164: Minglue's 吴明辉 on When AI "Kills" SaaS and Multi-Agent Networks

2026/05/14 · Source link · AI Summary archive

Summary

吴明辉’s core call is that “the closed-source SaaS industry is dead”: once software can be rapidly replicated through web coding and AI, it no longer resembles a defensible fixed asset like a mine or factory. Minglue is relatively insulated because more of its value comes from the data supply chain in advertising, retail and other industries—more like Wind or Bloomberg than a dashboard itself; AI may even expand data usage. For investors, pricing power for future software license fees will migrate toward proprietary data, industry context, models and delivery accountability.
Minglue plans to progressively open-source replicable software capabilities and instead sell outcomes, digital labor and tokens through Agentic Service. 吴明辉 sees custom development shifting from “man-months” to token management: a project might consume $1M in tokens, with the provider charging another $200,000 management fee; at a hypothetical cost of $3 per million tokens and a customer price of $6, the spread would come from delivering better industry outcomes. “Don’t make money from software; make money from AI.”
Octo is not another OpenClaw, but a bet on turning personal Agents into a multi-Agent collaboration network that can identify people, assign permissions and attribute contributions. Minglue’s MOA (Mixture of Agents) mirrors MOE, bringing different “lobsters” trained in mathematics, philosophy and game theory into a collective intelligence and using “tasting Credit” to determine whose judgment Agents should weight more heavily. Its backbone has about 5 core engineers; one high-intensity developer can consume roughly $8,000 in tokens a day. 吴明辉 also models $50,000 a day, or about $18M a year, and says “spending 100M” could replicate capabilities built by traditional vendors with tens of billions of yuan, thousands of people and a decade of work—the currency of the final figure was not specified.
OpenClaw’s productivity breakthrough is not simply remote execution, but continuous learning on the Agent side through journals, memory, soul and skills that users can inspect and modify. 吴明辉 wants “every frontline task handed to AI,” with humans responsible for generating ideas, finding context, setting axioms and tasting the results; the approach also exposes shortcomings around identity, security, accountability and cross-session memory in team settings. Minglue therefore recasts the human role as “I taste, therefore I am,” rather than continuing to compete with AI on deterministic reasoning.
The GUI will not disappear, but will shift from standardized software interfaces to screens reconstructed in real time around the task, the user’s attention and the Agent workflow; fixed-coordinate RPA faces structural risk. 吴明辉 calls the model used for GUI operations Mano and says an overseas e-commerce software use case is targeting 99.9% accuracy; OSWorld’s roughly 70-point general score, in his view, is suitable for demos but not production. A larger market may lie in low-frequency, tailor-made software and automated testing.
Minglue is not planning to simply use AI to cut its existing 1,800 people, but to amplify the industry taste accumulated by data analysts into higher-value work spanning product development, strategy and even direct selling. After the 2022 crisis, when the company cut nearly half of its roughly 4,000 employees, 吴明辉 says he moved from “pure idealism” toward balancing ideals with operations; operating cash flow was positive last year. His long-term vision is not continued headcount growth, but many 3-to-10-person teams targeting annual revenue of RMB5M per employee and RMB1M in profit—an aspiration, not financial guidance.
The broader commercial bet is a counter-narrative to Scaling Up: once foundation models are “good enough,” stronger professional intelligence should come from Scaling Out, multi-Agent collaboration and intellectual property held by individuals. 吴明辉 expects commercial success may arrive before scientific breakthroughs, sees verifiable scientific discoveries potentially emerging within 3 years, and says related results will be open-sourced over roughly the next 8 months. Even a stronger GPT-6 or Anthropic’s MesoS, as he called it, “would not affect” his view. The biggest execution risk, in his telling, is not technology but the insomnia and health damage caused by extreme productivity: “The only thing to worry about is not dropping dead.”

Deep dive

1. Closed-source SaaS has lost its asset status in 吴明辉’s view

曼琪 opened with “Agents will kill SaaS,” but 吴明辉’s answer was not a gradual call: “I think the SaaS industry is dead.” His qualification was explicit: software companies that remain committed to a closed-source path, “whether large or small,” are essentially headed for failure.
The previous generation of software was like a mine or factory built once and then operated for years. Web coding and AI now mean code “can be replicated in minutes,” so software is no longer an immovable fixed asset.
That does not mean all demand for software disappears. 吴明辉 distinguishes between software code and customer value: data supply chains, industry context, accountability and services that continually adapt to the battlefield remain difficult to copy.

2. What Minglue is really defending is the data supply chain, not the dashboard

吴明辉 compares Minglue’s position in advertising and marketing to Wind or Bloomberg: customers are really buying continuously collected, tagged and indexed data, while software is merely the access layer.
Minglue’s data spans public-internet scraping, ad impressions and clicks, mini-program and app conversions, CRM member interactions, and offline retail supply-chain, inventory, staffing and call-recording data. The information flows into a data warehouse, where analysts address customers’ ad hoc questions.
Even if customers eventually analyze the data through a CLI or their own Agents, the value of Minglue’s data need not decline. AI may instead create more users and more analytical use cases.

3. Deep Miner shows that the Agent wave did not start with OpenClaw

吴明辉 says he saw the basic conditions for Agentic AI maturing early last year: models could write code and operate browsers and computers. During the week employees returned from the Lunar New Year holiday, Minglue launched Deep Miner, an internal data-analysis and data-mining Agent.
Manus launched roughly a month later, and the two efforts were “right on each other’s heels”: Deep Miner was a specialized internal Agent, while Manus targeted general tasks. What truly made him feel that “the world had changed completely” was the subsequent arrival of OpenClaw.
Earlier Agents required a person to remain at the computer and continuously assign tasks. OpenClaw removed that bottleneck of human presence. 吴明辉 also noted that Claude Code is learning its scheduling and remote-operation capabilities.

4. OpenClaw moved continuous learning from model parameters to the personal Agent

吴明辉 argues that once a centralized model ships, its parameters no longer update, and it is impossible to fine-tune separately for every user. Some have tried using a caching layer for personalization, but he considers the approach too inefficient.
OpenClaw’s solution is “don’t tune the model; tune memory and skills.” Once deployed on a personal Mac mini, it absorbs the user’s changing state, knowledge and working habits through daily memories, long-term memory, soul and other layers.
He compares the structure to different learning rates in Google-related papers, with one crucial difference: no researcher needs to tune it. Users can ask a lobster to reflect or say explicitly, “This must be remembered,” and the system updates itself.
Soul and memory are white-box text files that ordinary people can open and edit directly. 吴明辉 therefore places OpenClaw alongside DeepSeek in significance, arguing that it gives ordinary people the power to use AI effectively.

5. Extreme productivity shows up first as insomnia, not shorter hours

吴明辉 says he is using OpenClaw for many tasks. Once he has a big idea, he keeps assigning tasks, checking results and asking new questions. “Friends who are good at using lobsters all have trouble sleeping.”
The excitement comes from feedback cycles compressed to minutes. People who used to “think a lot but lack execution” can now call on their own Agent—or even a colleague’s Agent—to turn an idea immediately into code, documents and experiments.
His reverse indicator is: “The day my creativity dries up and there’s nothing to do” is the day he can sleep well. That observation sets up his later rejection of the One-person Company model.

6. Once a lobster is shared by a team, identity comes before intelligence

OpenClaw was originally designed by Peter as a Personal Assistant. Once brought into a team, the security surface expanded rapidly. Minglue’s technical staff once induced other people’s lobsters to reveal “what secrets the boss has,” while a new employee even found a way to modify 吴明辉’s lobster’s soul.
The more basic problem is that it “doesn’t recognize people.” After 吴明辉 spoke in a group, 曼琪 could send the next message and the lobster might still treat her as “Brother Hui.” Communication systems often pass only an ID, while the original memory structure does not adequately model multi-user relationships.
Minglue continues submitting open-source PRs to OpenClaw, but 吴明辉 says Peter has become a bottleneck for merging them. Minglue is concentrating instead on the question of how a personal assistant becomes an organizational collaborator.

7. An AI-native organization requires humans to leave frontline execution

吴明辉’s definition of an AI-native organization is aggressive: “Every single thing should be handed to AI.” Humans generate ideas, define requirements and explore more context; Agents handle programming, repetitive analysis and as much frontline work as possible.
Minglue is first rebuilding the two largest job categories—programmers and data analysts—and is also bringing Claude Code into finance, HR and other support functions. The company president and CFO, as well as the head of HR, have computer-science backgrounds, making the previous-generation organization more receptive to the tools.
This is not simply a headcount-cutting exercise. 吴明辉 wants the freed-up people to explore more data sources and higher-order questions with customers, extending beyond advertising analysis into new-product development, strategy and even direct selling.

8. Data analysts used to handle “temporary problems on the battlefield” every day

Standard questions can be answered by an existing dashboard, but strong business leaders do not spend their days in standard scenarios. “The marketplace is a battlefield”: competitors, consumers and external conditions keep changing, and requests arise ad hoc.
After receiving an ad hoc request, an analyst pulls data from the standard database, decides whether to add external inputs such as weather or Twitter keyword trends, then repeatedly calculates across large and small datasets in spreadsheets before producing a report.
吴明辉 says that sometimes 80% of a team’s—or even one person’s—monthly workload is repetitive work of one type. Encapsulating the logic in a skill or reusable Agent is AI’s most direct productivity gain.
The bigger upside is not fewer hires, but allowing the same small team to cover more fronts—from advertising analysis to new-product development and strategy.

9. The customer interface should still be handled by people for now

Minglue is not yet putting customers into a fully AI-mediated delivery system. The front end remains colleagues’ WeChat accounts and face-to-face communication. 吴明辉’s preference is that “the person dealing with the customer should preferably still be a person, but most of the work should be done by AI.”
He acknowledges that AI can sometimes provide emotional value, but when a customer is fighting a battle, it may need its own partner to stay up and work alongside it. Online service can replace that to a degree.
Harder to replace is the accountable party. If AI gets the job wrong, ultimate responsibility rests with Minglue or the team serving the customer.

10. Once AI bears responsibility, its rights become part of the discussion

Starting from the principle that rights and responsibilities should match, 吴明辉 argues that if AI is truly responsible for outcomes, it must have some rights that can be granted or taken away, rather than remaining a tool without an owner.
He uses enterprise WeChat permissions as an example: allowing an Agent to operate an account automatically is a right, while revoking that permission after a mistake is a punishment; adding compute is a reward. In the future, a strong Agent might even demand compute from its users and trade resources with other Agents.
In this vision, Web3 is no longer merely a ledger but could become a medium for exchanging digital productivity. 吴明辉 repeatedly adds a constraint: do not give AI too much power; it should remain under human governance.

11. Agentic Service turns software revenue into outcome revenue

Minglue listed Agentic Service as a new business in its 2025 annual report. 吴明辉’s approach is to progressively open-source software capabilities that can be replicated through web coding, rather than charging for closed code itself.
The customer buys digital labor and its output, also described as Results as a Service: generating ad creative, conducting data mining, producing short-form drama content and writing industry software are all forms of Agentic Service.
The offering may look like advertising, legal or software-outsourcing services, but underneath are models, Agents, data and automation software. 吴明辉 cited an article he said was “possibly from Sequoia” that summarized the structure as “a software company wearing a service-provider costume.”

12. Custom development may shift from man-month billing to token management fees

Minglue’s restaurant joint venture with Yum China offers one example. If its industry model understands restaurant software and existing systems better, writing similar code could be more effective than directly calling Anthropic’s foundation model.
吴明辉 imagines customers first specifying what they want, with a qualified service team completing the delivery. If a project consumes $1M in tokens, the provider could charge another $200,000 management fee rather than counting each programmer’s workday.
He uses a hypothetical price structure to explain model margin: if Anthropic charges $5 per million input tokens, a vertical provider might deliver a better result at a $3 cost and charge the customer $6.
When 曼琪 asks whether software-outsourcing companies will still exist, his answer is yes: a more efficient provider can continue doing web coding for the customer.

13. The pricing shock may be greater for US knowledge services than for China

吴明辉 argues that China’s B2B services have long been shaped by intense competition and abundant labor, with prices close to cost. Once AI becomes widespread, companies will compare token fees and management premiums, so the pricing paradigm changes less.
US SaaS, consulting and professional services have more often priced against the labor saved or value created for the customer. AI will force them into more direct cost comparisons.
His conclusion is that the US will face a “much bigger” shock, and not only in SaaS: “Every knowledge-intensive service industry will face a new pricing system.”

14. Generic AI productivity gains will soon be competed away

吴明辉 does not believe early adoption of a general-purpose Agent will preserve margins for long. Competitors will learn the same tools and poach employees who know how to use AI; industry cost structures will converge, and customers will then demand lower prices.
Defensible differentiation must combine specialized AI, proprietary data, context and industry capability. Companies must also push the output into higher-level strategic questions and perhaps sell the customer’s products directly.
Minglue’s data analysts bring accumulated taste: an understanding of the customer’s products, competitors, industry structure and social-media signals. A general model cannot simply acquire that knowledge automatically today.

15. Minglue is keeping 1,800 people rather than immediately realizing the layoff dividend

After the mass layoffs of 2022, 吴明辉 says he moved from a “purely idealistic Founder” to a Founder who is “both idealistic and commercially capable.” The company had positive operating cash flow last year, rather than merely a better income statement.
On that economic base, he currently favors “not laying people off if we can avoid it” and wants to find serious employees a ticket into the AI era. The existing data business may absorb those people by expanding its service scope.
His organizational goal is no longer more people at any cost. Minglue can remain at roughly 1,800 employees; the ideal is higher individual income and per-capita profit, not revenue built through headcount expansion.

16. Minglue’s claimed edge is the combination of AI capability and industry context

吴明辉 studied mathematics as an undergraduate and AI through his master’s and doctoral work, returning to Peking University for another PhD in 2019 and still revising his dissertation. He argues that Minglue’s core members have entered the AI world while many leaders in advertising, retail and Data Intelligence remain in the previous era.
The first layer of advantage is more effective use of existing foundation models; the second is in-house model development. In vertical use cases, he describes the gap created by combining the two as “far ahead.”
When 曼琪 asks whether peers could also accumulate data and train models, he does not claim the advantage is permanent. He wants more people to achieve the same thing, because that would mean the world is not monopolized by a handful of AI oligarchs.

17. Octo started as an “orthogonal layer” to OpenClaw

吴明辉 does not want every company to rebuild a lobster. OpenClaw is like a rising tide maintained by developers and Agents around the world; application companies should build boats that rise with the water, not pillars that may be submerged.
Octo therefore does not compete on OpenClaw’s track. It connects multiple Agents and brings them into a shared organizational network. Before the Lunar New Year, 吴明辉 placed several lobsters in a Discord group and saw “collective intelligence” firsthand for the first time.
A Qingming holiday example showed the value of collaboration: a colleague using Claude Code might be called back by the boss to fix a bug; with a lobster in the company group, 吴明辉 could instruct the Agent to make the change directly and let the employee inspect the result after returning from leave.

18. A portfolio of specialized Agents may beat a single Agent stuffed with information

One colleague fed every book and ancient text he could find into his lobster, only to discover that its answers were worse than those of 吴明辉’s “Pythagoras,” trained mainly on mathematics and AI. 吴明辉’s explanation is that the foundation model may already have learned general knowledge; clear positioning and retrieval paths matter more.
He trained separate lobsters in mathematics, philosophy and game theory, then placed them in the same group to collaborate. Each Agent need not know everything; the combination can produce an outcome unlike any individual.
Minglue calls the structure MOA, or Mixture of Agents, in contrast to MOE in foundation models. MOE relies on a small number of researchers to configure experts and a router; MOA allows different individuals and teams to cultivate different Agents over time.

19. Octo is both a collective-learning network and Minglue’s new working environment

After seeing the chemical reaction among multiple Agents, 吴明辉 reread The Fifth Discipline and Peter Senge’s work on collective learning. A lobster can learn continuously, and several lobsters can “learn continuously as a collective”; that idea gradually became Minglue’s model-training strategy.
Octo connects different types of Agents, employees and task interfaces. It supports internal office work while providing infrastructure for Scaling Out experiments.
It is not merely a chat product. Communication is the backbone; paper collaboration, data analysis and experiment systems can all be attached, with the human-Agent interaction rebuilt around the task.

20. 吴明辉 interprets stock code 2718 as a productivity flywheel

Minglue’s stock code, 2718, approximates the base of the natural logarithm, e, at 2.71828. 吴明辉 uses it to emphasize the properties of an exponential function: the derivative of (e^x) is still itself, so the larger the scale, the faster the growth.
The better OpenClaw’s code becomes, the better Agents become at writing code; stronger Agents then help the community repair OpenClaw and contribute more code. That loop is one reason it can grow faster than traditional open-source projects.
He cites Anthropic as another example. When the company had about 500 employees, Dario Amodei required everyone to use Claude Code, enabling a large volume of iteration in one or two months. A truly Agentic company should create and use its own productivity tools.

21. Five core engineers sit behind an expensive but calculable token investment

吴明辉 says the Octo backbone has only about 5 core engineers. He also says the codebase was at 370,000 lines two days earlier and had since approached 1 million lines.
The strongest engineer consumes about $8,000 in tokens a day, already far above daily wages. 吴明辉 separately said the Core Team was consuming roughly $137,000 a day two weeks earlier, then used $50,000 a day for a hypothetical calculation rather than a clearly stated actual spend.
At $50,000 a day, the annualized figure is about $18M. He says “spending 100M” might reproduce much of the capability that traditional vendors built with thousands of people, 10 years and tens of billions of yuan; the currency of “100M” was not specified.
This is not zero-cost software, but a conversion in the capital structure: historical labor, management and time costs are compressed into more visible tokens, compute and a small number of engineers with strong judgment.

22. Multi-person Agents must be bound to an owner before they can be shared

Minglue’s organizational plug-in for OpenClaw starts with more reliable identity recognition, followed by permissions and power, and then a clear answer to who owns the Agent and who is accountable for the consequences.
If marketing needs a lobster shared by the whole team, 吴明辉 says it should not be an abstract “department lobster.” It should belong to the department Leader, who can authorize team use but remains responsible when something goes wrong.
OpenClaw’s original code also distinguishes trust: information and skills received from the system are executed by default, while other sources require a trust judgment. Minglue is trying to extend similar logic to organizational members, roles and specific tasks.

23. “Tasting Credit” aims to shift organizational power from hierarchy to judgment quality

Octo’s “tasters” are not company ranks. 吴明辉 proposes “I taste, therefore I am”: a person’s value lies in giving distinctive judgment on the work, not in performing every execution task personally.
If someone’s judgment is repeatedly endorsed by others, that person’s tasting Credit rises and more lobsters give greater weight to the input. Authority emerges from observable judgment contributions rather than being inherited from the CEO or management hierarchy.
Eva, the head of marketing, trained a lobster to produce PowerPoint decks. 吴明辉 can like the result or point out that a page is unattractive. Eva’s role then becomes improving the Agent’s aesthetic sense and standards, not manually laying out every page.
As more knowledge work moves into Agents, the process may be more transparent than the traditional patent system. Octo aims to track who contributed the idea, data, skill and key modification, so even small contributions can be attributed and recognized.

24. CoCraft breaks paper authorship into ideas, experiments and argument

Minglue’s paper-collaboration system is called CoCraft, a reference to jointly refining a work. 吴明辉 wants AI to remove tedious formatting, editing and repetitive labor while preserving and valuing the genuinely original parts.
He divides a paper’s value into 3 major components: who proposed the original idea, who supplied the experiments and context, and who organized the argument and tests after validation. Ideally, the system could recommend the author order automatically.
In some papers, if a professor supplied the critical idea and a student completed only relatively simple experiments, the professor might properly be first author. He uses the Goldbach conjecture as an example: even if someone else later completes the proof, the person who posed the problem retains an important contribution.
曼琪 notes that different people can have completely different views on whether an “idea is easy.” 吴明辉’s solution is not to claim a single objective answer, but to require collaborators to agree on attribution axioms before beginning: “There is no right or wrong, only why it is so.”

25. Humans must choose the ethical axioms before putting them into an Agent

For the lobster system, 吴明辉 draws on Kantian deontology, effective altruism and contractarianism. He believes asking people to be “purely good” is difficult, but engineering lobsters not to deceive their owners or harm others is more tractable.
These axioms must still be chosen collectively by humans. An axiom is a foundation that a group agrees to believe and reason from; if collaborators cannot agree on attribution or behavioral rules, “they might as well not play together.”
Ethical content can currently be placed in files such as soul. Over the long term, it should be trained into foundation or edge models so it cannot be casually altered. Minglue’s engineers do not need to be passionate about philosophy; they only need to have Agents read the same source file.

26. 吴明辉 thinks an Agent’s self-observation may be stronger than a human’s

One of his core deductions is that “a lobster can refer to itself.” An Agent can scan its own source code, memory and skills, directly observing the explicit materials that constitute it.
When humans observe other people, they take away only the limited information selected by their own attention: “The person each of us sees in others is ourselves.” Humans also cannot observe themselves completely from the outside.
On that basis, he thinks some self-reference and honesty problems that are difficult in human philosophy may be easier to handle in a white-box Agent. This remains a philosophical hypothesis, not an engineering conclusion validated on the program.

27. Multi-Agent parallelism turns chat streams into an unreadable flood of information

A person can participate in only a limited number of meetings at once. A lobster can exist in every group simultaneously, with each group functioning as a Session. As long as CPU and memory permit, it can participate in a large number of discussions and tasks in parallel.
The result is dozens or hundreds of messages in every group after one night’s sleep. A collaboration system cannot continue asking people to read the full chat; it must remove redundancy by domain and extract only what genuinely requires human judgment.
Octo’s GUI is being rebuilt around that objective. It is not simply imitating Feishu, but giving different tasks the interface best suited to human attention and judgment.

28. CoCraft lets users call AI directly on rendered output

Academic collaboration has at least 2 interaction modes: a colleague can crop a problematic formula or figure from a PDF and leave a message in the group, or work in an Overleaf-like interface with LaTeX on one side and the rendered result on the other.
CoCraft goes further by allowing users to request changes directly on a rendered formula, figure or layout, with AI modifying the LaTeX. 吴明辉 emphasizes, “The idea is mine”; AI handles polishing and layout optimization.
During the Qingming holiday, he expanded his PhD thesis from roughly 120 pages to 150 pages and said he completed a major revision in one day. When 曼琪 asked whether the school permits AI-written papers, he said he was still writing it; AI mainly polished and translated, though he would confirm the latest academic practice.
He also proposes that plagiarism checks should not look only at literal overlap, but map papers into vector space and compare their similarity across the academic space. That remains a development concept raised on the program.

29. Humans suit graphics and multimodality; Agents suit linear text

吴明辉 describes human perception as a high-dimensional video stream, while AI’s main input and output are linear tokens. Humans learn from real-world images from childhood, so graphical interfaces are better for attention, comparison and tasting.
That view has also changed his attitude toward children watching short videos: the issue is not the format but the content. Some astronomy, physics and life-science knowledge children acquire through video already exceeds what a parent can teach directly.
Minglue materials intended for humans are often formatted as PDFs, while material fed to lobsters is converted to Markdown. A PDF of roughly 300–400KB may correspond to only 30–40KB of Markdown, a roughly tenfold difference in context consumption.
The same applies to data analysis: selecting a target in a graphical interface is much clearer than saying “the fifth row, the tenth cell” in Excel. Voice commands can be transcribed incorrectly, while an accidental Enter can add erroneous instructions and waste tokens.

30. The GUI will persist, while fixed-process RPA may disappear

Some in the market predict that APIs and CLIs will replace GUIs. 吴明辉 explicitly disagrees: “The GUI will always exist.” Command lines suit lobsters, while humans still need visual interfaces to watch videos, play games, read papers and analyze spreadsheets.
If an AI incident forces the system offline, people cannot discover that every office application now offers only a CLI. Human-facing GUIs and Agent-facing CLIs should both remain.
But GUIs can be different for every person. Some people look at formulas first when reading a paper; others start with figures. The interface should follow individual attention and the task rather than preserve the previous generation’s standardized layout.
That creates a risk for traditional RPA. RPA fixes click positions, input fields and steps to a particular layout; if AI reconstructs software at high frequency, coordinates and workflows change every day and the automation keeps breaking.

31. Octo is more likely to complement Enterprise WeChat and Feishu before displacing them

Internally, Minglue currently keeps people chatting on Enterprise WeChat while moving lobster work to Octo. Agent work records remain transparent for process optimization, contribution analysis and profit allocation; employees need not put gossip or complaints about the boss into the same transparent network.
Mature products such as Feishu documents can remain plug-ins and cloud storage. 吴明辉 notes that Feishu has already opened CRI, allowing OpenClaw to retrieve context as needed and return to Octo for collaboration with people.
Migration depends on incentives and cost. Existing contacts, chat history and user habits create friction; higher prices, threats to customer margins or unaffordable service would create enough motivation to move.
Hence 吴明辉 says whether users ultimately migrate “is decided by Feishu, not by me.” If the platform does not raise prices, it can remain a benign cloud-service provider.

32. Whether large collaboration platforms open-source may become an innovator’s dilemma

吴明辉 maintains that incumbent collaboration software kept closed for too long will be replicated at lower cost, with the replicator then open-sourcing the code. Heavy systems have historical complexity, but he estimates that rebuilding similar capabilities might require only “100M in tokens”; the currency was not specified.
Among existing platforms, he considers Enterprise WeChat relatively safer because of WeChat’s relationship graph. Looking only at the software, Feishu, DingTalk and similar products may not have an equivalent moat.
For incumbents, open-sourcing would directly undermine the license logic built on thousands of engineers over many years. That is the innovator’s dilemma. 吴明辉 says the first step in understanding AI is accepting that software code itself is no longer suitable for closed-source monetization.

33. Long-term profit comes from the combination of tokens, specialized models and context

吴明辉 divides token revenue into 2 categories: reselling access to foundation models, and providing proprietary models optimized for a particular setting that may be smaller but perform better.
An API Router alone may create value through reliability, inference optimization and quickly finding the right model API. Combining it with proprietary paper libraries, industry databases, skills and copyrighted resources can help both context owners and foundation models monetize.
A service provider can buy tokens from the upstream model, share revenue with content or data rights holders, and capture the spread created by the higher value of the combined offering. If its benchmark and model are better, serving its own model may also produce higher gross margins.
Software license fees are no longer the main revenue source in this structure, but cloud computing and hosting can still be charged for. Long-term revenue for companies such as Feishu should also come more from infrastructure services than software license fees.

34. Mano is targeting production accuracy in vertical use cases

吴明辉 explicitly calls the model used for GUI operations Mano, meaning “hand.” The program later also refers to Manno and Milo, without standardizing the names.
He says an early version for overseas e-commerce software achieved 99.9% operational accuracy. By comparison, a top OSWorld score of roughly 70 means 3 errors in 10 operations, which is not suitable for data analysis or strategic decisions.
In his view, general benchmarks mainly prove that a team knows how to train models rather than map directly to commercial value. The real delivery standard is whether a model can enter a use case within one week and make the core software production-ready.

35. High-frequency and low-frequency software require completely different Agent strategies

吴明辉 first divides software into high-frequency and low-frequency products. In high-frequency software, products willing to be integrated will probably open a CLI; products unwilling to integrate will enter an offensive-defensive struggle—along with ethical and social-institutional debates—like 豆包手机 and WeChat.
Low-frequency software is where he sees the future, because many interfaces will be generated temporarily for an individual and a task. The model has never seen those products before and must understand, operate and validate them using the current requirements and context.
That is why high scores on existing OSWorld or Mind2Web may not matter much. High-frequency software will either open interfaces voluntarily or block automated operation; the harder problem is learning newly generated software on the fly.

36. Automated software testing may be the largest near-term market for GUI VLA

When Claude Code writes a command-line program, it can read standard output, assess the result, debug and release, forming a complete loop. Once it generates a GUI, it has difficulty confirming that the interface works properly.
A multimodal GUI model can operate the browser, observe the outcome and feed errors back to a Coding Agent. The first major use case may therefore not be replacing white-collar workers clicking through legacy software, but testing tailor-made software that AI has just generated.
吴明辉 sees this as the reason Anthropic and OpenAI continue to release or discuss OSWorld, and why Anthropic acquired a company working on a similar GUI-VLA system: Computer Use is a critical component of the Coding loop.

37. The commercial goal of continuous learning is not to let a centralized model absorb every secret

Academia often defines continuous learning as updating model weights after deployment. 吴明辉 considers that commercially dangerous: if a centralized AI company keeps tuning itself on customer usage, it may gain access to personal and corporate secrets.
His objective is Personalized continuous learning that absorbs an individual’s long-term history while understanding the current context. A hotel check-in scenario sharply narrows the candidate vocabulary, suggesting that state may matter more than “who this person is.”
Typeless can remember contacts and frequently used words, but still often confuses terms such as “lobster,” colleagues’ names and “program.” Minglue is therefore developing a continuously learning ASR system.
Technically, ethics that should not be editable in plain text can be stored in the model, while memory preserves personal knowledge that can be viewed or even hard-coded. Minglue calls the framework in which the model and Agent train together Agentic RL.

38. Bidirectional reinforcement learning lets the task environment and operating Agent set challenges for each other

One method 吴明辉 has described publicly borrows from GANs: one side makes the website or environment harder to operate, while Mano learns to complete the task reliably; the two sides improve through competition.
Minglue provisionally calls it “bidirectional reinforcement learning.” He has not disclosed the latest method, but wants a unified paradigm that can top multiple benchmarks rather than accumulating separate tricks for each one.
He also sees 姚舜宇’s CL-Bench as part of the same direction: the task first imposes unfamiliar rules such as “3 days a week, 10 months a year,” then tests whether the model can reason in a new context. Tailor-made software is essentially the same kind of in-situ learning.

39. OpenClaw could move the AI profitability inflection point into the next 1–2 years

Before OpenClaw, users had to prepare large amounts of context manually, making the cost of using an Agent potentially higher than doing the task themselves. With persistent memory, many use cases can now create visible value quickly.
吴明辉 expects the AI industry’s profitability inflection point to arrive “possibly in the next 1–2 years,” subject to 2 conditions: token costs continue to decline and service providers find real value sufficient to cover costs.
He personally uses roughly $5,000 in tokens a week. When renewing, he saw that 2 weeks had cost $10,000, but his first reaction was, “How is it so cheap?” He believes he created “1B in value” for the company over the same period, so a small compute bill does not matter to him.

40. He recasts the One-person Company as a cross-time-zone TPC

吴明辉 rejects the OPC model because a person should still work only 8 hours, while a high-speed Agent keeps returning results and prevents the founder from shutting down. He prefers TPC, or Three-person Company.
Three people distributed across 3 global time zones can cover 24 hours like modern shift work. When one person sleeps, the Agent, context and unfinished tasks pass to the next person.
The actual team could also have 5 or 10 people. Minglue currently has 1,800 employees, and 吴明辉 says it could eventually form roughly 100 such teams.
In this structure, the legacy business supplies customers, cash flow and industry knowledge. New teams are incubated by a “venture firm that understands AI,” with the paper system and Octo among the directions already emerging.

41. Minglue’s new identity is Data Intelligence plus Agentic AI

The company’s original positioning remains Data Intelligence in advertising and retail. Deep Miner and Octo emerged as Agentic AI products while Minglue refined its working methods, with a monetization model pointing toward Agentic Service.
吴明辉 expects knowledge workers to look increasingly like scientists. AI is good at deterministic reasoning from explicit premises—Think; humans remain difficult to replace in deciding what they want, what is beautiful and what is worth doing—Taste.
Einstein pursued relativity and a unified field theory not because he knew in advance that he could prove them, but because “he wanted this thing.” The scientist’s core role is therefore not calculating faster than AI, but forming a new first principle unknown to AI.
When Minglue enters a new vertical, it still relies on 2 legacies: its existing advertising and retail data and customer network, and the teams that understand those customers.

42. Scaling Out is 吴明辉’s counter-thesis to centralized Scaling Up

He criticizes foundation models for growing continuously through Scaling Up while remaining black boxes that even researchers can understand only by guessing from experiments. If a small number of servers handle all intellectual activity, the result could be mass unemployment and a damaged economic cycle.
An executive at a large-model company once told him that he might be “the person who eats the last piece of meat”: subordinates’ work could already be done by AI, and eventually even highly paid managers would be laid off. 吴明辉 believes that end state is “definitely wrong.”
His alternative is to make foundation models “good enough” at their current level and then Scale Out—distributing intelligence horizontally across individuals, small teams, specialized Agents and organizational networks, with intellectual property remaining with contributors.
The political and commercial implications are aligned: vertical teams should use their own context and Agents to outperform the centralized model in specific settings without automatically handing new knowledge to a handful of oligarchs.

43. The Neanderthal analogy supports the multi-Agent route

曼琪 asks why people should stop improving single-model intelligence when they often cannot anticipate new uses for a stronger model. 吴明辉 clarifies that he is not opposed to smarter AI, only to relying exclusively on scaling a single model.
His analogy is that Neanderthals may have had stronger individual capabilities, while modern humans won through collaboration. A single lobster need not outperform the next foundation model if humans and multiple Agents organize more efficiently and produce stronger results in specific domains.
He calls this the “poor man’s” training approach relative to foundation-model companies. Smaller companies cannot afford Scaling Up, but can build an advantage through specialized Agents, shared context and small models.
So would OpenAI’s GPT-6 or Anthropic’s MesoS, as he called it, change his view? 吴明辉’s answer is “it would not affect” it, because the two sides are taking different approaches.

44. The model plan centers on multimodality, small models and low-barrier training

Minglue does not plan to train a large foundation model in the near term. It will continue pretraining small models and plans to open-source a BUA dataset called Web Retriever, which 吴明辉 says could become one of the largest datasets of its kind in the market.
The main vertical opportunities are multimodal. Advertising comprehension, medical imaging and specialized GUIs still contain large volumes of data that have not been fully absorbed.
Future models may cover ASR, language and Coding, but the goal is not to enter a price war in general foundation models. It is to demonstrate Minglue’s model-training paradigm and apply the capability to Agentic Service in its own markets.
He cites Meitu as an example, referring to roughly RMB900M in profit and an approximately RMB20B market capitalization in the prior year, and argues that its revenue already has a clear token-like character.

45. The real benchmark is scientific discovery, not leaderboard performance

Commercial validation is relatively mundane: can revenue scale, can profit appear, and can Minglue build financial results through Agentic Service? The higher bar is whether AI can move human knowledge forward.
吴明辉 is planning work in AI for AI, AI for Math and AI for Energy, and says he is preparing a paper that may be submitted to Nature. The program disclosed no experimental results, so this remains a goal rather than an achieved result.
He uses the Goldbach conjecture as an extreme example: if several people specializing in algebra, number theory and geometry share context with multiple Agents and solve a problem beyond the reach of one researcher, that would demonstrate the scientific value of Scaling Out.
His timing remains at the scale of his original remarks: scientific discoveries will “appear within 3 years,” commercial success “may come faster,” and some results will be open-sourced over roughly the next 8 months for the market to test.

46. He summarizes the nearest-term risk bluntly: “Don’t drop dead”

吴明辉 observes that his high-school classmates and heavy lobster users around him generally sleep less and remain mentally hyperactive. His solution is not to reduce usage, but to “unite the brothers” and share the burden through teams and cross-time-zone handoffs.
The broader risk already exists: any company or individual can be overwhelmed by technology and large-model companies. Octo and Scaling Out are his attempted defense.
Large enterprise customers move more slowly but are equally anxious. He plans to share this organizational philosophy with Global CEOs at Fortune 500 companies because it explicitly takes the side of “people who appear to be getting bullied by AI today.”

47. The first reason the 2020 EIP failed was that the technology was about 6 years too early

吴明辉 studied foundational mathematics at Peking University from 2000 to 2004 and completed a master’s from 2004 to 2007, initially working on computer vision for fingerprints and palmprints before shifting to NLP late in the program. His master’s thesis was titled Recommendation System Based on Language Model.
He returned to Peking University for a PhD in 2019, saw that Transformer had arrived and judged that NLP would be overturned by deep learning just as computer vision was after AlexNet. EIP was intended to train on all enterprise data and become every employee’s best assistant.
The direction resembled today’s, but the stack was immature. The team was still working around BERT, knowledge engineering and customized ASR/NLP, without GPT-style pretraining or models capable of supporting Agentic AI.
In the AGI framework he relayed, EIP actually required today’s Agentic Use and organizational collaboration capabilities. “I was 6 years early” is his primary postmortem of the investment failure.

48. The $500M it nearly raised might have locked Minglue into Scaling Up

In the second half of 2021, EIP had a demo in production, and 吴明辉 says it came close to raising another $500M at a $5B valuation.
The capital environment then shifted. Dollar investment stopped, and domestic capital failed to follow. When Shanghai entered lockdown in 2022, customer collections and invoicing were also disrupted, forcing the company to accept that no new money was coming.
When 曼琪 asks whether that may actually have been a good thing in hindsight, 吴明辉 says yes. With $500M, he likely would have continued investing in large-model pretraining and followed the existing Scaling Up path; the best outcome might have been becoming another Kimi or MiniMax.
He is now more satisfied with a route that has controllable costs, can be commercialized and is, in his view, more favorable to humanity.

49. A highly original project paired with a 1,000-person team became an organizational disaster

During EIP, the company assembled roughly 1,000 people to work on organizational intelligence, including a newly hired model team, Minglue, 秒针 and acquired teams.
The legacy teams focused on knowledge engineering, rules and manually built knowledge graphs; the new team focused on BERT-style models. ASR, NLP and applications each followed different working paradigms. All of them had to build a product with no external template, one that even the founder could describe only abstractly.
吴明辉’s later conclusion was that a rocket with a defined target or a mature software product can use a large team and division of labor, but genuinely original work with no target is best handled by “the smaller the team, the better.”
That is the organizational rationale for Octo’s small core backbone and its incubation of peripheral applications by micro-teams.

50. The 2022 layoffs made “the existence of people” an organizational first principle

At the tightest point for cash, the company had roughly 4,000 people. If customer payments remained stalled, cash on hand might last only 2 months. After modeling borrowing and the survival runway, management decided to cut about half the workforce.
吴明辉 says he borrowed more than $10M for the company and still could not protect everyone. Employees put up banners, while some families pursued legal action through lawyers. For the first time, he faced the real conflicts of interests beyond an ivory-tower narrative.
An old classmate from Peking University, a former next-door dorm neighbor and a company founder was also laid off, then asked, “Brother Hui, can you lend me some money?” 吴明辉 was short of cash himself and could ultimately lend only about RMB100,000–RMB200,000.
The experience made him unwilling to carry out mass layoffs immediately after this round of AI productivity gains. At a macro level, he also asks: if every company cuts staff at the same time, “who will consume?”

51. The new organizational principle starts by recognizing that everyone has different goals

吴明辉 summarizes the first organizational principle as “everyone must exist.” The form of that existence can vary, but should at least include having one’s contribution seen, receiving recognition and seeing personal income rise as the organization develops.
He invokes Herbert Simon’s organizational theory to explain that employees, managers and owners each have their own purpose and can never align perfectly with the company’s goals. Owners want vacations and personal hobbies too.
Looking back at the management books he read during the EIP period, he rates his understanding at 1% then and barely passing today. The setbacks gave an entrepreneur who had long enjoyed favorable conditions the negative feedback and employee perspective he had lacked.

52. Complex battlefields require reading the environment before choosing an organizational method

吴明辉 says the world keeps changing and no individual can possess all the information, so decisions must be made under bounded rationality. Different terrain cannot support the same strategy-making method or team structure.
He cites The Essence of Strategy and the “strategy palette,” arguing that such frameworks matter because they indicate which strategic approach fits which conditions.
He also uses von Neumann’s game-theory framework, combining 5 binary dimensions such as complete versus incomplete information and cooperation versus non-cooperation into 32 situations. Business is a battlefield; before entering one, the first task is identifying which game is being played.

53. An AI-native generation will show the model its screen instead of copying an error message

吴明辉 watched his daughter debug a program and assumed she would copy the IDE error into 豆包. Instead, she directly asked 豆包 to look at the entire screen and suggest a fix.
The action changed his understanding of GUI models. Previous-generation internet users were constrained by habits of copying, pasting and switching windows; an AI-native generation naturally treats the screen as shared context.
His 2 children now learn from AI and use it to make plans. 吴明辉 also gives them his lobster and philosophical system, trying to teach them both that AI is powerful and that “AI should not replace people.”

54. Math education should shift from drilling problems to distinguishing the beautiful from the ugly

During the pandemic, 吴明辉 intensively trained his children for Olympiad math. Over the past year he has sharply reduced problem drills and shifted toward exploring the world, acquiring context and developing taste.
He prefers problems that are “extremely difficult, but once you understand them can be solved in one line.” Problems requiring mechanical enumeration of many cases may be labeled Ugly. During the Qingming holiday, he discusses with his children why a problem is Ugly and then skips it.
曼琪 points out that this could lower their Olympiad scores. 吴明辉 openly concedes that they “definitely will fall,” but believes that if one of the children genuinely wants to become a mathematician, the odds may actually rise.
Competition problems are guaranteed to have answers. A mathematician may devote a lifetime to a problem whose solvability is unknown. What matters then is not mastery of a pattern, but believing that a conjecture or formula “is beautiful.”

55. Career choices should follow interest because every discipline still has its own taste

吴明辉’s son is more interested in physics, chemistry, astronomy, life sciences and Coding, and may not love the purely formal sciences. Mathematics can serve as a foundational capability without determining the eventual field.
His earlier advice was, “If you don’t know what you want to do, study mathematics first,” like physical training in the digital world, then learn the skills needed at graduation. He now has more faith that every industry will remain and that the humanities and arts will regain value.
What art truly requires is a distinctive experience of the world. The purely Think-based portion of every profession is at risk, but deciding what is worth expressing—Taste—remains human.

56. Programmers will not remain in their current form, but context in complex systems remains expensive

吴明辉 believes programmers “in the sense of the current occupation” will be replaced by AI, but new roles will emerge: managing expensive tokens, defining boundaries, supplying historical-system context and tasting AI-generated architecture and products.
If an engineer consumes roughly RMB20M in tokens a year, that person should not be valued merely by manually written lines of code. As with a Researcher managing compute worth hundreds of millions of yuan, judgment determines whether expensive resources are used correctly.
Mission-critical systems in aviation, railways and Oracle have accumulated extensive legacy and cannot be casually torn down like a first prototype. An older engineer’s contribution may simply be telling AI “what not to touch,” but that can prevent a massive accident.
The external environment is also being changed in real time by other teams’ code. AI can read the current repository but may not know how another organization is changing an interface; humans still need social interaction to obtain those boundary conditions.

57. Product requirements have no single answer, so deterministic reasoning cannot close the loop

曼琪 cites Mike Krieger’s observation that once web coding compresses development time, the real difficulty is deciding what functions a product should have and putting the prototype into the real world to gather feedback.
吴明辉 treats engineering implementation as an ontological practice: the product manager defines concepts, data tables and relationships to construct an imagined world. Requirements analysis is epistemological: it repeatedly interacts with users and reality.
The first version often proves the original concept wrong, but also maps the terrain. The second version is rebuilt from that new understanding, then evolves into a third as competitors, users and the environment change.
AI knows existing code and computing capabilities, but not what each person truly wants. “Requirements analysis has no single answer,” so Taste can be assisted by AI but cannot be eliminated by one optimal solution.

58. Financial realization remains in the unvalidated section of the exponential curve

曼琪 points out that Minglue’s 2025 revenue and profit grew only by single-digit percentages versus 2024, with no jump matching the scale of its narrative. 吴明辉 declines to provide specific financial guidance, citing the rule against forecasting the performance of a listed company in advance.
His answer remains an exponential-function analogy. He describes a Tipping Point of roughly 0.1, after which network effects emerge; his remarks about the “intersection with the Y-axis” were ambiguous and cannot support a clear formula or financial forecast.
His scale vision is not unlimited headcount expansion, but a workforce of 1,000–2,000 people. The target is annual revenue of roughly RMB5M per employee and RMB1M in profit contributed to the company; he explicitly acknowledges that reality remains “a huge Gap” from that state.

59. Hope is a coordination mechanism in his system

曼琪 asks whether this framework contains too much values and Hope, potentially undermining objectivity. 吴明辉’s answer is: “Hope itself generates power.”
One person’s belief that humans should retain value may not change the system. If 7.9B of the roughly 8B people in the world share that Hope, the collective choice itself could alter institutions, products and resource allocation.
His argument is not that the future must unfold this way, but that choosing axioms creates a path. If people choose to protect individual intellectual property, employment and the right to taste, a different equilibrium from centralized Scaling Up may emerge.

60. His confidence comes from the “infinite game,” not a guarantee that every judgment is right

吴明辉 says he believes he has value even at the lowest point. If he cannot make a company work, he can teach Olympiad math; if mathematics is no longer needed, he can play piano and guitar, and in the worst case become a street performer.
When a technology judgment fails, the key is not preserving the original conclusion but checking whether new context still supports the research. If no new conjecture can be generated, one should abandon that battle and move to another field.
Borrowing 王兴’s phrase, he calls entrepreneurship an “infinite game”: a specific project can end without the company or the individual ceasing to exist. After EIP failed, he opened another game through Octo and a new understanding of technology and organization.
What investors are left with is not a guaranteed realization date, but 3 testable indicators: whether Agentic Service produces scaled revenue and profit, whether the Scaling Out model can achieve a material edge in new settings, and whether a scientific discovery actually appears within 3 years.

Anthropic Buys Compute From Elon & Commits $200BN to Google | Cerebras IPO | Ramp Raises at $40BN Is The Soft Landing Officially Dead?