate continuously.
The second startup used the same standard. If the team spent 6 months setting up an overseas company and learning English, it might gain “a CEO with better English” while losing a market window that already existed.
21. The Strengths of the First Startup Pre-Embedded the Weaknesses of the Second
The team was good at building complements on large platforms, finding a business inside seemingly modest ceilings and anticipating platform changes. Monica’s Day One monetization directly converted that old experience into a new discipline.
But platform dependence limits both ceiling and control. Being too good at timing can also cultivate opportunism—chasing short-term windows while resisting sustained investment in long-term capabilities.
Xiao therefore does not separate strengths and weaknesses: “Your weaknesses are often defined by your strengths.” Avoiding foundation models means giving up control over when capabilities arrive, but it also buys freedom to use models from around the world.
22. The Second Startup Initially Only Wanted to Keep “Making Good Tools for Humanity”
ZhenFund became the first and, at the time, only investor again. Liu Yuan invested all the proceeds he had earned from Xiao’s first startup into the new company, supporting the original team’s launch. Years earlier, Xiao had said that with ample resources he most wanted to “make good tools for humanity,” and Liu remembered.
In September 2022, the team read Sequoia U.S.’s article about generative AI changing the world and began systematically researching the direction.
Its first experiments involved Stable Diffusion and an anime model leaked on Twitter. It also built an image-generation app that has since been discontinued. Xiao felt images, and future video, were closer to entertainment and “killing time,” while the team’s strengths remained in tools, so it did not continue betting heavily there.
23. GPT-3 Made Natural Language Look Like a General-Purpose Software Interface for the First Time
In November 2022, Xiao tried the GPT-3 API in OpenAI Playground. He pasted sales and customer conversations into it and used natural language to request names, contact details and project stages; the model returned structured output without a dedicated NLP engineer.
He also asked the model to generate a search link for secondhand refrigerators in Wuhan and draft replies to English emails. The experiences felt “magical” and made him realize this would become a capability used in daily work, not merely entertainment content.
The key change was not the accuracy of one particular feature. It was that a product team could build, at extremely low cost, experiences that had previously required data, labeling and collaboration with algorithm engineers.
24. Jasper Exposed 3 Weaknesses That Became Monica’s Starting Point
Before ChatGPT launched, Jasper was the leading GPT-3 application. Its founder publicly celebrated expanding the team from 9 to more than 200 people in a year. Xiao instead saw rapid expansion as an anti-scale effect: the more people there are, the greater the management friction and the slower execution may become.
Jasper targeted enterprises and charged relatively high prices. Based on the simple belief that computing-industry costs would keep falling, Xiao expected a market in which individuals or professionals purchased software themselves rather than relying entirely on corporate budgets.
The team deliberately set its goal at “RMB500,000 a month.” Xiao says “lowering expectations is the source of happiness in life”: if they had initially demanded a super-platform, the project might have been canceled the moment ChatGPT appeared.
25. “5×7×0.1” Gave a Team That Had Never Gone Overseas Enough Courage
The overseas choice had a practical reason first: domestic models and regulatory conditions were not yet mature, and overseas models serving Chinese users could create regulatory risk as the scale grew.
The more intuitive internal formula was that overseas willingness to pay for software was roughly 5x, the dollar-to-RMB exchange rate roughly 7x, or “5×7=35.” Language, culture and experience reduced the team’s fighting power to 0.1, but the result “seemed to be 3x more earnings.”
Xiao admits this was not rigorous mathematics, but it helped the team reach consensus quickly. No founder had lived overseas long-term and their spoken English was poor, yet they chose to capture the opportunity first and solve the corporate entity, payments, language and local operations afterward.
26. ChatGPT Forced the Product from a Web Form into a Contextual Plugin
The project began before ChatGPT, and the team reconsidered whether to continue after its release. The low target made the answer “RMB500,000 a month should still be achievable,” but the product form had to change.
A Web-only product invited the question, “Why wouldn’t users just open ChatGPT?” Jasper-style forms also rapidly lost their advantage once conversational interaction arrived. The team therefore moved the Chatbot into the product and chose the less crowded browser-plugin format.
The plugin could be summoned at any time and automatically receive the current page’s context. While writing an email, posting to Twitter, watching YouTube or using GitHub, it could provide the relevant capability without copying and pasting between a webpage and ChatGPT. This extended the team’s first-startup experience with Chrome plugins.
Asked why it did not build an app, Xiao admitted that the team misread one window. Roughly 2–3 months before ChatGPT’s official app launched, wrapper products such as ChatOn and Ask AI did gain many users and revenue; the team lacked the bandwidth and underestimated the opportunity.
27. Indie Developers Were High-Signal Samples for Finding PMF
The team focused on indie developers because they lacked large-company budgets and strategic mandates. If a product was still alive and still had users, that was some evidence it had solved a real problem and found a distribution channel.
Early products such as Perplexity and ChatGPT for Google showed the same pattern: small teams, rapid iteration and Twitter marketing. The latter displayed ChatGPT’s answer only on the right side of Google results, but matched the user intuition of the moment extremely well.
ChatGPT for Google was built by a HUST senior, the developer known online as wong2. It grew quickly after launching on Hacker News. The team initially only wanted to buy advertising from him, then realized it could directly acquire the product and a controllable initial user base.
28. Acquiring ChatGPT for Google Bought Time, Not Code
During the 2023 Spring Festival, Xiao checked Monica’s progress every night, but his greatest anxiety was go-to-market. The product could iterate continuously, but “PMF is like mathematics: if it isn’t there, it isn’t there. You cannot solve it simply by working harder and harder.” Burying oneself in R&D would not necessarily create users.
On the first working day after the holiday, he did not go to the office but flew to Shanghai to negotiate the acquisition. For the indie developer, continuing commercialization, growth and team-building meant quitting his job and becoming a founder. For Xiao’s team, those were precisely the skills it had already built during its first startup.
The transaction’s key value was controllable initial volume at a time when AI supply was scarce and global demand was surging. Waiting for KOL replies and buying traffic channel by channel were unpredictable; the window itself could not be paused.
29. Redirecting Traffic from an Asset Made Experimentation More Aggressive Than Buying Ads
After Monica launched, ChatGPT for Google imported thousands to tens of thousands of users a day. The team used them for surveys and interviews, quickly identifying what worked, what did not and whether a demand was representative.
Xiao distinguishes 2 types of spending. Ad money disappears once spent, so managers hesitate. An acquired product remains an asset: even if a traffic experiment performs poorly, its underlying users and brand remain, allowing the team to compress the cold-start period more decisively.
He admits that without the acquisition, he does not know how the team would have completed the cold start. Other small teams built similar products during the same period, but their growth was clearly slower than Monica’s, and the window advantage quickly hardened.
30. Monica’s Name, Pricing and Revenue All Began with Friction Reduction
Many browser shortcuts were already taken, but M was available. The team wanted the product to feel like an AI assistant, so it looked for a common but not overused human name beginning with M—one people could spell after hearing it. It ultimately chose the 3-syllable Monica.
The name did not come from Friends, although many users later felt the character’s personality fit the assistant image. Xiao emphasizes that a team with weak English especially cared about a name that could be written as soon as it was spoken.
Once overseas payments and pricing were ready, the product reached its initial RMB500,000-per-month target within half a month of launch. It validated Day One monetization and gave the team rapid positive feedback instead of making it spend 2 years teaching itself how to charge.
31. 2 Independent Products Forced the Team to Learn Growth
The team did not rename ChatGPT for Google as Monica. The original product’s growth came from the search term “ChatGPT for Google,” word of mouth and an existing brand; renaming it abruptly could have wiped out organic additions.
The deeper organizational reason was that if tens of thousands of users arrived automatically every day, no one would value a colleague’s method for adding 200 users daily. Building a new brand from zero made small wins large enough to create positive feedback for the growth team.
The independent-product strategy captured the traffic dividend for AI content on Twitter while preserving a safety cushion: the products could be merged later if the experiment failed. By 2024, ChatGPT for Google had several million active users, while Monica had several million member users, and Monica had already surpassed ChatGPT for Google.
32. Serving Only Overseas Markets Concentrated Effort on the Missing Capability
As domestic models became more usable, the team repeatedly debated entering China. Its 2023 conclusion remained to focus only on overseas markets. The team’s overseas understanding was already limited; serving 2 markets would dilute Monica during its most important window.
Core users were concentrated in North America and East Asia, including the United States, Canada, Japan and South Korea. The product supported dozens of languages, most translated by AI. The result was “seemingly not bad,” but it solved only the most basic localization.
Overseas markets are not one market. Currency, exchange rates, time zones, language and aesthetics all differ. Japanese websites have far higher information density than U.S. sites; if a product does not feel locally made, American users may not interpret it as an international brand, but may instead trust it less.
33. Fragmented Markets Add Complexity but Give Startups Strategic Depth
Xiao compares China to “one large flat plain”: winners in a unified market become huge, so every team is willing to fight head-on. Overseas markets resemble multiple small hilltops, with Europe, Japan and Korea beyond the United States.
A loss in one market can be followed by focus and recovery in another language, culture or region. This fragmentation gives startups strategic depth, but also requires more granular user understanding.
He describes the first step overseas as “courage and conviction.” Corporate entities, payments and local operations are concrete problems that can be researched and discussed with friends. After attending a SaaS conference in the United States, his biggest takeaway was that early U.S. startups “were not that different,” making him more willing to compete.
34. After the 2023 Tailwind Disappeared, Growth Returned to Fundamentals
In 2023, AI application supply was scarce, and doing a handful of simple things correctly could generate growth. Later, startups, big-tech applications and comparable products multiplied, and the same actions no longer produced the same results.
Xiao believes an organization must first acknowledge that earlier success included a tailwind. Otherwise it will only complain: “This clearly worked last year—why doesn’t it work this year?” The end of a tailwind is not a sudden decline in execution; it is a change in the competitive baseline.
The new phase offered no dramatic shortcut: deepen understanding of global users, anticipate technical capabilities, strengthen the organization and processes, and improve the product step by step. He calls this “doing one’s proper work,” while admitting that he once extrapolated difficulties linearly and allowed his own state to deteriorate.
35. Monica Wanted to Be the Unified Entry Point for Models and Endpoints
The browser plugin was only the starting point. Monica later added a Web product and mobile app, while its website positioned it as an “All-in-one AI assistant.” Users could access different models across devices and workflows.
Xiao listed entry points including Chrome, the app, WhatsApp, Messenger, Telegram, Slack, VS Code and JetBrains. The core idea was to “integrate all models and appear across every ecosystem.”
Big-tech applications are usually tied to their own models, while model startups tend to serve only their own models. Neither group is likely to seriously build cross-model, cross-endpoint integration. Xiao believes that if users need something and the giants do not want to build it for strategic reasons, there is room to try—“we can discuss how big later.”
36. With a Model, You Control When Capability Arrives; Without One, You Gain Freedom of Choice
Xiao acknowledges that first-tier Foundation Model companies have enormous advantages. They not only predict when capabilities will arrive; they can decide when to release them and directly expand the boundaries of human ability.
Monica’s decision not to build a foundation model on Day One was primarily resource-constrained, not a matter of philosophical purity. An application team asks how much existing models have improved and how to turn that improvement into better service for users.
It is a two-sided trade-off. Without a model, the team lacks control over timing but can use any model in the world. With a model, it has more control but often must prioritize its own capability.
After Pica, Mem0 and Supermemory joined at the beginning of the year, the team also trained several “small and economically practical” models. Xiao does not rule out going further into models, but only if resources are available, user demand is clear and external vendors cannot solve the problem better.
37. “Trade-First, Industry-Second, Technology-Third” Offers a Path from Product to Technology
A senior industry figure used “technology-trade-industry” and “trade-industry-technology” to describe competing routes. Outsiders often see Huawei as technology-first, but the senior believed it also started by making the business work, then genuinely strengthened engineering and technology.
Xiao used the framework to understand his own path: make the product and business viable first, then bring necessary technology in-house once models become easier to access and task requirements clearer.
This does not deny the value of model companies. As the industry matures, it may stratify. Applications and models are not necessarily an either-or choice; the answer should change as technical boundaries and organizational resources change.
38. Brand May Outlast a Temporary Model-Performance Advantage
Xiao believes the technology industry has long underestimated brand. DeepL built strong awareness through translation that was better than Google Translate in its early years; even if LLMs can now do better, DeepL continues to grow in the new AI wave.
Grammarly even uses OpenAI models directly, while retaining the user perception that it is the best tool for writing and grammar. What truly compounds may not be one generation’s algorithmic edge, but which product users think of first when they have a task.
Zhang Xiaojun connected the Founder’s podcast appearances to brand-building. Xiao agreed that brand is one reason a product can outperform others, but did not define podcast appearances as a moat by themselves.
39. AI Applications Look More Like Consumer Electronics Businesses
The team compares Monica to consumer electronics. Its website identifies ChatGPT, much like a phone highlights Snapdragon 888 or “Intel Inside.” The core model is a critical chip, but not every task needs the most expensive chip.
Specific scenarios can call small models, proprietary models or open-source models. Access to upstream models also resembles a supply chain: when GPT-4 was scarce, the team had to speak with Microsoft, promise more consumption and negotiate earlier quota.
Front-end features are easy to imitate, so differentiation shifts toward product design, brand and go-to-market. Kimi’s use of niche Bilibili creators is an example of translating the same capability into language a particular audience understands.
Model calls generate recurring costs, making an AI product more like an “atomic business” than a near-zero-marginal-cost “bit business.” Subscription pricing, gross-margin management and supply chain are all part of the product definition.
40. Foundation Models Are Putting Small Teams Back on the Board
Xiao uses calorie estimation from a photograph as an example. Previously, it might have required data collection, labeling and an organized algorithm team. Today, a multimodal Foundation Model lets 1 or 2 people build a usable experience, even if accuracy is still imperfect.
He cites a framework with 4 types of players: large companies have capital, distribution and ecosystems; model startups such as OpenAI and Anthropic enjoy the technical dividend; indie developers can rapidly create “magic” through creativity; the hardest position is the mid-sized company with neither abundant resources nor strong revenue.
Monica positions itself as “a large-scale indie developer”: bigger in headcount, but still focused on users, rapid experimentation and trusting external models rather than imitating big tech through organizational scale.
41. AI Business Models Remain Traditional, and APIs Lack Cloud-Style Lock-In
Xiao believes no surprising business model had emerged yet. ChatGPT and Perplexity were mainly subscription products, with Perplexity beginning to test ads; it was still unclear whether domestic products could generate sufficient revenue from memberships.
He does not consider Foundation Model APIs an especially good business. Cloud migration is painful and creates natural lock-in; model APIs are easy to replace. Monica switched from GPT-4 to Claude in roughly 1 day of testing, which the team called a “second-level switch.”
Suppliers therefore risk falling into price-performance competition. If perceptible model differences narrow, speed and cost become the main variables, benefiting model-agnostic applications. Xiao still hopes the industry does not “hit a wall,” but keeps creating real user value.
42. OpenAI and Anthropic Show 2 Top-Student Routes for Model Companies
Xiao believes ChatGPT’s consumer value remains underestimated. OpenAI’s valuation far exceeds Anthropic’s largely because ChatGPT has completed a massive advertising campaign in the minds of users worldwide.
Citing the revenue mix at the time, he said OpenAI’s consumer subscriptions were significantly larger than its API and enterprise revenue, while Anthropic might derive roughly 80% from APIs and 20% from Claude subscriptions. The former takes an end-to-end route from model to application; the latter relies on a sufficiently strong model serving many enterprises and application companies.
Within Monica’s usage, Claude 3.5 exceeded GPT-4o, and the team considered its performance excellent. Xiao concluded that model companies could at least choose between deepening a consumer application and making an API third parties cannot avoid.
Llama’s continued open source and Gemini’s big-tech ecosystem will also persist. Xiao believes there will not be infinitely many model companies, but did not claim that domestic and overseas structures must converge.
43. The First Class of AI Applications Grows in Gaps Left by the Main Scenario
Xiao divides the applications of the past 2 years into 3 categories. The first is main-scenario supplementation. Early ChatGPT lacked browsing, and attaching a search API with citations was not difficult, giving Perplexity a window of more than 1 year.
Perplexity later went beyond answering searches and tried to produce content directly on its own platform, resembling Xiaohongshu’s combination of content creation and search. Capital and its user base let it expand the boundary after filling the original gap.
Monica also belongs to this category. A browser plugin naturally carries page context, filling the gap created by the official Chatbot’s need for copying and pasting and its difficulty embedding into workflows.
Xiao warns that such opportunities may shrink. Once OpenAI and other model providers seriously complete Chatbot and browsing capabilities, a window based purely on the leader’s neglect will not last forever.
44. The Second Class Turns Model Capability into a Shareable Visual Format
Images and video more readily produce the “model as application” pattern. A new model first creates an effect that did not previously exist; the product then packages it as a social-media format users can imitate, share and distribute.
Xiao cites Pika’s “Crush It,” where users tried the product because of the novel effect, and Viggle, which makes one character imitate another’s movements. The generated content is naturally suited to TikTok distribution.
The causal chain is model capability leap → new format → social distribution → users pouring in. Product value is more directly tied to progress in the underlying modality.
45. The Third Class Uses a Container to Catch Capability Spillover into Vertical Fields
Cursor represents the third category. Foundation Models can already write fairly good code, but the capability lacked a container that could handle files, diffs, execution and errors.
Xiao once conversed with a model inside Monica, manually created folders, copied code and compared differences, and spent several days building a Chrome plugin. Only later did he realize that all of this was engineerable edge work, which Cursor Composer had automated.
He calls this a “huge mistake.” The hardest part—writing code—had already been solved by general models such as Claude 3.5 Sonnet. An application company only needed to put the remaining workflow into the right shell to create a new vertical product.
46. Cursor’s Distribution Was Lit by Outsiders, Then Taught Insiders in Reverse
Cursor first excited product managers, operators and technology KOLs who normally did not write engineering code. For the first time, they could build small tools themselves, so they recorded screens, posted on Twitter and recommended the product.
Once distribution reached engineers, professional users discovered capabilities such as completion. Some had long existed in VS Code and GitHub Copilot, but Cursor’s interaction and brand expression made them visible.
Xiao recounts an internal conversation: an engineer marveled at a Cursor feature, and the CTO replied that the VS Code plugin had it too. The chain suggests that senior users of a vertical product may underestimate the leap perceived by users at the edge.
47. Product Transition Should Not Be Decided Directly by Macro Fear
Zhang Xiaojun asked whether Monica might die during a transition in product form. Xiao answered that companies usually die because “the money runs out” or “they no longer want to do it.” Neither was happening, and current growth and revenue proved the product had at least gotten some things right.
The real distinction is between 2 errors: underestimating a demand that could have become much larger because of insufficient understanding, or mistaking a temporarily correct window for a permanent truth.
The solution is not a vague judgment about whether “plugins will disappear,” but continuous observation of specific user needs and a Founder who can adjust. Macro conclusions made too early can obscure verifiable facts.
Xiao also rejects the premise that one can “build commercial products but not necessarily a big company.” No successful entrepreneur knows on Day One that the business must become large; “obsession with being big can itself be a problem.”
48. From Jasper to Agent, Each Model Generation Requires a New Shell
In the second interview after the 2025 Spring Festival, Xiao connected a product-evolution line: Jasper used fill-in forms; ChatGPT shifted to natural conversation; Monica and Doubao added context; Cursor went further by writing code and directly executing requests.
The line moves in 2 directions at once: interaction becomes more natural for ordinary people, while capability becomes stronger. Cursor was already viable, but did not truly explode until Claude 3.5 Sonnet appeared in July or August 2024, showing that model capability can arrive first and product PMF then jump suddenly.
Xiao summarized it this way: “Model capability is evolving rapidly, but the shell itself also needs to evolve.” Model providers may not immediately expose the full value of new capabilities; third-party applications can translate them into experiences ordinary users can feel.
49. A “New-Era Andy–Bill Law” Describes the Chase Between Models and Applications
The PC-era Andy–Bill Law held that new computing power supplied by Intel would be consumed by the next version of Windows. Xiao applies the analogy to LLMs: no matter how cheap or powerful models become, the application layer will consume the capability through more complex interaction, tools and tasks.
Models will progress from writing, Q&A and retrieval to writing code, calling APIs and operating browsers. Applications will not simply enjoy lower prices; they will “ruthlessly consume huge quantities of tokens” by turning capability into longer execution chains.
The semiconductor industry offers another analogy of specialization. Without TSMC, the division between specialized manufacturing and fabless design would not have formed. Models and applications may follow the same path: vertical integration first, then gradual specialization as the industry matures.
50. Model Providers Will Not Build Every Application, and Players Can Change the Window
Nearly every model company is building both a Chatbot and an open platform. The Chatbot matches the public’s image of AGI as “one sentence that can do everything,” so it appears to be an unavoidable entry point for model providers.
But AGI also has public-good characteristics. Model companies cannot cover every industry, long-tail workflow and “dirty, tedious work.” Specific industries require substantial engineering and operations, which may be better left to the ecosystem while the model provider charges for API access.
The window is even more complex because it has no fixed endpoint. The provider may eventually build the product, or may see a third-party leader become successful enough and choose not to enter. Founders can be wrong to assume either “it will definitely come” or “it will never come.”
Xiao uses WeCom SCRM as an analogy. The platform could logically build it itself, but if an ecosystem company does it well enough first, the platform may decide to leave it to the ecosystem. That is the game in which players change the endgame.
51. DeepSeek Made Reasoning a Publicly Visible Experience
Xiao believes DeepSeek’s key product innovation was displaying the model’s thinking process. Perplexity increased credibility with webpage citations; DeepSeek let ordinary users directly see “how it thinks.”
OpenAI’s o1 adopted the Reasoning path earlier, but had a paywall, showed only a simplified reasoning process and could not browse at the time. Zhang Xiaojun admitted that he knew it represented a new technical route but could barely feel an improvement in user experience.
DeepSeek also browsed the web, and domestic users felt its outputs were more detailed. Xiao observed that people once assumed users preferred short answers; after Chatbots became excessively summarized, users were instead willing to watch them “say a little more.”
Underlying quality mattered as well. DeepSeek Reasoner lifted many domestic users from an ordinary baseline to first-tier quality, with more detailed writing and more “emotional intelligence” in conversation. A technical leap combined with visible design was what created a generational experience.
52. DeepSeek’s Hardest-to-Copy Advantage Was Doing Things at Its Own Pace for Years
Xiao believes DeepSeek’s app only appeared around December 2024. Its form was simple, and the team did not seem to operate it as a super-app. But without that “own shell,” ordinary users worldwide might not have directly experienced the model’s progress.
DeepSeek had remained open source before it became famous and remained open source when it was not famous. After its breakout, other companies began reconsidering open source, technical branding and consumer products. Xiao believes the most valuable lesson is not mechanically copying open source, but “be yourself.”
DeepSeek paid nothing for KOL distribution yet reached No. 1 on app-store charts in more than 100 countries. For an application founder who lived through the mobile internet, it was like seeing a 100-meter runner break 10 seconds: proof that the wildest product distribution can still happen.
As for whether it will continue toward a Super App or how it will monetize, Xiao said clearly, “I’m not sure.” A strong position today does not mean the organization’s goal is fixed; OpenAI’s path is also closely tied to the choices of its CEO.
53. A Real Agent Is Not an API Feature Phone but Something That Finds Its Own Tools
The Agent concept spread widely through projects such as AutoGPT in 2023, but remained below expectations for a long time because models were not smart enough to complete multi-step tasks reliably.
Monica had previously connected APIs manually for each demand: search, knowledge bases, files, PPT, image generation and more. Xiao increasingly felt this resembled a Feature Phone: “The user needs a function, so add another function.”
The ideal Agent should plan, write code, call APIs and process feedback on its own, without developers enumerating every function in advance. A senior figure, Bai Ya, gave Xiao the turning point: “Extremity is not enough; personalization is enough.”
Cursor, Bolt.new, Lovable and Windsurf have made code execution increasingly accessible to non-engineers. Senior engineers can still identify problems, but ordinary users keep experiencing more delight, showing that the foundational capabilities for general Agents are maturing.
54. Windsurf’s Automatic Error Correction Created a “Lightning Strike” in Tool Use
When Xiao tested Windsurf’s YOLO-like mode, a missing library or code error was automatically returned to the model for further handling. On the show, he explained YOLO as “You Only Look Once,” emphasizing that the model could continue processing errors of this kind.
The more shocking moment was the model proactively saying it would download code from GitHub and continue the task. GitHub stores a vast number of tools created by humanity; once a model can find and combine them, its capability boundary is no longer the API list preinstalled by the developer.
Xiao described the moment as “being struck by lightning”: the Agent was finally not just answering, but using the outside world, accepting environmental feedback and continuing to act.
55. Cloud VMs, Browsers, Code and APIs Form the Agent Foundation
Local programming tools still ask whether to execute commands or install libraries, forcing users to click yes or no. Xiao questions this form of safety outsourcing: true beginners do not understand the risks, so confirmation provides no meaningful judgment.
He therefore imagined placing the execution environment inside a cloud VM. Code, commands and browser operations would run on an isolated computer; once the task was complete, the instance could be released, and a damaged environment rebuilt without endangering the user’s local machine.
His complete architecture consists of a virtual server, a browser and a model capable of writing code and calling APIs on its own. The browser covers knowledge and services with no API that can only be accessed through the Web; code connects both existing and temporary tools.
Xiao believes this architecture can handle a large number of long-tail tasks and is closer to a mass-market Agent than Devin, which was designed mainly for engineers.
56. Agent Interaction Must Shift from Synchronous Q&A to Asynchronous Collaboration
A traditional Chatbot follows the waterfall pattern “A–B–A–B”: the user sends one sentence and the model replies with one. New input during a reply may interrupt the previous answer. Complex tasks require waiting, branching, retries and mid-course corrections.
The new product first breaks the request into a Planner and tells the user which steps it intends to complete. It reports progress continuously during execution; if the user sees the direction is wrong, they can add information or request a redo, and the system adjusts the final deliverable.
Xiao compares it to “the best possible intern.” The intern will not answer every question instantly, but will explain the plan, research, and synchronize results step by step. The user does not need to remain in the chat window the entire time.
The new product was separated from Monica because the original product already had synchronous usage habits. A “no-burden” entry point was better suited to testing the new assumptions around asynchronous tasks and consumer Agents.
57. Multi-Step Tasks Let Users Touch the Model’s Hidden Capability Boundary for the First Time
In the GAIA test, one question required identifying the animal shown at a specified time in a YouTube video. The team found that the Agent actually opened the video, used YouTube keyboard shortcuts to jump to the exact point and answered based on the image rather than the subtitles.
Another task involved scraping all of Elon Musk’s posts and analyzing patterns semantically. The Agent could call the Twitter API, process the data and output a report. Such work exceeds traditional chat but remains a task users can describe naturally.
Xiao does not believe search and Q&A will disappear because of Agents. The new form simply pushes the Chatbot’s boundary outward, turning tasks that once required copying and pasting, running scripts and asking multiple follow-ups into a single delegation.
The model does not stop after completing one step. It encounters errors, feeds the feedback back as input and continues trying. This execution-feedback loop is what makes an Agent more human-like than a pure Reasoner.
58. Models Had Just Crossed the Usability Line, but Agentic Capability Was Still Uneven
During testing, Xiao said that at the time Claude 3.5 Sonnet was the only model in the world showing long-horizon planning and step-by-step problem solving: it would first list 1, 2, 3 and 4, then feed the result of the previous step into the next.
Traditional Chatbots are aligned to solve as much as possible in one round. Agents require different data and training methods to learn planning, tool use, environmental observation and retries. Xiao emphasizes that the detailed training know-how belongs to Foundation Model companies; he was only observing differences from the application side.
He judged that model capability had become “just good enough” in late 2024 and early 2025 for the team to build a basic version, but success rates were still too low. Products such as Operator also showed that models needed to become smarter and cheaper.
59. Consumer Positioning Must Work in Both Price and Appearance
The new product should still look like an ordinary Chatbot rather than a programming tool. Under the hood it can write substantial amounts of code, but users should see natural-language tasks, plans, progress and results.
Xiao did not want to copy Devin’s engineer-focused positioning at roughly $500, and saw OpenAI’s roughly $200 plan as another opening for a lower-cost product. “When a company starts charging a very high price, you can always beat it with a lower price.”
The base price determines whether the product is a mass-market consumer good or an enterprise tool. Heavy users paying more because of real compute costs is understandable, but the entry threshold cannot exclude ordinary people first.
The team planned a small-scale test using concrete examples to control expectations. Xiao does not assume the product “can do everything,” because one of a product manager’s important responsibilities is telling users where the current boundary lies.
60. 3 Spring Festivals Showed That AI Windows Always Bring Surprises
During the 2023 Spring Festival, the team had already started Monica before ChatGPT launched. ChatGPT’s sudden appearance made the team even more anxious; it held meetings through the holiday and acquired ChatGPT for Google immediately afterward.
The surprise of the 2024 Spring Festival was Sora, but the rest of the year looked more like a linear extrapolation from 2023. Multimodality, images and video were impressive, but the direction was imaginable, so the team mainly continued building Monica.
In 2025, the team had planned to work overtime only on the Agent, when DeepSeek became another phenomenal event. Xiao’s summary was: “AI will not let any Spring Festival pass quietly.”
He distinguishes predictable from unknown dimensions. Once context, multimodality or Agents are introduced, one can estimate quantitative growth; the truly important leaps often come from dimensions that have not yet been proposed.
61. Once Leadership Is Confirmed, Operating Discipline Must Switch to Offensive Discipline
Xiao admits Monica was “not aggressive enough” in 2023. Even though the second startup had improved recruitment and organizational maturity, the team expanded as if it were running a well-managed company, possibly missing the window to hire at scale and capture users.
When difficulties arrived in 2024, he again extrapolated them linearly and assumed things would continue to be hard. Looking back after technology jumped again, the old state represented only the absence of a new dimension, not the future.
The new rule is: without a tailwind, make users, product and organization solid; once innovation and leadership are confirmed, move faster and more aggressively. “When big tech can understand your innovation, that is when it becomes dangerous.”
62. Entrepreneurial Pace Comes from Believing in One Thing, Not Chasing Every Hot Topic
Xiao describes the ideal mindset as remaining optimistic about technical progress, reducing attachment to gains and losses, and moving at one’s own pace. Models continue to create tailwinds, so today remains a good time to start a company.
He rarely talks about “passion,” but uses an extreme thought to explain his commitment. When exhausted while driving, he imagines that if he had an accident, he could make only one call to his co-founder and would say: “Don’t worry about me. Keep moving forward.” At that moment, the thing itself matters more than personal gain or loss.
This commitment cannot be manufactured by a capital narrative. It must come from the excitement of a product suddenly completing a task beyond expectations. When the team saw the Agent operate YouTube and call GitHub, it felt like “we were really creating something like a life.”
The workload remains brutal. Xiao and his co-founder often work from 9 a.m. to 11 p.m. His answer on work-life balance is: “There is only work, no life, and certainly no balance.”
63. The World Is Not Logic; It Is a Game in Which Players Rewrite One Another
Xiao distinguishes logical deduction from game thinking. If Baidu has the best algorithm engineers, static logic would imply that Baidu must win recommendation. In reality, a new player can arrive and change every participant’s strategy.
DeepSeek’s open source and global distribution pushed Sam Altman to publicly reconsider whether not open-sourcing had been a mistake. This could not have been linearly derived from any single company’s resources; it was the result of a new player changing the environment.
He even imagines that if ChatGPT had first been built by a third-party application company, OpenAI might have chosen a pure-platform route. Conversely, an ecosystem product successful enough might cause the model provider to abandon entering the market itself.
The entrepreneur’s 3 steps are therefore: believe the world is dynamic, try to make yourself an important variable, and achieve it through “be yourself”—doing what you believe in well enough.
64. The Age of the Era Is a Better Bet Than the Founder’s Biological Age
Speaking to high school students and university hackers, Xiao shared: “Think in the age of the era, not the age of your biology.” If AI’s age is counted from late 2022 or early 2023, then 2025 is only AI’s 2nd or 3rd year.
When he graduated in 2015, he did not realize that the mobile internet was nearing its end. Founders who started 2–3 years earlier captured a much larger dividend. This time, he did not want personal age to obscure how early the industry still was.
The organizational goal is not to write “Super App” first. It is to let a group of partners ride the technology dividend and enjoy the process. Commercial returns depend on too many external variables; learning, technical progress and completing products are more controllable.
65. Chinese AI Founders Should Compete More Aggressively in Global Markets
Xiao sees the present as the intersection of 2 main lines: AI and globalization. He is not a geopolitics expert, but believes all sides appear to be becoming more conservative and isolationist. Against that backdrop, Chinese founders should enter global markets and gain experience.
No one on the founding team had lived overseas long-term, and their English may have peaked in high school. If Xiao were compared with a Founder who had lived in the United States, he admits he might initially choose the latter. But real competition should not stop at résumé comparisons; it should be decided by who builds the product.
His simple judgment remains that “the global market is larger.” The market can pay the founder’s tuition. Overseas offices, language and local user understanding all need to be built, but a founder should not spend 6 months proving readiness before a huge opportunity arrives.
66. The Most Extreme AI Possibility Is to Never Short a Capability That Has Not Yet Appeared
Xiao does not know where stacking more compute will ultimately take models, but believes one cannot assume progress will slow over the next few years. When Pichai was asked at an Nvidia customer appreciation event what might surprise Jensen Huang, the answer relayed was “basically nothing”—any breakthrough could happen.
His slogan is: “Civilize the mind, toughen the body, and leave the rest to AI.” Reasoning is still mainly thinking; an Agent can execute, use tools and incorporate environmental feedback into the next step, bringing it closer to human action.
He even believes the white-collar lifestyle “may be a detour in human history.” Humans spending long hours sitting, performing intensive mental work and exercising little is only a roughly century-old abnormal state.
If AI takes on a large share of intellectual labor, white-collar working hours may shrink substantially and people may return to physical health and spiritual life. What society does with the released time will have to be answered collectively once the change occurs at scale.
67. If You Cannot Control Geopolitics, Make the Days You Can Control Solid
The second startup made Xiao more accepting of what lies outside his control. Model release timing, geopolitics and competitors’ actions are inputs, not variables he can schedule. A Founder can make the product, organization and judgment solid.
He uses DeepSeek to explain “greed, anger and delusion.” Greed is attachment to favorable conditions; anger is dissatisfaction with adverse conditions; delusion is ignorance of reality. Founders easily want to return to a tailwind while resenting a difficult period, and both attachments distort action.
His ideal day is not luxurious: read in the morning, handle work after colleagues wake up, then continue learning. The Spring Festival is enjoyable precisely because external communication declines, leaving more time to focus on controllable tasks.
In the rapid-fire questions, he chose hot-dry noodles, the company office, The World’s Great Age of Stars and The Way to Happiness: the first gives people energy but also more anxiety, while the latter helps bring anxiety down. His final advice to new founders is to first build complex vertical workflows well, then wait for the model’s “magic” to arrive.