116. Wu Minghui on 19 Years, Enterprise Agents, and the IPO

2025/10/09 · Source link · AI Summary archive

Summary

Minglue Technology’s IPO is not a linear-growth story, but a cash-flow repair after 19 years of repeated mergers, breakups and near-starvation. The seemingly 27 funding rounds actually combine the histories of 秒针, Minglue Data and acquired companies; cumulative funding exceeded $1B, the latest valuation was $1.5B, and the company is now profitable. At its most dangerous point, it had about RMB100M on the books while spending RMB100M a month—just “one month of runway.” Wu Minghui says the past 3 years turned him “from a pure idealist founder into a founder who knows how to run a business.”
Minglue is betting its next phase on enterprise Agentic Models, not general-purpose foundation models. Deep Miner uses a foundation agent, MOA and specialized models. Wu Minghui says its 7B model scored about 44 on the relevant Specialized leaderboard, ahead of Kimi in second and third at about 34, and ranked fourth overall on OSWorld—at pass@1. The full model-and-application stack is currently being built by about 40 people. After the IPO, the company will have more room to invest in AI R&D; Wu Minghui repeatedly cites past constraints on compute and GPUs.
His view of the core moat has shifted from code volume to vertical data that is unavailable publicly and compounds over time. He summarizes it as (y=f(x)): once general-purpose models make (f) increasingly similar, competition among companies will move to (x)—private context, historical sequences and real operating environments. Anyone can scrape a webpage snapshot, but nobody can retroactively reconstruct years of daily KOL follower changes, advertising behavior or store-level people-goods-place data. “Whoever has the better context and the better data” earns the durable premium.
If general-purpose models mainly monetize by selling tokens, Wu Minghui expects the market to “compete itself down to electricity bills”; the real value lies in the loop formed by models, data and scenarios. Public data can train multiple companies, and competitors can quickly catch up through question-answering and distillation. Minglue therefore trains only 7B, 8B and other specialized models, mainly for its own agents. Once deployed in an industry, real tasks, expert labels and feedback environments can improve today’s roughly “30-point” L3 capability to a “90” in specific scenarios.
AI’s deepest rewrite of enterprise software is not adding another chat box, but rewriting the connections and entry points between software products. Many legacy IT systems have no API, while external systems such as tax platforms will not open MCP. BUA and CUA let agents operate webpages and computers directly. Auto-filling CRM from recordings and automating expense claims from photos are only the starting point; a recruiting agent could execute tasks across Zhaopin, BOSS Zhipin and 51job. “It creates a new kind of link, even a new kind of production relationship,” and makes post-acquisition product synergies far easier than before.
Wu Minghui’s view of scaling laws is close to contrarian: a general-purpose model cannot economically handle every specialized task. His evidence is that a general multimodal model can lose to a tiny specialized model even on OCR, because “describing what you see” and recognizing text character by character require different processing. Just as elevator cables in a skyscraper eventually spend most of their weight pulling themselves, a model becomes burdened by excessive “basal metabolism” once it grows too large. The answer is not infinite scaling, but division of labor among expert models and agents, with MOA assembling “the strongest board in every section.”
Minglue’s most expensive lesson was EIP in 2020: its direction was close to ChatGPT and Microsoft Copilot, but it arrived a generation too early. Wu Minghui at one point organized 1,000 people to build applications, data integration and BERT-based task-specific models simultaneously, while imposing an unsuitable IPD process. The problem was that “when L1 was at 30 points, I was doing L3 work.” A sudden capital freeze, Shanghai’s lockdown and interrupted collections eventually forced mass layoffs. His postmortem: the battlefield was too broad and his management understanding too weak. “Even if we had raised another $500M, we might still have died. Dying earlier would have been fine.”
In his vision of the future, companies become multi-agent organizations where humans retain responsibility, standards and taste while agents handle most execution. Business owners control the company’s memory and benchmarks. External experts plug in agents with identity, credit histories and IP as “partners,” and share in operating results. Fully autonomous agents may actually fail to access the tacit knowledge inside human brains, so Deep Miner supports both autonomous and human-in-the-loop modes. “The test isn’t of the student, but of the agent they trained”—the same thesis he applies to the restructuring of education, organizations and productivity.

Deep dive

1. Nineteen years of company history ultimately came down to survival through profitability and an IPO

The episode takes place on the eve of Minglue Technology’s listing: the company received its overseas issuance and listing filing notice on August 29, 2025. Wu Minghui calls the listing “a new milestone,” first and foremost because it addresses the extremely tight funding position of the past 2 to 3 years.
The quick-fire profile is specific: the company has 19 years of history, a latest valuation of about $1.5B, and is already profitable. Wu Minghui is 43, his MBTI is INTJ, and he simultaneously serves as founder, CEO and CTO.
He does not frame the IPO as an endpoint, but as the starting point for another round of R&D investment. After listing, the company will have more ample funding to invest in AI; in the past, it was particularly short of GPUs and compute.

2. The supposed 27 funding rounds are really the combined history of 3 companies

Wu Minghui explains that 秒针 itself raised multiple rounds; Minglue Data, established around 2013-2014, had an independent financing history; and 秒针 also acquired competitors, so the prospectus presents several histories together.
The company continued raising money after the merger, so “27 rounds” does not mean a single entity completed 27 sequential rounds. By his estimate, cumulative funding reached more than $1B.
This complex capital history corresponds to an equally complex organizational history: 秒针, Minglue Data and AdMaster grew separately before merging into a group in 2019.

3. Li Guangmi found Minglue Data while researching Palantir

Wu Minghui met Li Guangmi while working on Minglue Data, not during the 秒针 phase. At the time, the company believed its To B business could support itself through revenue; it had quietly raised only one angel round and was not actively fundraising.
Li Guangmi was researching the U.S. big-data market at Sequoia. After seeing Palantir achieve a high valuation, he began looking for “which Chinese company was most like Palantir” and proactively sourced Minglue.
The two hit it off, and Sequoia ultimately invested. The episode shows that Wu Minghui’s second startup was initially driven by technology and market research, not created by capital.

4. Acquiring Xiao Hong was about conversational data, not a CRM

Xiao Hong was introduced by Zhou Li, Wu Minghui’s Peking University classmate and founder of Answer AI. After their first extended conversation at Wu’s home, he immediately concluded, “I definitely had to invest in her.”
Xiao Hong’s company Nightingale owned Weiban, a social CRM on WeCom with a large installed base, containing conversational data between customers and sales or customer-service teams.
Others in the market also wanted to invest in or acquire Nightingale. Wu Minghui’s differentiated pitch was: “I won’t invest in you—I’ll acquire you directly.” He was interested not in the application itself, but in the higher ceiling offered by using the data to train AI.
The acquisition took place around the end of 2020. Even after massive subsequent investment, layoffs and Xiao Hong’s departure, Wu Minghui says clearly that he would do it again, because the data and team were top-tier assets.

5. The pandemic turned online work into an enterprise AI data gateway

Wu Minghui observed that online-work tools such as Zoom had existed for years before the pandemic but were underused. Tencent Meeting’s original target for 2020 was, in his recollection, about 50,000 daily active users; after the pandemic, it quickly surged into the tens of millions.
Documents, meetings and WeCom systems consequently generated huge volumes of conversational data, coinciding with the direction of his 2019 engineering PhD research at Peking University: conversational intelligence.
His inference was direct: if a company saves all its documents, meetings and chats and uses them for AI, that AI “will definitely be better than most people in the company.”
Weiban’s data thus formed a loop with his doctoral research and showed him an enterprise AI opportunity in office scenarios that were relatively standardized and offered continuous feedback.

6. In 2015 and 2020, Wu Minghui crossed two identity thresholds

Around 2015, when Minglue Data had just been founded, he truly shifted from CTO to CEO. Before that, at 秒针, he had long viewed himself as the technical lead rather than the operator.
By 2020, 秒针, Minglue Data and AdMaster had become Minglue Group. He was thinking simultaneously about models, data platforms and office applications, with organizational ambitions far larger than 5 years earlier.
After Xiao Hong joined, he even hoped she would become his CEO successor. Wu Minghui says she had high emotional intelligence and would not reject him directly; she later told him explicitly, “Brother Hui, I can’t be the boss of a To B company. I definitely want to do To C.”

7. Wu Minghui’s first company was itself a merger startup

In 2006, while a second-year master’s student at Peking University, Wu Minghui led fellow students in software outsourcing. The editor-in-chief of Sina Education introduced him to Li Feng, who was looking for a technical co-founder, and the two companies quickly merged.
Li Feng initially wanted to build online education, while Wu Minghui had a pure-tech team. After the merger, Li Feng handled fundraising, and Wu Minghui learned for the first time that one could “not earn money from customers, but raise money from VCs to build for the future.”
In the first financing, the company was valued at about RMB3M; raising RMB1M meant giving up roughly one-third of the company. In the initial structure, Li Feng held about 40% and Wu Minghui about 30%.
After angel investors including Zhu Wei came in, some shareholders held their stakes through the eve of the IPO. Because of Li Feng’s New Oriental connections, Luo Yonghao and Li Xiaolai were also early shareholders.

8. Web 2.0 infrastructure made recommendation systems his original technical ideal

Wu Minghui’s team provided Web 2.0 infrastructure for Sina and Sohu. Web 1 content was identical for everyone and could be handled with layered caching; once users had IDs and saw different content, system complexity rose sharply.
The team built the ability to serve personalized content in scenarios such as BBS forums. When fundraising, it described the technology as an adaptive-learning system that “recommends different textbooks to different people.”
After investors rejected the online-education narrative, recommendation systems became the first formal business plan. Advertising analytics was initially just the first step toward acquiring data, followed by advertising and general-information recommendations.
He later admitted that after making money from advertising-data analytics, the company “forgot its original purpose” and failed to restart the original recommendation-system path when mobile internet arrived.

9. His master’s thesis had already applied language models to recommendations

In his first and second years of graduate school, Wu Minghui studied computer vision, fingerprints and palmprints. In his third year, he switched to recommendation systems for the company. His thesis was titled “Recommendation System Based on Language Model”—though there was no “large” at the time.
Related research by Peking University’s computer science and Chinese-language departments used the Chinese Concept Dictionary, CCD, to connect words and word senses: polysemy links one word to multiple meanings, while synonyms link multiple words to the same meaning.
Traditional recommendations mainly relied on “people who viewed this product also viewed.” His thesis went further by calculating webpage-content similarity, allowing users who viewed A and users who viewed similar B to enter the same recommendation relationship.
The earliest real user was Luo Yonghao’s Bullog.cn. The product predated Toutiao by many years, but lacked the stable identity and mobile infrastructure that did not yet exist.

10. Recommendation systems arrived too early; what was missing was a stable ID, not an algorithm

When he discussed recommendations in 2007, the iPhone had not yet launched. PC internet had no default stable unique ID, while new websites struggled to persuade users to register, making it impossible to accumulate personal historical behavior.
Cookies could tag users temporarily, but software such as 360 often cleared them, breaking the historical chain. Wu Minghui’s summary: “The infra wasn’t ready.”
Recommendation systems only truly worked once mobile devices supplied stable identities. By 2011-2012, however, 秒针’s advertising business was already doing well, and he did not return to the original path.
He believes that if he had restarted recommendations then, 秒针’s data advantage might have exceeded Toutiao’s. Two or 3 years later, competitors had already dominated and the window was closed.

11. Li Feng’s departure showed that complementarity matters more than having “two top students”

Li Feng left in 2007 not because of conflict, but because the company had moved away from education and he felt his contribution as CEO was insufficient. He preferred to enter the capital markets.
He retained part of his stake and transferred another portion to later partners, making Wu Minghui the largest shareholder, while further dispersing the company’s ownership.
Wu Minghui’s view of “top-student partnerships” is that people in the same field may look down on each other, while people in different fields recognize each other’s strengths. Hardware, software and supply chains each have their own geniuses, and people are more willing to trust areas they do not understand.
He and Li Feng were sufficiently complementary, but later “the direction Li Feng understood best was no longer necessarily valuable to the company,” so the partnership ended naturally.

12. Peking University’s mathematics department taught a lifelong top-ranked student to compete through differentiation

Wu Minghui began entering mathematics competitions in fifth grade. A teacher taught him a semester’s material in a week, giving him his first sense that the school timetable could be compressed dramatically.
He finished high-school mathematics and sciences in middle school. In high school he barely attended regular classes, studied with a laboratory key and often cycled to the city library to read mathematics books.
After the mathematics competition, he planned to attend Tsinghua for computer science. A Peking University admissions teacher promised that if the top 3 students from Shandong all chose Peking University, the other 2 teammates would receive offers, so he chose mathematics at Peking University.
After entering the department known as the foremost of “Peking University’s 4 madhouses,” he went from being a routinely outstanding student to ranking around 20th or 30th. He then turned to entrepreneurship, making money and coding: “I couldn’t beat you at math anymore, so I started a business and made money.”

13. Olympiad training made failure an everyday condition

Wu Minghui acknowledges that entering Peking University was a setback, but says he adjusted quickly. The essence of competition training is facing extremely difficult problems for a long time—“failing every day”—so one does not keep suffering after failing to solve a single problem.
He liked the reminder on the first page of the British Mathematical Olympiad paper: the problems are very difficult, and “if you can completely solve even one problem, that is already very good.”
After speaking with Wang Hao, the men’s team’s head coach, he compared athletes with competition students: repeated wins and losses every day prevent a single game from breaking one’s psychology.
The training shaped his “super-optimistic” personality. Bad news may upset him for 1 or 2 seconds, but he quickly moves on. He also acknowledges that this optimism can become a bug if he bets on the wrong deep pit.

14. His first money came from teaching olympiad math; software entrepreneurship began with a bug fix

From middle school onward, Wu Minghui would explain complex problems to classmates when competition teachers could not teach them clearly. At Peking University, an olympiad-training institution recruited him as a coach.
A parent rented a classroom for him at Tsinghua and brought in a full class of students. Around 2001, he taught twice a week and earned RMB20,000-30,000 a month, while housing near the Fourth Ring Road cost about RMB5,000 per square meter.
He entered software by accident. An informatics-competition teammate took an outsourcing job, encountered a large number of bugs, and asked Wu Minghui for help. Wu read up on the problem and fixed everything in a day. The teammate then suggested, “Why don’t we start a company together?”
He had no business-family background and did not plan to become an entrepreneur. His path was always to solve problems nearby first, then turn the capability into a business.

15. 秒针’s business model came from advertisers’ need for real traffic

The initial plan was to sell traffic analytics and recommendation systems to publishers. But media companies commonly exaggerated traffic to sell ads and did not want a third party seeing the real numbers.
Only after a media friend adopted the system did Wu Minghui understand the real source of resistance. He then turned to advertisers who spent money on ads every day, because they were “effectively being cheated all the time.”
Helping advertisers verify real traffic was both a genuine need and a way to obtain the data required to build recommendation systems. That became 秒针’s long-term direction.
Small publishers on the supply side later disappeared in large numbers, while major publishers could build in-house and were more resistant to external monitoring. Wu Minghui believes the original publisher route never had a sustainable market.

16. 秒针’s original moat was stable computation over massive logs

秒针 went live monitoring ads on MSN and received tens of millions of data points on its first day. The data center had not expected a company renting only 2 racks to saturate the bandwidth; the system appeared to crash, but the real issue was an egress limit.
Once the bandwidth was opened, the load arrived instantly, yet 2 servers still held up. Wu Minghui says very few teams in China—whether startups or large internet companies—had that kind of big-data capability at the time.
The server configuration also differed from video platforms: rather than mainly stacking disks, the company fully loaded CPU and memory to perform large amounts of cross-computation in memory. Its later technology stack resembled the Spark and Databricks path.
Advertising anti-fraud is not simple counting. Every IP log must be traced through historical behavior and compared for similar patterns to determine whether it represents bot traffic, sharply increasing computational complexity.

17. The 2008 crisis nearly froze financing; the commercial inflection came only in 2010

Wu Minghui recalls that the first VC wave entered China around 2005. He met 30-40 accessible funds at the time. When he went out to raise money again in 2007, he was still almost a “nobody”; many Chinese teams lacked decision-making authority and had to report to partners at headquarters in Silicon Valley.
He was a newly graduated technical student talking about data processing, analytics and future recommendations. Capital found it hard to believe that someone without a work history could build the company.
The 2008-2009 financial crisis made fundraising even harder. The company survived by occasionally taking outsourcing projects worth several hundred thousand RMB and borrowing about RMB1M. The benefit was that Peking University graduates still earned only several thousand RMB a month, so the burn was low.
After 2010, internet advertising accelerated, while broadcast-policy changes pushed budgets from television to the internet. 秒针 already had the data and technology, and its product began selling at scale in 2010 and 2011.

18. Missing mobile internet was the wrong feedback from his “value network”

Wu Minghui admits that he “regretted it quite a lot” when mobile internet arrived. Advertisers believed phone screens were too small for advertising, and he was led astray by existing customers’ judgments.
He uses The Innovator’s Dilemma to explain it: everyone receives ongoing short-term rewards from their own value network, but that network’s feedback can be wrong as a whole.
As a technology obsessive, he did not buy his first iPhone until the iPhone 4. He never installed Douyin on his phone and has watched less than 1 minute in total. If a message can be answered on a computer, he will never use a phone.
He prefers the largest laptop available and multiple large screens, believing that a phone shows too little information at once. That personal preference, combined with advertising-industry feedback, made him unusually slow to recognize the mobile opportunity.

19. The advertising-recommendation business was split off because the “referee entered the game”

After seeing PC traffic growth bend around 2012, Wu Minghui restarted the second phase of the commercial plan: establishing a new BU to handle ad recommendations and conversion for mobile games and e-commerce customers.
He was both 秒针’s CTO and the BU’s general manager, effectively running a business end to end for the first time. Within about 2 years, the business was indeed working.
The market then questioned whether 秒针 could both monitor neutrally and buy traffic for delivery, calling it “both the referee and the player.” Once it bought traffic for media, it lost its third-party identity.
The company ultimately split in two. The original partner took the new delivery business incubated by Wu Minghui, while Wu stayed with the monitoring business and became 秒针’s CEO, creating the key transition of 2014-2015.

20. One lean-startup course split a company into 3 companies

After studying entrepreneurship at CEIBS in 2014, Wu Minghui brought back concepts including the second curve and lean startup, asking the team to assess whether its existing product and R&D organization was redundant.
The team concluded that at most half the people were needed to maintain the existing business. He optimized part of the workforce and allocated resources to establish Minglue Data and Yunji Robotics.
One company became 3: 秒针 continued advertising, Minglue focused on big data, and Yunji built robots. Wu Minghui was Yunji’s co-founder but later retained only a non-executive director role.
His reflection is that the problem was not a shortage of talent, but too many excellent people piled onto one mature product. Once split, the organization generated 3 companies with independent prospects.

21. Minglue Data was a second startup, not an extension inside 秒针

Minglue Data initially had no direct equity relationship with 秒针. Wu Minghui trusted his big-data technology but did not trust his ability to sell to the IT market at the enterprise level.
He therefore recruited a former Oracle executive as the initial CEO while continuing as 秒针’s CTO. The executive was soon poached by Alibaba Cloud to become a vice president.
秒针 was already stable, but its ceiling had become clear after the advertising-delivery business was separated. Minglue represented Wu Minghui’s return to mathematics, data analytics and AI.
This was also the state in which Li Guangmi encountered him around 2015: not a first-time founder, but a technical CEO reopening a big-data track from within a mature business.

22. Copying Palantir brought a market, but also exposed the slow feedback of To G

Minglue initially studied the U.S. company landscape, with Palantir as the clearest reference. It therefore entered To G first before expanding into manufacturing, finance and other industries.
The technology evolved from big data into knowledge graphs and ontology. Early ontology work relied on humans labeling symbols, rules and relationships; today AI can generate some of those rules.
Wu Minghui acknowledges that the To G market was large and well-funded, but its feedback was “extremely, extremely slow.” Without timely feedback, it is difficult to keep iterating and strengthening AI.
During the pandemic, Minglue shut down or spun out all To G businesses, with the original teams operating independently. The decision did not reject the market’s size; it judged that To G was incompatible with the feedback mechanism AI required.

23. The 2019 merger was meant to create a unified “brain” for enterprises

After moving from To G into To B, Minglue began handling complex data from finance, manufacturing and offline retail. 秒针 had enterprise customers including Procter & Gamble, Coca-Cola, BMW and Mercedes-Benz, as well as advertising data.
Wu Minghui was CEO of both sides. Whenever a new business emerged, he kept asking, “Which side should this go into?” Placing the same opportunity in either company would have been enough to raise its standalone valuation.
He eventually chose the simplest solution: in 2019, 秒针, Minglue Data and AdMaster merged, with the goal of consolidating all enterprise data and training AI on top of it.
The merger also solved a management-bandwidth problem. Previously, about 20 people on each side reported directly to him. Unification allowed the organization to be redesigned instead of having one person maintain 2 separate systems.

24. Returning to Peking University for a PhD was about catching up with the AI paradigm shift

Chief scientist Wu Xindong believed Wu Minghui, with only a master’s degree, needed to return to school to access the latest technology if he wanted to build an AI company. Wu Minghui himself had also seen deep learning solve the computer-vision problems of his early career.
In his master’s years, the mainstream approach was feature engineering, wavelets and Fourier transforms, followed by support vector machines. After AlexNet topped ImageNet, neural networks changed the field completely.
Wu Minghui bought textbooks on deep learning and habitually “looked at the formulas first,” deriving the differences between each new method and university mathematics. He returned to Peking University in 2019 for an engineering PhD.
His judgment was that deep learning had broken through in computer vision, so the next “crown jewel of artificial intelligence” was naturally NLP. He chose natural-language processing for his doctoral research.

25. GPT overturned BERT, forcing him to rewrite his thesis proposal

When he began his doctoral proposal, NLP’s SOTA still centered on BERT. GPT appeared in the literature review but was not yet the main research line.
Once GPT suddenly became mainstream, much of the existing task-specific work was no longer worth pursuing. Wu Minghui says NLP PhD students at the time were “crying themselves unconscious in the bathroom.”
He responded by rewriting the proposal instead of defending sunk costs: “Then I’ll stay at Peking University for a few more years.” The new direction shifted to advertising-video understanding and multimodality.
His second proposal coincided with Meta releasing multimodal work. Reviewers warned him not to let a major tech company “come chasing with a tank” again, and told him to find a topic on which general-purpose models would not compete.

26. “Subjective video understanding” replaced a wrong answer key with a population distribution

Wu Minghui chose to study the subjective emotions different groups experience when watching the same video. In related work at ACM Multimedia, he challenged the foundations of existing subjective benchmarks.
His philosophical division is simple: if everyone has one unique answer, the task is objective; emotions and feelings have no single ground truth. He disliked the Chinese-language question “What emotion does this essay express?” from childhood, because personal feelings should not be graded by a teacher.
In a video of a war ending, the winning side may feel happy while the losing side feels sad; the video cannot be assigned only a positive or negative label. The correct method is to build audience-profile groups and model gender, age and region separately.
The model could answer, “How would Asian men aged 30 to 40 respond to this ad?” That is particularly useful for Chinese brands going overseas, because a Chinese CMO’s taste cannot automatically represent the Middle East or other markets.

27. Different model rewards ultimately create different product personalities

Wu Minghui believes GPT models are rewarded for pleasing users, while Anthropic’s Claude, constrained by a constitution, prioritizes following instructions. The difference is not capability, but different reinforcement targets.
He compares it with Google and Baidu: one reward may seek a single click and departure, while another wants users to remain on the webpage longer. Both teams may be excellent, but different values produce completely different outcomes.
Models therefore will not follow one universal evolutionary path. Enterprise agents, coding agents and consumer chat products operate in different environments and will make different trade-offs between behavior and reliability.
That leaves room for vertical companies. They do not need to copy every general-model capability; they can continuously reinforce a particular value function through business environments and proprietary feedback.

28. EIP tried to build the model, data platform and entire office suite at once

After online work exploded in 2020, Wu Minghui named the product line EIP and sought to build 3 layers simultaneously: applications, enterprise data integration and AI models.
He initially asked HR to organize 100 people. One month later, he raised the target to 1,000. The company shut down 2 potentially problematic product lines and quickly assembled a team of about 1,000.
The application layer included CRM and other office products. The middle layer was responsible for cleaning and governing enterprise-environment data into a unified platform. The bottom layer used BERT to train task-specific models for customer service and other tasks.
The concept was called “Little Ming Assistant.” Looking back at the 2020 promotional video, Wu Minghui believes it looked very much like the ChatGPT and Microsoft Copilot that came later.

29. A huge organization and the wrong process pulled Xiao Hong away from product detail

Xiao Hong initially managed only part of the applications. Because Wu Minghui admired her product instincts, he gradually put her in charge of all applications and eventually product across the entire organization.
She had never managed so many people and was a product manager who cared deeply about detail. Once the organization became layers of reporting, she could no longer refine products personally.
Wu Minghui also introduced Huawei’s IPD process, which he later judged to be the wrong management logic—perhaps suitable for hardware, but not for an innovative business at that stage.
He does not blame the outcome on Xiao Hong being a poor fit. Instead, he acknowledges: “The core issue was that what I was doing was wrong.” The direction was too early, the scope too broad and the management method incorrect.

30. The biggest technical misjudgment was attempting L3 before L1 was mature

Wu Minghui borrows Sam Altman’s tiering framework: EIP required an L3-level Agentic Model, but even L1 chat capability was not ready at the time.
He had expected these problems to be solved within 2 or 3 years, which was not an unreasonable estimate. The regret was that the breakthrough ultimately came from U.S. model companies, not inside Minglue.
Scholars including Tang Jie taught at the company or served as advisers, showing that the team had access to the frontier. The problem was that the general-purpose technology base was not yet strong enough to support the product vision.
He summarizes the period as: “When L1 was at 30 points, I was doing L3 work, so it died.” Today L3 may still be at 30, but optimizing it for a scenario is what it means to be “half a step early.”

31. Capital and operating shocks arrived together in 2022

In 2020 and 2021, Wu Minghui still believed the company could build a radically different product. In retrospect, he thinks that even another $500M might not have saved it, given the simultaneous battle across models, applications and data infrastructure.
Capital conditions cooled at the beginning of 2022. Investors increasingly stopped investing in projects or merely continued talking. Shanghai’s lockdown also prevented major customer locations from issuing invoices and making payments.
Most revenue came from Shanghai, while the company also had many international customers. The lockdown affected about half of revenue, while the 1,000-person product team and other company groups continued burning cash.
From May or June that year, mass layoffs became unavoidable. Wu Minghui’s brutal conclusion: “Even if we had raised more money, we might still have died. Dying earlier would have been fine.”

32. At the most dangerous point, only 1 month of runway remained

The company’s cash balance had fluctuated around RMB100M for several months, while monthly spending was also about RMB100M—“which means 1 month of runway.”
Investors repeatedly accused him of wasting the huge sums previously raised. He was still internally “unconvinced,” but first had to keep cash flow alive: “Even if it’s a down round, I have to raise this money on my knees.”
Now that the company can generate its own cash and is profitable, he believes it does not need a large cash balance to be safe. Conversely, even RMB2B on the books remains dangerous if monthly burn is RMB100M.
His most important personal change over the past 3 years was learning to apply his mathematical ability to operating and financial calculations rather than assuming capital would be supplied indefinitely.

33. The other side of a high valuation was underestimating the founder’s responsibility to capital

Looking back at the previous cycle, Wu Minghui says many technology companies treated raised capital as inexhaustible and spent it wastefully. Before experiencing a crisis, “none of us had enough reverence for capital.”
The higher the valuation, the less dilution the same funding amount creates. But if revenue is only RMB200M-300M and a software company cannot list, high valuation and large fundraising can eventually become a dispute over buybacks between investors and founders, with the founder potentially becoming a deadbeat.
Fundraising also performs a pricing function beyond replenishing cash. Once there is a market price, secondary transfers, employee equity incentives and subsequent transactions have a basis.
His conclusion is not to reject capital, but to recognize that when China’s M&A market is still incomplete, the financing route ultimately needs to lead to an IPO or another exit arrangement.

34. AI could push enterprise software into an M&A wave

Wu Minghui believes enterprise services could become the base for future acquisitions, allowing companies to continue buying more businesses like Xiao Hong’s. He acknowledges that China’s enterprise-software M&A market has historically been inactive.
In the past, integrating 2 products required rebuilding underlying interfaces, accounts and data. Today, if both products are owned, they can be turned into sub-agents and coordinated by a multi-agent system.
Because the cost of creating synergies between products is falling, he expects more consolidation in enterprise services in both China and overseas.
This also echoes Minglue’s own 19-year history: past mergers depended on hard organizational and system integration; the next round may primarily use agents to reorganize the delivery chain.

35. AI will make every employee feel like a boss with an assistant

Traditional enterprise software is friendly to bosses because bosses only view dashboards. It is painful for salespeople, finance staff and other operators; a salesperson may spend “40% of their time” filling CRM and other forms.
A boss does not feel that expense claims are difficult because a secretary connects invoices, forms and multiple systems. Wu Minghui believes AI will distribute the same assistant capability to every employee.
Salespeople could give compliant, anonymized meeting recordings to AI to fill in the CRM automatically, while expense claims would require only a photo of the invoice. The output of one system could automatically become the input of the next through an assistant.
AI therefore changes not only software usability, but “the entire software industry, enterprise services and even the industrial internet”—the way systems connect and the production relationships built around them.

36. Before Xiao Hong left, she had already learned the operator’s hardest lesson

After the layoffs, Weiban, which Xiao Hong managed, shrank from a peak of about 300-400 people to around 100. She had to personally lay off colleagues she had recruited and considered highly capable.
Wu Minghui believes the experience taught Xiao Hong cash flow, P&L and organizational contraction early. She brought Nightingale’s P&L to break-even, and by the time she left she was a “qualified CEO.”
She was no longer merely the strategy and product owner, but also the operating owner. Wu Minghui says he had not been fortunate enough to experience such a “large-amplitude” cycle at her age.
The independent project she initially wanted to try involved gene-editing tools. Its name, “Butterfly Effect,” also came from her interest in complex systems and life sciences.

37. He missed the investment in Manus because he lacked cash, not because he failed to understand it

Xiao Hong once offered Wu Minghui an opportunity to invest at a low price, and he was preparing to sign a term sheet. But Minglue’s cash flow was tight, and a separate stake he planned to sell failed to close. In the end, he could only say, “Let’s talk again when I have money.”
The day after Manus launched, before its $500M financing, Wu Minghui told the IR team that if he had the money, he would personally invest at a $1B valuation.
The judgment came from using it himself. Xiao Hong gave him an account with near-founder privileges, allowing unlimited tasks and high concurrency. He called the product “the best of the best.”
Minglue and Manus had not communicated in the months before launch, yet each independently arrived at several similar directions. Wu Minghui admits Manus was better at To C and says he learned many new ideas from it.

38. Tencent, personal loans and a $100M financing bought only limited breathing room

During the cash crisis, Tencent “gave us a hand.” Wu Minghui also invested more than $10M of his own money, but most of his holdings were unlisted stakes in Yunji and other companies, which he did not want to sell at what he considered the bottom.
He preferred to borrow from several prominent figures on the strength of his historical reputation and invest the money back into Minglue. In his view, the robotics era had not yet arrived, and selling robot-company shares too early could mean missing 10x or 100x potential.
Even after investors, Wu Minghui and Tencent put in a combined roughly $100M, severance payments alone exceeded RMB200M, with additional funds used to repay bank loans.
Some banks had said they would renew loans after repayment, but once they collected the money, they stopped lending. After these expenses, little remained in the company, and its cash balance never became comfortable.

39. Resource constraints compressed a thousand-person project into a 40-person product

EIP used about 1,000 people to build one product. Deep Miner, with comparable strategic importance today, started with about 20 people and has only 40 after expansion.
The model team has more than 20 people and the application team has more than 10. Shareholders saw that the company could still deliver multi-agent systems and high-scoring models with limited resources, and concluded that Wu Minghui had become a mature CEO.
The company’s historical compute resources were also extremely limited. Before the pandemic, it purchased about 1,000 GPUs, including consumer-grade cards. Wu Minghui says it may have been the Chinese company capable of building models with the fewest resources.
The organizational principle this time is a small team fighting end to end, without treating headcount and code volume themselves as moats.

40. Code volume was once a moat; in the AI era, it no longer works

Wu Minghui once believed Chinese software did not make money because investment was insufficient. A product built by 5 people over several months looked to customers as if they could build it themselves.
Hardware requires prototypes, tooling and supply chains. By the time a finished product appears, customers are unwilling to repeat the investment. Software more easily makes customers feel they can build it in-house.
The 1,000-person EIP plan was based on the assumption that “1,000 people working continuously for 3 years” would create sufficiently complex code, which could then be sold for RMB1M per customer while creating RMB10M of value for each.
Today AI can generate code quickly and an engineer can write tens of thousands of lines in a day. Code volume has lost its meaning; the new moat must be vertical data, real environments and continuous feedback.

41. What cannot truly be copied is the time axis of data

In advertising, Minglue continuously monitors ad opens and clicks across platforms, while also using data partners and its own crawlers to save short videos, posts, shares, comments and likes from social media.
General-model companies often obtain a snapshot from a single point in time. Minglue focuses on continuous sequences such as daily changes in a KOL’s followers. The difference is not capturing something once, but saving it continuously for years.
Wu Minghui compares it with Tianyancha: a company’s current ownership structure is only a snapshot. Saving it daily allows the historical evolution, relationship changes and disputes to be reconstructed.
A customer once asked in 2012 for an analysis of Olympic advertising from 2008. By then it was impossible to scrape the material again. Once historical data is not saved, it can never be rebuilt.

42. Customer first-party data has value, but also clear ownership boundaries

Minglue also hosts first-party data including brand-member records, customer-service interactions, store POS data, inventory, pricing, SKUs and conversations between salespeople and customers.
Wu Minghui repeatedly emphasizes that this data “does not belong to us.” Minglue only helps customers govern and use it. Yet the data is the most important context for judging real business outcomes.
A store’s “people, goods and place” is not a single table. It is an environment formed jointly by salespeople, products, space, inventory and IoT, making it well suited to training vertical agents.
This data advantage also means products need private-deployment versions. In a SaaS environment, many customers will not connect highly confidential internal databases.

43. As models converge, enterprise competition moves to the (x) in (y=f(x))

Wu Minghui writes decision-making as (y=f(x)): (f) is the model and capability, while (x) is the evidence, data and context needed to make a judgment.
After DeepSeek O1 and similar models appeared, he heard people in Silicon Valley and Peking University’s mathematics alumni network discussing whether quantitative work would no longer need humans. If (f) is sufficiently intelligent, the difference shifts to (x).
Ten companies using the same model and the same long-context window will not therefore have the same competitive strength. What separates them is enterprise context that models cannot contain and the public internet does not have.
Minglue’s strategy is therefore not to compete with foundation-model companies on general knowledge, but to accumulate data and build environments in existing and new industries before training its own specialized capabilities.

44. Selling only general models and tokens may ultimately “compete itself down to electricity bills”

Public internet data cannot belong exclusively to one company. Once 2 models are close in capability, price competition will push general-purpose capability toward infrastructure cost.
Even if a new model leads temporarily, competitors can design prompts to obtain its answers, turn the Q&A into training data and catch up quickly through distillation.
Wu Minghui therefore believes companies that “train foundation models on public data and rely mainly on selling models and tokens” will face intense commoditization.
Minglue trains only specialized small models, mainly for its own agents. Pricing should reflect scenario value rather than charge more simply because more tokens are consumed.

45. Small models provided the first validation on the BUA and CUA leaderboards

After Minglue entered the Browser Use-related Mind2Web leaderboard, Wu Minghui said its small model ranked first and that larger models might not beat it. This validated the specialized training required to operate legacy enterprise systems.
On Computer Use’s OSWorld, its 7B Specialized model scored about 44, while the Kimi models in second and third scored about 34. Overall, it ranked fourth.
The teams ahead were mainly large groups from OpenAI, Anthropic, ByteDance and Qwen. The appearance of a small company for the first time prompted evaluators to ask, “Where did you come from?”
What made him prouder was pass@1: a single submission produced a high score in the organizer’s online test, avoiding the controversy of submitting repeatedly and using test trajectories for reverse training.

46. Deep Miner was born from an SLG game experiment

One source of the project was Minglue’s recurring problem of losing access to data. Advertising results sat in the backends of ByteDance, Alibaba, Amazon and Google, sometimes with APIs and sometimes without, forcing employees to inspect pages manually.
Another line came from an SLG game company Wu Minghui invested in. About 40% of users were especially wealthy “big brothers” who recharged daily to build islands, while about 60% were companion players who spent almost nothing.
He initially wanted AI NPCs to replace the companion-player traffic, but quickly realized the AI would need to know external news such as Trump’s latest activity to avoid exposing itself in conversation, while also genuinely operating the game and coordinating battles.
A comment from the game company—“Our games are all numerical games”—changed his thinking. Companies also analyze numbers in software every day and make decisions from them. “Why start with games? I should first do my own business.”

47. Business itself is “a numerical game in the real world”

During the Lunar New Year, Wu Minghui played the game while calculating the cost of building houses, mining, raising troops and defending territory. Becoming a regional lord might cost RMB1M-2M, followed by geopolitical competition and persistent inflation.
He consequently reinterpreted business as “a numerical game in the real world”: companies read states across multiple software systems, allocate resources and compete with rivals, with no fundamental difference from an SLG.
He figured out the full loop on a ski lift. He immediately assigned the industry team to research CUA, BUA and AI planning, and launched the project on the first day after the holiday.
The project was initially called Deep Mining. His son later suggested Deep Miner, and the available domain made the name official. “Mining” came both from the company’s English name, MiningLamp, and from the tradition of data mining.

48. Agents are hard because of the action space—and that is the opening for vertical companies

Go has a transparent board with 361 possible points and a clear reward. A computer has multiple software products, each with numerous buttons, states and dependencies, making the action space far larger.
Enterprise tasks also require subgoal decomposition. If the goal is to build a trillion-dollar company, planning, tool calls and feedback cannot be defined as completely as in a board game.
Wu Minghui therefore believes today’s Agentic Models are “far from meeting customer needs.” This does not reject agents; it leaves room for companies with industry data, environments and experts to optimize them.
Minglue does not pursue a general-purpose agent. It narrows the task distribution in controllable domains, fills in context, establishes benchmarks and turns unstable capability into deliverable results.

49. OCR and elevator cables provide 2 arguments against infinite scaling

OCR was once a classic task that a small model could handle. But general multimodal models are mainly trained to “describe what they see,” favoring global understanding and potentially making character-level recognition errors.
Wu Minghui gave a Tencent Cloud conference agenda to a general-purpose product and asked it to study the speakers’ views over nearly 6 months. The first OCR step was wrong, so all subsequent research failed.
If OCR must become an independent expert inside a MOE, the world may contain tens of thousands of expert tasks. The more experts there are, the more complex routing, training and basal metabolism become.
His analogy from Finland’s KONE elevators is that cables in a skyscraper eventually weigh more than the car and mainly pull themselves, so tall buildings must be divided into transfer sections. “Scaling laws will hit a bottleneck at some point”; general capability is only relative.

50. Enterprise AI must retain text, GUI and human intervention simultaneously

Text interaction resembles assigning a task to an assistant: flexible and suitable for mobile settings. But when the task involves a local image detail, a file or precise manipulation, language alone cannot fully express intent.
Wu Minghui expects GUI and CUI to coexist for a long time. Enterprise work should also expand I/O bandwidth; financial traders use multiple screens precisely to absorb more information simultaneously before making decisions.
Even when the data source is correct, Minglue says it still has about 1% hallucination. Other To C agents may hallucinate in the low teens to 20%-30% range when data is available, and approach “pure fabrication” when it is not.
The task process must therefore be transparent, inspectable and interruptible. Deep Miner expands data into CSV and other formats so people can modify, check and trace it.

51. Cross-border e-commerce is the clearest initial commercialization scenario for enterprise multi-agent systems

Current target customers are mainly advertising, retail and cross-border consumer companies. They need to operate local ad platforms, retail channels, logistics, supply chains and customer-service systems.
Companies with RMB1B to several billion in revenue may employ 100-200 operators, yet still cover only major platforms such as TikTok and Google.
Wu Minghui envisions agents with multilingual ability, local cultural understanding and numerical computation reaching long-tail markets in Denmark, Germany and even Hokkaido.
In the near term, products will still begin with data analytics. CMK, product, supply-chain, marketing and social teams at large companies can connect internal and external databases to identify consumers, suppliers, keywords and KOLs.

52. Tool Use will shift software entry points from platforms to user agents

Legacy IT systems are old and lack APIs. Their original developers have long since left, and nobody dares wrap the database with MCP. External systems such as tax authorities are even less likely to open the interfaces enterprises need.
BUA and CUA allow agents to imitate people operating browsers and computers, connecting production resources and tools that previously could not be connected.
Recruiting illustrates the change in entry points. A traditional agent can search positions or résumés only inside Zhaopin’s own database, while a powerful Tool Use agent can continue operating BOSS Zhipin, 51job and other websites.
Users will ultimately trust an agent that completes tasks across platforms, not a single website. Zhang Xiaojun said, “The entry point has changed,” and Wu Minghui agreed that this would create a new production relationship.

53. Platforms that only provide connections may be bypassed by higher-level agents

Platforms such as Didi historically provided links between supply and demand without necessarily producing supply themselves. If device entry points from Apple or Xiaomi can directly dispatch services, the original platforms risk being bypassed.
Wu Minghui’s defense is for platforms to own irreplaceable supply-side capabilities. If Didi is strong enough in autonomous driving, for example, it is no longer merely a replaceable hub.
Enterprise software faces the same logic. It cannot simply defend historical entry points; it must possess the best delivery capability, data or model in a specific scenario.
He emphasizes that the revolutionary variable is not the abstract word “AI,” but the agent as an interaction technology: planning determines how to interact, while Tool Use determines what it can connect to.

54. The boundary between workflow and a true agent is who does the planning

Wu Minghui cites Anthropic’s framework from “Building Effective Agents”: many products marketed as agents are actually workflows, with steps predefined by humans and AI executing only at fixed points.
A true agent system does the planning itself, breaking a task into subgoals and choosing tools independently. Manus is closer to an agent than an ordinary workflow in this sense.
Workflows are deterministic and suitable for routine tasks; agents are open-ended and suitable for exploration. Enterprises need to determine which work can be fixed and which must retain trial and error.
Writing a program is itself exploratory. Once debugging is complete, the code becomes a workflow and tool. Employees should keep turning boring work into fixed processes, then use existing tools to explore higher-level questions.

55. The enterprise AI roadmap is not layoffs, but scaling the capabilities of the strongest people

The first step is still to add AI to existing work. Minglue has about 300 data analysts; if they analyze faster, company efficiency and gross margin will improve.
But true transformation is not merely “faster.” It means letting AI achieve higher performance in every industry and multiplying the output of the best people many times over. Go has already shown that AI can surpass humans; enterprise scenarios lack only complete context and inexpensive rewards.
Advertising strategy must be tested with real money. The real environment is far more expensive than a board game, and a standard GRPO reward in a simulator may diverge sharply from business feedback.
Wu Minghui believes the key question is how to amplify the capabilities of top employees by multiples and make them scalable. Only then will a company be “completely transformed.”

56. MOA turns the company itself into a multi-agent architecture

Manus is closer to a single agent, while Deep Miner began as a multiple-agent system. Minglue calls the architecture MOA, or Mixture of Agents, corresponding to MOE, or Mixture of Experts.
Wu Minghui acknowledges that as CEO of a company with about 1,800 people, he cannot hold every piece of context and cannot write a good single global planner. A real organization must let each team build its own benchmark and optimize end to end.
Optimized agents can then join shared teamwork. The organization may use supervisors, self-organization, democratic negotiation or centralized command, resembling human organizations.
The company has about 1,800 people and roughly 800 engineers; historically, perhaps only 100 truly worked on AI. His requirement is for ordinary engineers and data analysts to switch roles and become AI engineers or data product managers.

57. The future enterprise may consist only of an owner and a group of agent partners

Wu Minghui cites a “provocative but not necessarily wrong” scenario in which an enterprise consists only of the owner and partners, with those partners not full-time employees but providers of a function to multiple organizations.
The owner controls the company’s memory, benchmarks and ultimate responsibility. Experts in advertising, pricing, psychology and other fields connect their own agents to the organization.
The difference between a “partner” and an ordinary SaaS subscription is that the partner shares in final business performance: strong sales mean more income, while weak sales still earn a base fee rather than a fixed salary.
This structure requires agents to possess IP. If the private context inside a top pricing expert’s brain is packaged into a personal agent, it is worth far more than an anonymous general-purpose model.

58. Optimism, trustworthy AI and a data bet form Wu Minghui’s new starting point

Wu Minghui wants future agents to have identity, historical results, credit and reputation. When companies buy services from McKinsey, BCG or 秒针, they are often buying the IP that “gives decision-makers confidence after the stamp of approval.”
He has set Minglue’s new slogan as “data-driven trusted productivity.” The goal is no longer to train AI “smarter than me,” but the most honest, safest and most trustworthy AI—one that can detect contradictory citations, incorrect sources and data problems.
His R&D pipeline includes Tool Use, data-problem identification, forecasting and human-machine collaboration. He sees forecasting as the endpoint of Data Mining, because product selection, operations and investment ultimately all lead to prediction.
His personal belief did not change after 2022: “I’ll definitely figure it out. It may just take a little longer.” He recommends that AI practitioners read On the Origin of Species and Elements, sees Chinese chips as a good opportunity, is currently betting on data-related AI, and is bullish on making Chinese intelligent hardware AI-native.

20VC: OpenAI's Multi-$BN AMD Deal & Does Kingmaking Work in VC Education in the AI Age: a Teacher Rethinks Learning & Purpose, w/ Johan Falk of Graspable AI