Summary

This personal episode is not an annual launch retrospective, but Koji’s 9 Aha moments, distilled on June 20 from frontline conversations and product trials; the views are intensely personal.

DeepSeek and Qwen3 turned model capability into a public resource, giving AI application startups a “somewhat fair” starting point. 妙鸭 founder 张月光 raised more than $300M and still complained, “This isn’t a fair competition,” because founders cannot guarantee that the model they use is the best available; after DeepSeek open-sourced its model, intelligence and cost both moved downmarket, and US H100 rental prices even reversed their full-year decline to rise 10%. Qwen3 then added a full range of model sizes, easing the problem of using R1 or V3 directly when they are too heavy, or distilling them independently when the technical bar is too high. The playing field is only relatively level, however, and will stay that way only until a newer closed-source SOTA model appears.
When models iterate monthly and technical moats can be wiped out overnight, the real accumulation point for To C products is execution speed and momentum. The line on Manus’s wall reads, “There are no secrets, only pure execution speed.” After struggling to raise money and being rejected by many dollar funds, Manus launched to massive attention and secured Benchmark’s investment at a $500M valuation, showing that products like Perplexity, Cursor and Monica can build large user bases without owning their own models. Koji reduces the moat to four things done repeatedly: understand users, pursue an exceptional experience, compound word of mouth and iterate fast—in other words, “focus, excellence, reputation and speed.”
Founders should not stay out of the arena just because a product looks like Talking Tom Cat or a “wrapper,” because understanding a model’s boundaries can only be built through action. Perplexity confronted the objections head-on in its Series B deck: “Conducting a symphony orchestra is not easy,” “Building a wrapper also takes a lot of work,” and “The devil is often in the details.” A good wrapper is, at its core, product craft that makes the most of a new capability. Koji endorses the path of “questioning Talking Tom Cat, becoming Talking Tom Cat, then surpassing Talking Tom Cat”: no one can skip straight from age 1 to 17 and grow into 18. When the wave arrives, “learn to swim early and get used to the taste of seawater early.”
The opportunity in AI Agents remains underestimated because an Agent does not simply produce a one-off answer faster; it understands, plans, asks follow-up questions, calls tools and self-checks like an employee. At the end of 2024, Koji paid $500 to try Devin and saw it expose its reasoning and work process while proactively reporting progress, leading him to bet that 2025 would be the Year of the Agent. More than 100 days later, Manus, Jan Spark, Fellow, Flow With and Lovart appeared in succession. Given only the instruction “McDonald’s × giant panda full VI,” Lovart planned 15 pieces of collateral and wrote “Bite into harmony,” showing the product potential of model orchestration and task decomposition.
AI’s biggest entrepreneurial spillover may not be that every product adds AI, but that any sufficiently nuanced need can finally be worth productizing. Sunlight, 魂旅 and Focus Flight barely use AI themselves, but benefit from lower development costs through AI design and Coding, recommendation algorithms that precisely reach niche users, and stronger user willingness to pay. The math is straightforward: a one-person product with 10,000 users each paying RMB10 a month generates RMB100,000 in monthly revenue—and can sell directly to users worldwide.
Once generative AI fills the gap in ad-creative production, platforms may evolve from traffic distributors into “ultimate business Agents” that handle creative, targeting and data optimization end to end. In theory, a merchant would only need to upload its product, budget and acceptable ROI; Meta, TikTok or 小红书 would do the rest. NextAd’s angel round, backed by a16z, Point72 and Prosus, also reflects capital’s growing consensus that AI will reshape advertising. Koji is unsure whether this will let companies focus more on their products or push competition further into the supply chain and toward ever-lower ROI, but believes the change “will almost certainly happen”—possibly overestimated over 2 years, yet still underestimated over 10.
As AI Coding matures, design, taste and distinctiveness are more likely to become scarce forms of productive capacity, giving designers more opportunities to build companies end to end. Y Combinator is particularly keen to see founders with design backgrounds because they combine user empathy, taste and the ability to make something different; tools are also erasing the boundaries between product, design and engineering. Designer 祝以南 built 日落岛 independently, extracting colors from sunset photos and turning them into collectible color cards—a concrete example of “beauty as productivity.”
Beyond efficiency, consumer AI products are also delivering emotional value, even “creating moments of love.” Moflin creates the feeling of a living thing through humming, slight movements, wireless charging and body warmth, and has recently sold for more than RMB4,000 on 闲鱼. 独享 already has 50,000 DAU and 40,000 organic 小红书 posts, with more than 10,000 people using “AI陪你睡觉” every day. Koji admits he may not personally empathize with the need, but the numbers show that AI can deliver real emotional value; Rapet founder 何家斌’s declaration—“I want to create moments of love”—became the year’s judgment that moved him most.

Deep dive

This is not an interview with a founder or a launch chronology, but a personal talk Koji recorded at the AWS China Summit on June 20, drawing 9 intensely personal Aha moments from frontline conversations and product trials.

1. Six Months Was Enough for Outsiders to Rewrite the Model Landscape

Looking back on late 2024 on June 20, Koji recalled that roughly 95% of big-model retrospectives at the time focused on the “Six Little Dragons” of AI, ByteDance and Alibaba. DeepSeek was merely an obscure outsider mentioned in passing. What attracted him was not its intelligence, but its unusual combination of being “legendary,” internationally known, low-cost and the first to ignite a price war.
The story that stayed with him most was 梁文锋’s firm refusal to commercialize, on the grounds that any commercial move would distract the team from its sole mission of pursuing greater model intelligence. Koji called it “a very pure and very luxurious team.” Six months later, DeepSeek and 梁文锋 had become archetypal figures of an era full of heroes.
妙鸭 founder 张月光 raised more than $300M and still worried: “The model you have isn’t the best, so this is unfair. This isn’t a fair competition.” What surprised Koji was that even an application founder with such abundant resources believed the competition was determined by the underlying model.
DeepSeek’s open-source release made SOTA models a relatively usable public resource. All-in-one machines were a mixed bag, but they did help bring models downmarket. H100 rental prices, which had been falling throughout 2024, instead rose 10% in the US after the release, showing that lower-cost models did not reduce demand for compute.

2. Qwen3 Extended “Model Equality” to Size and Cost

Using DeepSeek R1 or V3 directly is overkill for many applications; distilling a smaller model independently requires technical expertise. Qwen3 offers a complete range of parameter sizes, making it more likely that founders can find the right model for each use case—closer to “more choices, more fun.”
One investor even said, “It was only after Qwen3 that I truly dared to invest in AI applications.” Koji’s judgment is not absolute: DeepSeek and Qwen3 merely made the competition “somewhat fairer,” while the next updated closed-source SOTA model could recreate the capability gap.
This relative leveling of the field is a prerequisite for applications to bloom: model capability and cost are no longer controlled only by a handful of major technology companies.

3. Manus Proved That Speed and Momentum Can Precede a Traditional Moat

A piece of paper on Manus’s office wall reads, “There are no secrets, only pure execution speed.” Robert Scoble relayed a Silicon Valley VC’s reaction: Manus reminded him of an earlier Silicon Valley era that relied not on secrets, but on moving fast.
Koji has invested in the Butterfly Effect team’s previous company since 2016 and is now its advisor. Manus’s pre-launch funding round was not going well; the team had met nearly every Chinese dollar fund. Aside from supporters including 真格, Sequoia and Tencent, many investors said no. On launch day, however, Manus drew massive attention and secured Benchmark’s investment at a $500M valuation.
The prediction that “rising model capability will drown every application” has not come true: Perplexity, Cursor and Monica do not have their own models, yet have already attracted large user bases. For To C products, the momentum created by continuous innovation can itself become a moat.
真格 partner 刘媛 stresses that solving user problems from the bottom up and discussing moats from the top down are two different perspectives. Koji expands the former into understanding users, pursuing a finely tuned experience, compounding word of mouth and iterating fast enough. Over time, the answer returns to the internet’s seven-character formula: “focus, excellence, reputation and speed.”

4. Being a “Wrapper” Is Not a Sin; Refusing to Become Talking Tom Cat May Mean Missing the Wave

Talking Tom Cat took off in 2010 by mimicking speech, then faded as its gameplay proved too limited and retention fell. Later products including 哄哄模拟器, Perplexity and Monica faced similar criticism that they were flashes in the pan or uninteresting wrappers.
Perplexity addressed the criticism directly in its Series B deck: “Conducting a symphony orchestra is not easy.” Building a wrapper also takes substantial effort, because “the devil is often in the details.” Orchestrating different models, recognizing their respective strengths and weaknesses, and turning new capabilities into an experience is itself product craft.
Koji even offered a deliberately provocative take that he expected people to “go easy on”: the iPhone is, to some extent, also a wrapper around the new technology of multitouch. “Wrapper” is not a slur. The key question is whether the product fully and reliably releases the underlying breakthrough into the user experience.
An article by 吴秉坚 provided the corrective: founders cannot stare at the end state and try to skip the history in between. “No one can jump from age 1 to 17 and grow straight into 18.” The real path is to understand Talking Tom Cat, become Talking Tom Cat and surpass Talking Tom Cat. When the wave arrives, “the key is to stand in the sea”—learn to swim first and get used to the taste of seawater.

5. Product Locusts Can See the Agent Interaction Paradigm Before the Market

The first users of a new product are often industry professionals, product managers and investors. They consume resources without being the target users and generate noisy data, which is why they are called “product locusts.” Koji, however, believes that in periods of rapid change, trying products early lets people glimpse the future: “The future is already here—it’s just not evenly distributed.”
By the end of 2024, Devin was already showing the paradigm that later became common among Agents: fully exposing its reasoning and work process, proactively making plans, reporting progress, calling tools, and engaging in deep thinking, self-reflection and verification. Even with a $500 paywall, Koji decided to try it.
The experience gave him the confidence to title his New Year conversation “The Critical Year for AI: The Year of the Agent Begins.” More than 100 days after the episode was released, Manus fired the first shot, followed by Jan Spark, Fellow, Flow With and Lovart. Early products may not be mature, but they can serve as templates for understanding the next generation of interaction.

6. Lovart Showed How an Agent Can Expand One Request into a Complete Deliverable

Koji believes the change brought by Agents is still underestimated. Alongside Rock Flow’s financial Agent Bobby, he chose Lovart as an unusually complete example in design and asked it to create a McDonald’s and giant-panda collaboration, with a logo as only the starting point.
When the instruction became “Based on the logo above, create a complete VI system,” Lovart decided for itself what the VI should contain and delivered 15 pieces of collateral in one pass: an app UI, food packaging, highway billboards, and even the color and graphics for a delivery rider’s motorcycle. The user had not provided that list in advance.
The slogan on the motorcycle, “Bite into harmony,” particularly impressed Koji: “Bite” maps to the burger, “harmony” to the panda, compressing both into a line of creative work that sounds like it came from a professional advertising strategist rather than a simple image generator.
Mechanically, the Agent first understands intent, decomposes the task, then asks scope questions such as what the store’s visual identity should be before expanding the ideas and orchestrating multiple models. Lovart can connect to GPT Image 1, FLUX Pro, OpenAI o3 and Gemini’s Imagen 3, while also calling Kling for video, Tripo for 3D and Suno for music. “What you learn from books is ultimately shallow; to truly understand it, you have to do it.”

7. Once AI Cuts Costs, Niche Needs Can Finally Support a Sustainable Commercial Loop

An early image of a “horse-less carriage” made Koji laugh: it retained every feature of a horse-drawn carriage and simply removed the horse. When a new technology appears, the first tools often fail because they copy the old way of working. Creation therefore looks clumsy at the beginning, while criticism always looks clever.
After 小红书’s indie developer competition, he borrowed 多抓鱼 founder 猫柱’s line to write: “Criticism always looks clever, just as creation always looks clumsy.” The article was reposted more than 1,000 times. Koji attributed the resonance to a trend and an emotion: vertical needs are worth productizing, and everyone wants to become a super-individual.
Sunlight tracks whether Apple Watch users have had enough sun; 魂旅 lets people whose “bodies are at their desks but souls are traveling” simulate taking a green-train journey from Shanghai to Dali; Focus Flight creates distraction-free focus time through a virtual flying state. None uses AI, yet all would have been difficult to build at low cost before the AI era.
The commercial loop comes from three structural changes: AI design and Coding lower development costs; the recommendation algorithms of 小红书 and 抖音 find precise users; and Koji believes AI products have also trained users to pay. If 10,000 users each pay RMB10 a month, a solo developer can generate RMB100,000 in monthly revenue and sell directly to users worldwide.
Koji further stresses that founders should not add AI for its own sake, but return to user needs. He has long viewed himself as a product manager: he worked on 饭否 and 海内, then built 街旁, 新世相 and 糖岛; 糖岛’s 猫肚皮枕 and 呱呱凉被 later became category-leading products on 天猫, 抖音 and 京东. These experiences have only reinforced his belief that nuanced user insight, product design and marketing can still create value.

8. AI Advertising Could Rewrite the Power Relations Among Products, Platforms and Supply Chains

In 1996, Bill Gates reflected on Microsoft’s slow response to the internet by saying: “We always overestimate the changes that will occur in the next 2 years and underestimate the transformation over the next 10.” Koji suspects people are repeating the same mistake in their time horizon when thinking about AI.
In his Stratechery conversation, Zuckerberg put one of AI’s core implications for Meta bluntly: build the “ultimate business Agent.” Platform data analysis and audience targeting are already so powerful that manually selecting “women aged 18 to 24” can sometimes perform worse. The key unsolved problem is creative generation.
The theoretical end state is that a merchant uploads only its product, budget and acceptable ROI, while Meta, TikTok or 小红书 automatically generates creative, targets audiences and tracks the data. Koji does not know whether this will let merchants focus single-mindedly on refining products or force everyone to squeeze “every fraction of a cent” out of the supply chain to tolerate an ROI lower than competitors'.
NextAd, the AI-native ad-tech company where Koji serves as an advisor, raised angel funding from a16z, Point72 and Prosus; the latter two rarely invest this early. Koji sees the deal as a signal of industry consensus, but remains measured on timing: the scenario may not arrive within 2 years, while the change over 10 years could be deeper than imagined.

9. The More AI Coding Spreads, the More Valuable Designers’ Differentiation Becomes

When recruiting companies for its next batch, Y Combinator said it particularly wanted to see founders with design backgrounds: as products become easier to build, excellent design matters more. Designers bring three entrepreneurial capabilities—empathy with users, aesthetic judgment and the ability to create distinctiveness. Generic products rarely make users crave them.
The traditional boundaries among PRDs, UI/UX and engineering development are being broken down by tools including Lava Ball, Boat and Figma Make. Koji does not see AI only as a threat to designers’ jobs; he believes designers who know how to use the efficiency leverage are most likely to independently complete the entire journey from idea to product.
Designer 祝以南 built 日落岛 independently. Users upload sunset photos, from which the product extracts colors and generates collectible color cards. It bears out Steve Jobs’s judgment: “Design is not just what it looks like and feels like. Design is how it works.”

10. Consumer AI’s Value Also Lies in “Creating Moments of Love”

Moflin does not converse with people; it only hums and wiggles slightly, and it cannot walk or crawl. It deliberately eliminates an exposed charging port in favor of wireless charging through its belly, while retaining a small amount of body warmth. Koji’s daughter knows it is an electronic pet but still feeds it Yakult, saying, “We’re eating, and Mufflin is eating electricity.”
Moflin has recently sold for more than RMB4,000 on 闲鱼. 王登科’s 独享 already has 50,000 DAU and 40,000 organic 小红书 posts. More than 10,000 people use “AI陪你睡觉” every day: they place a phone with its screen gradually dimming face down beside the pillow, removing the temptation to keep playing, and can check the next morning what their AI pet dreamed about and how well it slept.
Koji’s honest reservation is: “I don’t empathize that much either. I don’t seem to need this feature.” But whether he empathizes does not change the fact that 10,000 people use it every day—the emotional value users need is something AI can already deliver very well.
Behind Rapet, the AI pet that became a hit at CES this year, founder 何家斌 gave the company a mission that was not market share but “I want to create moments of love.” Koji ultimately brought the 9 Aha moments back to the same crossroads: experience the future and jump into the sea to act, but also put down the phone and return to nature and everyday life, because “great products are often born at the crossroads of technology and the humanities.”

107. A Catch-up with Mengqiu: Venture Capital Is Pretty Boring—Plus Travel, Reading and Feminism Life Is One Big Sales Job: How to Become a Better BD? | A Conversation with 戴安琪 of a Tech Family Office