An AI Founder’s Vanity, Posturing, and the Peak of Foolishness | A Conversation with invoko.ai Founder 梦琪
Summary
- 梦琪 has gone from believing that “anyone building software has rocks in their head” to cautious optimism about AI software. Coding agents eliminate the barrier to shipping a first version, not the work of stable delivery, maintenance, or building a brand. Of the software products that went through PR over the past year, she estimates fewer than 10% were genuinely usable, stable in practice, and smooth. The new client took 3 months to build and launched its 49th version the night before release, with bugs still outstanding. The investment takeaway is not that model risk has disappeared, but that moats are moving from ideas and code to “15,000 details” and the continuous work of tending the lawn. 曲凯 adds that products including Manus, Jace, and Genspark have already begun to build brand effects.
- Six months of firsthand experience overturned her strongest fundraising thesis: that professional users would keep specifying an Agent until it became a SaaS product, while enterprise customers who did not use the product would force the team to become an Agency. Sourcing represented only about 30% of influencer marketing, and 70%-80% of that depended on API data quality. The real bottleneck—communication and delivery—remained unsolved, leaving some peers’ AI products with little more than a “spectacle effect” designed for investors. 曲凯’s question was sharper: what teams call expanding into new scenarios is often just a way to postpone admitting that the original direction does not work.
- The core of the pivot was not changing sectors, but changing the founder’s objective function—from fundraising, an impressive roster, and new concepts back to whether users actually get results. 梦琪 acknowledges that her early talk of RL and scientist hires was about “chasing the trend, showing off, and satisfying vanity,” and calls the moment just after raising money “the peak of a founder’s foolishness.” She now argues that most pure application companies do not need algorithm engineers. Funding, valuation, and a star team are neither necessary nor sufficient conditions for success.
- The To C pivot did not produce a proprietary idea; it reduced the distance from user intent to an outcome to a single click, which is why users still choose a plugin like Click despite there being “a dime a dozen” substitutes. Users once had to take a screenshot, send it to ChatGPT, write a prompt, and copy the result back; now they press a button. She now sees “experience” as the core of consumer software and believes products in this generation may need to devote roughly 70% of their effort to interaction and emotional value. Commercially, the team chose a “gym” subscription model rather than serving a handful of users consuming $50K a month in tokens, with gross margin preservation as the goal.
- Real product discipline comes from high-density user research, not from stuffing products with buzzwords such as Conversation UI, GUI, TUI, or digital employees. The team breaks every decision into a funnel, posts roughly 40-50 threads and comments on Reddit for each question, and tests pain points, willingness to pay, and priority. The process has made a founder who once generated “800 directions” a day more restrained. First principles have been rewritten as “the shortest path to solving the problem,” with the strongest pull reserved for the people who actually keep the company alive: its users.
- The abandoned influencer-marketing direction found a way to deliver outcomes as a skill after the team used it continuously in-house: the output was not a dashboard, but an Excel sheet with screening, negotiation, and scheduling already completed. Because the team’s own context was continuous, it added 7-8 communication and metric details every day. The current result may already outperform the To B product on which the whole team spent 5-6 months, and the shared tool has begun generating revenue. It points to a narrower but more credible path for vertical AI: build a digital employee internally first, then let it spill over into an external product, rather than selling a platform upfront.
- She still believes there is enormous room in AI software, not because of technical superiority but because penetration remains low, user segments are deeply stratified, and commercial competition is never settled overnight. At a San Francisco event, she estimated that only about 40% of attendees knew Wispr Flow, and perhaps only about 20% of those people used it or paid for it. Many enterprises have yet to deploy even a real workflow. Even after the major platforms take most of the profit, applications may still capture more “soup” than SaaS companies used to capture in “meat.” The biggest risk is that product lives shrink to just over a year and planning horizons collapse to next month. But “when Douyin came out, Kuaishou did not die that same day.”
Deep dive
1. After the Software Funding Boom Fades, Products Face a Real Audit
曲凯 describes the current environment as another trough for software, but not the same kind of trough as the industry-wide weakness in 2023 and the second half of 2024. Embodied intelligence was still hot in 2024; embodied intelligence and AI hardware surged in 2025; and embodied intelligence is even hotter this year than last year. Software, by contrast, is colder than last year, with capital and attention clearly diverging.
He summarizes the standard playbook of the past 2 years: ByteDance backgrounds, high-P hires, a marquee team, and an Agent narrative, followed by several fundraising rounds before the product even ships. When the small product fails to gain traction, the company pivots again and again toward the next hot concept. After hearing the description, 梦琪 says bluntly: “You put me on the grill the moment you started describing it that way.”
梦琪 credits Manus with making it possible for software companies to raise serious money after last year’s May, month 5: “Everyone should kneel and thank Manus.” But most of those products were To B Agents. The product form changed, while the sales logic did not: “In the end, you’re still flattering the old guard.” High valuations did not automatically produce a new business model.
2. Coding Agents Compress Code Production, Not Software Delivery
By February or March this year, 梦琪 could barely answer 3 questions: how her product differed from Claude Code, what would happen if Claude Code handled the task itself, and what would happen if someone in the Claude ecosystem built a competing skill. For a while, she concluded that “anyone building software has rocks in their head.”
She relays what a post-training colleague at OpenAI told her: at the end of last year, the company barely used coding agents internally; by February this year, agents were writing roughly 80% of the code. Engineers were not working less, however, because they were simultaneously being asked to complete more. Capability gains and workload gains arrived together.
She then offered a counterexample: countless software products went through PR over the past year, but fewer than 10% were genuinely usable, stable in practice, and smooth. Many people say General Agent is just a sandbox or that they thought of it years ago, yet more than a year after Manus launched, only a small number of products have achieved comparable engineering quality and stability. 曲凯 believes products such as Manus, Jace, and Genspark have also built some brand equity; users may not want to keep trying a third product.
3. Avoiding the “18th Manus” Does Not Automatically Produce the Right Vertical
Before incorporating, the team first had to choose between a General Agent and a vertical Agent. Seeing an “18th Manus” already on the market, it decided General was too crowded and committed to vertical. At the time, 梦琪 thought the logic was airtight and that she was “brilliant and invincible.” She later found that the decisions she felt most certain about were often the ones that failed most completely.
Her preferred use case was a sourcing Agent: brands finding influencers, companies finding candidates, enterprises finding experts, and businesses finding customers all looked like variants of “finding the right person.” The team chose influencer marketing as its first wedge, hoping to test whether sourcing could transfer across scenarios.
曲凯’s interruption is worth preserving: the truly homogenized thing today is not the product but the narrative. Sourcing is already “all over the street”; similar stories have been told for years, yet few products have delivered good results. Ideas nobody has thought of are not scarce. Products that can execute are.
4. Professional Users Will Specify an Agent Until It Becomes SaaS
2 months after launch, the team was receiving 2 completely opposite types of feedback. Founders said, “I don’t want to click anything—just give me the result.” Professional marketers wanted every step to be auditable, every metric definable, and every filter controllable; they required confirmation of the intermediate process before allowing the Agent to continue.
The team decided to distill expert knowledge first, then sell “high cognitive capability” to non-experts. After another 1-2 months, the product increasingly resembled the SaaS products that had been built in the industry for 8 years. Without proprietary high-quality data, the team could only call on cleaned APIs from others and build an approximate interface on top. 梦琪’s verdict: “That’s just insane.”
She later understood this as a structural outcome. Ask a professional video editor what Agent they need, and they will keep adding requirements until the product looks like Adobe. Professional users need rigorous process control, so the product inevitably turns into SaaS. Eventually, it may be worse than a traditional SaaS product with an AI bar added on the right.
5. Sourcing Captured AI’s Easiest 30%, Not Its Highest-Value Work
After breaking down the workflow again, the team found that sourcing accounted for only about 30% of influencer marketing. The real waste of time and labor came from repeated communication afterward. Worse, 70%-80% of sourcing performance depended on API data quality. The team had selected AI’s low-hanging fruit in the value chain while skipping the higher-value work.
The company originally wanted to replace an Agency: whatever the Agency delivered to a client, the Agent would deliver as well. But once the organization began revolving around professional users’ process-heavy demands, delivering an end-to-end result became difficult. Publicly rejecting those demands also looked “politically incorrect.”
曲凯 asks why the team did not first make a small scenario work. 梦琪 initially answers, “Purely to make more money by expanding.” After 2 more rounds of questioning, 曲凯 points out that the team had already concluded sourcing was not good enough, and was using expansion to postpone rejecting the old direction. 梦琪 maintains one distinction: the concept was not invalid in a binary sense; its value was simply “not strong enough.”
The idea of expanding horizontally into influencers, candidates, and customers was also rejected. 梦琪 compares it to flour and eggs: the same ingredients can make a pancake with an egg or pasta, but nobody walks into a restaurant and orders both. The needs are difficult to cross-sell, while a Founder who has all of them does not personally use the product. The team therefore went deeper into growth channels such as GEO and Reddit, but by December had begun questioning whether the vertical Agent thesis worked at all.
6. The Structural Endpoint of a Vertical Agent Is Often an Agency and a “Spectacle Product”
The first trap 梦琪 observed was that when customers do not use the product but the company still has to make money, the company must send people in to provide the service. During the transition, the vertical Agent is forced to become an Agency. Some peers take the idea further: they build the Agent to satisfy investors, while real revenue comes from human services. Even the internal service staff do not use the product. The product is left with only a “spectacle effect.”
The second trap remains To B. She says To B in China has become almost a “derogatory term.” After spending more than a month researching the US, she concluded that Chinese founders also face a clear ceiling when doing To B in America. 曲凯 does not attribute that to Sino-American tensions: “It would be just as hard for an Indian to come to China and do To B.”
Asked by 曲凯 why the team did not pursue Pro C, 梦琪 admits it had confused Pro C with OPC. OPC is a trend, not a reality; in reality, OPCs are rare, and even fewer make money. “If he hasn’t made money, why would he pay you?” Pro C users with the strongest information-access capabilities may simply use Claude Code effectively, leaving limited room for an intermediate product.
7. AI Subscriptions Divide Roughly into “Heavy Users” and Gyms Betting You Won’t Show Up
梦琪 divides the available businesses into 2 categories. The first is a high-ARPU token business: serving a small number of “big R” users who consume more than $50K a month. After studying Lovable, she believes its user base is not as large as outsiders imagine.
The second is the “gym”: charge a subscription fee and bet that users will not come every day, or at least will not consume all their tokens. Credits preserve gross margin, but intense competition makes it difficult to build especially thick margins. 梦琪 says the team has chosen this model rather than serving a tiny number of users with unlimited consumption. 曲凯 adds that many AI companies have high average contract values, and 10,000-20,000 Power Users may be enough to support the business.
Her To C learning process offered another possibility through Steam. Pomodoro timers, GTD, habit-building, and ADHD tools can become companion-like, gamified products that users open by default every day, even though it becomes difficult to define whether they are games or productivity tools. As the cost of creating software falls, “categories will blur. Your ceiling is your imagination.”
8. To B Was Stopped Not Because Orders Were Missing, but Because the Odds and Model Upside Were Insufficient
By February, the team put the To B product into “low-power mode,” retaining only a small maintenance staff. 梦琪 believes that if she were simply going to continue doing To B, she could do it using her past customer base and industry relationships. She would not need to raise money, and “shouldn’t be starting a company.”
She compares an opportunity that does not make her excited to selecting a square from above a chessboard. After running through the logic, she may find that an opportunity suits her, but that is entirely different from saying, “I’m doing this today, whether you fund me or get lost.” Founding a company should pursue high odds, and should not become a proof exercise designed to demonstrate whether she is good enough.
The more practical problem was that To B Agents had not fully captured the benefits of model improvements. The biggest help from new models may have been higher internal development efficiency through coding agents, not a 10x or 20x improvement in solving customer problems. The team had previously used “too much left brain,” deriving concepts every day without seriously asking what it actually wanted to do.
9. The Beautiful RL Story Ultimately Exposed the Founder’s Vanity
曲凯 revisits 梦琪’s early narrative around RL, algorithms, and new technology, then compares it with a product that later looked much like traditional SaaS and asks whether this was a process of “gradual decline.” 梦琪’s answer is more direct: the sexy logic may simply become unsexy when deployed, and she initially chased it mainly to “follow the trend, show off, and satisfy my vanity.”
She did hire scientists who later joined US research labs. Her current “hot take” is that most pure application companies should not hire any algorithm engineers: “Let them go train models.” Application businesses generally have no use for them. More hiring often serves to show investors a prestigious lineup or add leverage for the next funding round.
The RL logic sounded airtight at the time. Growth content naturally generated feedback signals such as exposure and engagement, making it easy to construct a reward function. 曲凯 points out that a foundation-model vendor making a small change could create more value than a startup’s own training. 梦琪 admits that she now finds the story laughable when she hears someone else tell it. Immediately after raising money, she thought she was exceptionally good at creating concepts. That was “the peak of a founder’s foolishness.” Reality demanded that she “bend over and deliver one thing at a time.”
10. San Francisco Disenchanted Her with Both Vertical Agents and Revenue Padding
At her most anxious point, 梦琪 spent more than 2 months in San Francisco trying to determine whether she lacked the ability and imagination, or whether the vertical Agent itself had a problem. After speaking with multiple companies, she concluded that they were facing similar structural constraints, not merely suffering from poor execution by individual teams.
The trip also thoroughly disenchanted her with the US startup ecosystem. The simplest version was, “I’ll pay you $50K a month, and you drive the revenue.” More complex versions involved networks of 6 or even 10 people who bought one another’s products in a loop, making it difficult for background checks and audits to identify “revenue padding.”
She finally understood why so many people suddenly became unusually warm and repeatedly invited others to meals near the end of each month: they needed someone to help push that month’s revenue. The team did not participate because payment was cumbersome. More importantly, she felt that it “didn’t make much sense” and did not want to plant a land mine for herself, preferring to preserve its ethical standards as far as possible.
梦琪’s conclusion was that if a startup exists only to serve vanity, the objective function is clear and the constraints are few: funding, star hires, and attractive customer logos are all achievable. If the goal is to make oneself honestly admit, “I did a really good job,” the task becomes much harder.
11. Click Validated a Minimal To C Need, While 49 Versions Shattered the “One Week to Build” Illusion
Before shutting down To B completely, the team let some members test To C first because “team morale and confidence are nonrenewable resources.” A founder cannot repeatedly consume everyone’s trust. 梦琪 preferred a gradual transition to announcing one day that everyone should bet on an unvalidated new direction.
The first version of Click was only a browser plugin. It pulled information across large numbers of tabs and let AI follow the cursor to write. One user in South Africa said he could finally become immersed in writing documents because he could directly @ a tab or Gmail context instead of allowing fragmented information to occupy his mind. That also hit home for 梦琪’s own ADHD experience.
Click was confined to the browser and could not cover desktop windows such as Lark and Slack, so the team expanded it into a client. 梦琪’s principle is that “all application companies are wrappers”; the key is not to wrap the previous generation of capability. The team therefore chose to “firmly wrap Claude Code.” Its people were also better suited to To C: they were “fiery” and willing to argue with customers, while the reality of To B was that “if you’re being kept, don’t talk about an independent personality.”
The client sounded simple but took 3 months to build, with version 49 launching only the night before release. During the public beta in late March and April, it could open and run, but 梦琪 still considered it basically unusable; early users had “the worst luck in the world.” She describes software as “15,000 details”: solving one more makes the experience a little better. That realization made her less afraid of coding agents.
12. The Value of a Consumer Product Is Closing the Gap Between “I Want to Do This” and “It’s Done”
梦琪 initially looked down on Click. It had only a few buttons, and substitutes were “a dime a dozen.” User feedback changed her view: an idea does not need to be unique. If it shortens the distance between user intent and a solution, it has already created value.
Previously, a user who wanted AI to rewrite the content in front of them had to take a screenshot, send it to ChatGPT, explain the task, write a prompt, and copy the result back. Click compressed the whole sequence into one button press. The button itself encoded the intent and the preset, so the user did not have to describe what they wanted a second time.
Translation plugins provide the same evidence. Even when a foundation model is powerful enough, users will still install a dedicated plugin because a deterministic entry point is closer than opening a general-purpose model. Once a product carries more intents, the real challenge becomes orchestrating its functions and preserving balance without letting information complexity overwhelm the user.
She uses transportation choices to summarize To C. Why do some people choose a premium ride instead of a luxury sedan, a carpool, or occasionally Shouqi? The answer is “experience.” As long as the experience is better than going directly to ChatGPT or Claude, an application has a reason to exist. She estimates that products in this generation may need to put roughly 70% of their effort into interaction, including speed, comfort, and the emotional value provided by the interaction itself.
13. Emotional Value and Reddit Research Became Product Infrastructure
To break through the wall between productivity and emotional value, the team brought in several US designers to refine the product. 梦琪 did not want to build a product that merely offered comfort. It had to solve the problem reliably first, then make the user feel that the product “cared about you.” Otherwise, it would be nothing more than “an unreliable subordinate who provides emotional value.”
Privacy is the client’s most direct emotional barrier. Users become frightened when they see extensive permissions; they do not know where their data is stored and will not read a long agreement. The team uses comics to explain how data is stored and adjusts its privacy policy wherever possible to make the disclosures easier to understand.
She believes this generation of Chinese founders generally conducts less user research than San Francisco teams. Relying only on personal intuition is risky when serving consumer users from different cultures. Reddit data mining can produce static conclusions, but it cannot dynamically pursue questions along the team’s line of thought. The team therefore also has to enter the discussion directly.
The team breaks every product decision into a funnel, with each layer mapped to relevant posts and comments. It publishes roughly 40-50 posts for each question and collects the feedback. Directly pitching the product tends to generate little traffic, so the team focuses more on needs and pain points, using the responses to judge whether the need is essential, whether users will pay, how much they will pay, and what is wrong with existing products.
14. First Principles Are Not About Using the Newest Interface; They Are About Driving the Nail In
梦琪 describes herself as an ENFP. In the past, she wanted to propose “800 directions” a day and pack every fashionable capability into the product. Reddit research has forced her to be more disciplined: having many needs does not mean every need belongs in the current version.
During the To B phase, the team was obsessed with Chat UI, GUI, Conversation UI, TUI, and digital employees. A user would paste in a product link, the system would automatically analyze the need and configure an employee, and chat would trigger a GUI. The team thought it “understood AI better than anyone,” while users had only one response: “You must be bored. Can you solve my problem first?”
She is now willing to spend roughly 70% of her time confirming what the problem is, whether the need is real, and which path is shortest before asking whether a new concept can help solve it. The order cannot be reversed. Nearly a year into founding the company, the strongest force pulling the team forward is no longer peers, investors, or employees, but the people who actually keep it alive—its users.
梦琪 compares the old approach to showing users an engraved titanium hammer. Users care only whether the nail has gone in. Her sharper analogy is that AI is like long underwear: founders use the condescending attitude of “I’m your mother” to force users to put it on, while the users are not cold at all. 曲凯 points out that, from the customer’s perspective, token consumption may cost more than an intern, while traditional SaaS is already good enough. 梦琪 says that was also how it felt at the time: AI’s economic value might simply not have been high enough.
15. Abandoned Influencer Marketing Became a Real Digital Employee Through Continuous Internal Use
While building growth for the new To C product, the team had to revive its old influencer-marketing capability and turn it into an internal skill. It added 7-8 new metrics or communication details every day, gradually filling in the corner cases. 梦琪 believes the current result may already outperform the To B product on which the entire team spent 5-6 months.
This time, “delivering an outcome” was no longer an abstract slogan. The deliverable was an Excel sheet containing the partner list, quotes, negotiated prices, value assessment, content review, shooting dates, and upload dates. Once the spreadsheet was delivered, the Agent’s mission was over; the client did not have to operate the intermediate workflow.
The internal product improved more quickly because its context was continuous and complete. External customer needs were fragmented: 2 sentences today, no time tomorrow, and 3 more sentences a week later. The team kept refining the skill through its own real work, then shared it with friends. Some responded, “This thing is actually pretty useful,” and it began to generate revenue.
The same logic is being applied to R&D. Software development has been reduced to 3 steps—raising a problem, solving it, and accepting the result—with attention concentrated at the 2 ends. A multi-agent system plans P0 and P1 tasks and distributes them; each member’s coding agent executes. 梦琪 calls it “Jira for a new era,” but the interface remains unresolved. She only knows that it “definitely won’t be chat,” so the team is not rushing to release it.
16. Models Shorten Product Lives, but Low Penetration and Maintenance Costs Still Leave Room for Applications
Model development is the biggest force pulling every software founder, repeatedly pushing teams to pivot. A client may require 3-6 months of polishing but have a life of only slightly more than a year, leaving just over 6 months to make money. Planning horizons have therefore shrunk from next year to next month. 梦琪 once pounded a table in frustration after debugging failed at a Blue Bottle in Palo Alto; the neighboring table instead asked about the product because her emotions were so intense.
She cites Bret Taylor, who builds a customer-service Agent, comparing software to a lawn: maintenance may cost more than laying it down in the first place. After using Claude Code to build a small tool, an individual has already enjoyed the creative payoff; maintenance and corner cases are the painful part. AI might eventually discover and maintain those issues itself, but she explicitly leaves that as a possibility rather than an established fact.
Her San Francisco fieldwork showed the other side. She asked everyone she met about products such as Wispr Flow and Tablets; roughly 40% knew them, and perhaps only about 20% of that group used or paid for them. She also cites another random survey asking whether people used ChatGPT, emphasizing that users cannot all be assumed to be typical early adopters. An application does not need to call itself an all-purpose AI; it only needs to solve an unsolved problem.
Once, because of technological bias, she insisted on choosing an Agent rather than an agentic workflow. She later admitted that most enterprises would solve the majority of their problems simply by deploying the workflow. After Harness appeared, 梦琪 felt overwhelmed by the information flow. She opposes reducing dynamic competition to an “AI-sector Pomodoro” or “AI-sector Hongguo” that can be understood in a few seconds. The world is highly stratified, and even after large platforms take most of the profits, the soup available to applications may still exceed the meat available to SaaS companies in the past.
17. The Real Pivot Was Not the New Product Launch, but the Founder Finally Believing in Her Product
The team did not stage a “world’s first” launch. Partly it wanted to stay quiet, and partly 梦琪 believes most launches buy only a wave of traffic and weak retention, like paying the market to complete an assignment. The deeper reason was that she was neither satisfied with nor convinced by the old product. Now, even though the new product is still ugly and has problems, she is willing to give it everything because “this is my spark.” The task is a short-answer question, not a proof exercise.
Asked again whether she should have left ByteDance, her official answer is that she would still found a company because it was “meant to be.” Her unofficial answer is that in March and April she considered flipping the table and thought software was meaningless. Her current position is the opposite: “If I can’t build it, nobody can; if even one person in this industry can build it, it will be me.”
曲凯 contrasts the present with more than a decade ago, when an angel round might invest RMB1M at a RMB10M valuation, followed by another round after the product launched and generated data. 梦琪 says that getting funded in the current cycle was both lucky and “shameless,” but she has confirmed that funding, valuation, and a star team are neither necessary nor sufficient conditions. Entrepreneurship cannot be reduced to scientific variables; you can only “play the cards you have as well as you can.”
What she ultimately wants to build is not merely an efficiency tool that looks at the problem. It should solve the problem while caring about who the user is, how they feel, and what relationship they are in: “DAU is not a number to me. Every number represents a person.” 曲凯 summarizes the past year as a move from abstraction to execution. Their shared conclusion is that good founders still need idealism—just not the kind that uses abstract concepts as a substitute for delivery.