Original · Unique Research · 2026-06-01
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the complete narrative analysis, all three interrogation sections, the dynamic moat conclusion, and the full roundtable Q&A transcript. All named speakers, companies, roles, numbers and claims are preserved. Company, personal and product names are transliterated where official English forms remain unverified. Market projections, performance claims and company-specific figures are source or speaker attributions, not independently verified findings.
Unique Awards
The moat for AI going-global startups is not technology—it's how far you are from the big players' firing range and how deeply you embed into business workflows.
When the big players' firing range extends every six months, where do you plan to run?
"
Frankly, startup teams have no moat at all.
The person saying this is Phil, a Singaporean who sells AI digital employees. None of the three guests beside him contradicted him.
This was the opening atmosphere of an AI going-global roundtable at the Unique Awards. Moderator Wang Chaochao (CC), partner at Unique Capital, opened with a soul-searching question: The technology is cool, but how do you make money?
And the first thing they discussed was not technology, but the distance from the big players.
Four Ways to Make Money, but Only One Underlying Logic
Let's start with the most vulgar fact: all four of these people are making money now, but in completely different ways.
Angel Lam has worked at Google Cloud for 10 years, helping startups build their full going-global chain. Her logic is simple and brutal: Google has cloud, advertising and monetization—from building your product to going global to making money, it's a one-stop package. "Gaming, e-commerce, AI startups—we talk to all of them with one-stop services." This is the aircraft carrier posture, not competing on single points.
Liu Kun, born in 1998, graduated from Tongji University's architecture department, did peacekeeping in Lebanon during college, and now startups in AI marketing. His company Chengguo Shijie (橙果视界) was founded two years ago with a model called "outcome-based pricing"—helping brands acquire customers via social media on TikTok in Southeast Asia, Europe and America, earning commission. "When we defined this term, it hadn't caught on in Silicon Valley yet—it probably became hot there only last year."
In fact, he wasn't bragging. Outcome-based pricing in Silicon Valley indeed only became a buzzword in AI startup circles in 2025. Liu Kun was already doing it in 2023.
Yin Tianming, former product VP at 37 Interactive Entertainment (三七互娱), has been in gaming for 20 years. His company Senluo Zhihui (森罗智绘) is essentially an entertainment company—doing game publishing and short dramas. The AI creation platform was forced out by business needs. "As we went along, to empower our business, we developed our own AI creation platform. We didn't expect it to work this well."
Phil, founder of Canlah.AI, backed by a Singapore listed company, sells "digital employees"—ten AI Agents corresponding to ten positions in a marketing department. Data analyst, operations, content production, Amazon store optimization, each handling its own area. "The OpenClaw (小龙虾) on my phone—I don't even need to check what it's doing every day; it just produces what it should produce."
It sounds like each has their own path. But strip away the jargon, and there's only one underlying logic: using AI to help clients reduce content production and marketing costs, and taking a cut.
This raises a question that must be punctured.
Interrogation 1: Strip Away the AI Cloak—Isn't This Just Outsourcing?
Liu Kun admitted it himself.
"FTE is the hottest word lately. Essentially, in Chinese terms, isn't it just on-site staffing? It's just that during the on-site process, you distill customized client service into standardized SOPs, then use those to serve other clients."
Phil didn't hide it either. Moderator CC asked him directly: Are you selling headcount or selling outcomes?
"Right now, we're selling headcount."
So here's the question—what's the essential difference between Liu Kun, who earns commission by running ads for clients, and Phil, who charges Agent headcount fees, compared to the advertising agencies and labor outsourcing companies of ten years ago?
The difference lies on a steep curve.
Liu Kun gave a key number: AI's participation in delivery has already climbed from the initial 50% to 95%. "Maybe only 5% relies on humans. From the company vision perspective, we firmly believe that in half a year or a year, it can be 100% completed by AI."
The meaning of this statement needs to be unpacked.
If it stays at 95% forever, it's a labor-intensive outsourcing business—you still need to hire people, train them, and face endless revision demands from clients, with profits eventually squeezed to the floor. But if it truly breaks through to 100%, it becomes a money-printing machine with marginal cost approaching zero. The same AI system serving 100 clients versus 1,000 clients makes negligible cost difference.
Phil used a more intuitive metaphor: "When you first hire an employee, you pay a salary, right? But when they become exceptionally good and become my CTO, I can give them profit sharing. When they become even better and become a partner, I go and sell outcomes for them."
Translated: today's AI marketing companies are using the meager cash flow of outsourcing to buy an option on the future "sexy model" of pure AI delivery.
The bet is on when that last 5% goes to zero.
Interrogation 2: Beyond the Big Players' Firing Range—Is the Ceiling High Enough?
The most interesting undercurrent of this roundtable was that everyone was calculating their distance from the big players.
Phil put it most bluntly: "Google is very friendly right now, but every model iteration actually knocks down a whole bunch of startup teams, which means they're too close to the firing range."
Liu Kun's wording was gentler but meant the same thing: "Focus on doing things beyond the big players' firing range—things they're too lazy to do or don't deign to do."
Even moderator CC summarized: "Startups don't need to charge into the aircraft carrier's lane—the farther away, the better."
OK, so here's the question—the corner you hide in, how high can its ceiling reach?
Phil gave a technical-level answer. He said the essence of large models is outputting a "normal distribution"—you ask it to generate a "premium image," and what comes out is necessarily the greatest common denominator of internet aesthetics: mediocre, safe, not making mistakes but not standing out either.
"What kind of images truly go viral on Instagram and Pinterest? They have many characteristics in them. The boss doesn't need to tell me what a premium image is—you tell me the brand and website, and I'll automatically generate it for you."
His moat is: letting clients get good results without writing prompts. This sounds simple, but think deeper—every time a large model gets a bit smarter, the prompt threshold lowers a bit. How long is the shelf life of your "no prompt needed" advantage?
Liu Kun offered another approach. He doesn't compete with big players on technology—he competes on granularity. "In the past 48 hours on TikTok Southeast Asia, say Thailand or Indonesia, what kind of content has higher ROI? Under the same budget, how many KOLs and KOCs of what follower scale should you deploy? Go ask a big player, and they'll say they can do it too, but you might have to wait a month or two."
This is a "capillary-level" moat. Big players can build the blood vessels, but they won't manage what flows through every capillary.
Yin Tianming's moat is the most special, because it doesn't come from AI.
"We might be different from others making AI creation platforms. Others might just make a tool to sell under the AI wave. We're here because we understand the business—we understand game publishing, we understand what gives short dramas their thrill, we understand the rules of going global."
He used a harsh phrase: "Many people write recipes, but not everyone can cook."
What he means is: Senluo Zhihui is not "made an AI tool then sold it to the content industry," but "did content for 20 years, incidentally made an AI tool for our own use, then found others wanted to use it too." The direction is completely reversed.
Angel, listening nearby, added: "If the future goes the high-quality route—because eventually everyone can generate images and materials, but how to achieve high quality requires taste and very detailed workflows."
She recently helped a YouTube music platform make a virtual singer MV. Lip sync, character consistency, visual quality—all fine work. "It requires very long communication."
My judgment is: startups that only make tools have an extremely low ceiling, because every large model iteration is eating into your territory. But if you go deep into specific business workflows and embed into private data—your ceiling is no longer determined by the big players' firing range, but by how large the commercial pie is in your vertical industry.
Phil himself said a key sentence: "Your data isn't on Google, so Google can never train what you have."
Data and Know-how. Beyond the big players' firing range, only these two things can stop bullets.
Interrogation 3: How Long Is the Shelf Life of Doubled ROI?
Yin Tianming gave a set of data.
After using the AI creation platform, their marketing spend ROI improved by 30% to 50%. How exactly? "In the past, a good art original painter might produce two or three a day at best. Our team now produces over a hundred high-quality materials per day." Productivity exploded dozens of times, plus their self-developed sampling optimization algorithm—"others on the market might sample five times to get a satisfactory result; we average four times, with resource savings of at least 20%."
On Liu Kun's side, conversion rates under the same budget improved by 1.2 to 1.5 times or even higher. "Previously, a 1 million budget got 50 million impressions and 100,000 to 200,000 conversions"—now with the same money, the results jump significantly.
The numbers look good. But what's the reference point?
Yin Tianming's baseline is his own ad-spend performance from six months to a year ago—that is, comparing with "himself without AI." Liu Kun's baseline is pure manual or old systems under the same budget.
This raises a less comfortable question: when the entire industry standardizes on AI tools, will your "multiple improvement" still exist?
In 2025, the overseas micro-short drama market already exceeded $4 billion. AI comic dramas are projected to grow from $100 million in 2025 to $650 million in 2026—6x in one year. Production costs are 80% to 90% lower than traditional live-action filming. What does this mean? It means everyone is using AI. The money you save with AI, your competitors are also saving.
The multiple dividend Yin Tianming is enjoying is essentially a first-mover advantage dividend. He's been in gaming for 20 years, and AI tools are incidental; but when the competing company next door also starts using a similar AI platform, the 30% ROI improvement will gradually fall back to the industry average.
At that point, what's the competition? Yin Tianming himself gave the answer, though he may not have realized it was the answer.
He developed a tactic called "drama-game linkage." "When I launch a Journey to the West-themed game, I make a Journey to the West-themed AI short drama for linkage. On the platform, I'll cut the short drama into countless clips, which in turn provide user-acquisition materials for the game."
This was called "film-game linkage" in the traditional film and TV era—heard for ten years, with successful cases countable on one hand. Why? Because filming a drama was too expensive and couldn't align with game release cycles.
After AI drove costs down, this suddenly became executable. Three to four days for a premium AI short drama—want Journey to the West paired with Journey to the West, want xianxia paired with cultivation—it's all at your fingertips.
This is the structural change AI brings—not making a certain step 30% faster, but making things that were "thought of but couldn't afford" become "worth a try, no loss."
Dynamic Moat: There Is No Safe Corner, Only People Who Keep Running
Liu Kun said something very honest.
"For us startups, our moat is actually a dynamic moat. At each stage, we do things a little differently or a little more than the industry big brothers, using that slight lead to win clients and stickiness, then using those to accumulate the next moat, forming this flywheel."
He even said: "If I have the opportunity to stand here again next year, I might propose a new moat."
The subtext of this statement is: this year's moat may not work next year.
My strongest feeling after listening to this roundtable is—none of these four people feel safe. Angel at Google is thinking about how to let the ecosystem serve more startups; Liu Kun is betting on AI participation breaking from 95% to 100%; Yin Tianming is feeding an AI platform with 20 years of gaming intuition; Phil is calculating when his digital employees can be promoted from "intern" to "partner."
They share a clear awareness: on the AI going-global track, there's no possibility of digging a moat and then lying back to collect rent. The so-called moat is that you run half a step more than others—and then keep running tomorrow.
Phil's statement may be the most memorable criterion of the entire roundtable: "In the end, it still comes down to whether we have data and Know-how, and whether we're hiding outside the large models' firing range."
When the big players' firing range extends every six months, where do you plan to run?
More Conversation Details
Guests:
Google Cloud Business Development Manager — Angel Lam
Chengguo Shijie Founder — Liu Kun
Senluo Zhihui Founder — Yin Tianming
Canlah.AI Founder — Phil
Moderator: Unique Capital Partner — Wang Chaochao CC
Wang Chaochao CC: I'm CC, a partner at Unique Capital. Thank you all for coming to this roundtable. "Sensory Reconstruction"—this word sounds very grand, but when it lands in the commercial world, it's a very concrete thing. What does it mean? AI can not only write copy, it's also learning to understand images, listen to music, possibly read videos, and ultimately may take over the entire content production workflow. But there's an old question that has always existed: the technology is strong and cool, but how do you make money? Especially in a money-making city like Shenzhen, everyone cares deeply about making money and commercialization. OK, today we've invited four guests from different ecological niches to discuss together how multi-content production makes money overseas. Let's have our guests introduce themselves and their companies in order—Angel, let's start with you.
Angel Lam: Maybe I'll talk about Google's entire ecosystem. I think the best thing about Google is that it has different ecosystems. For example, Google Cloud is one ecosystem, Google Ads is another, and Google monetization is yet another. If a business builds its product on Google from the start, we have resources to help them build it. Then when they really want to go global, to Go to Market, they'll go to the advertising department, and we can provide different Insights—for example, if you're going to Japan, which product do people in Japan search for more? We can provide all these Insights. OK, we all want to make money, right? How to make money—we also have a monetization team. Telling everyone which ad placement is better for you, or providing opinions on IAP, IAA and other monetization methods. I think one of Google's best points is its one-stop ecosystem. For example, the gaming mentioned earlier—gaming is also many of our clients. From creating a game, how to use AI scenarios in it—the cloud side might handle that, and finally advertising to monetization, providing one-stop service. So different clients, such as gaming, e-commerce or AI startups, especially video platforms, we talk to clients with one-stop services, mainly helping them create products and go global. I believe everyone has also cooperated with Google in different scenarios.
Liu Kun: I think I might want to share two points. First, our business model has undergone changes over the past two years, and I'll also share what hasn't changed. Let me start with what hasn't changed. I remember first meeting CC about a year and a half ago. At that time, our company had just been founded for half a year, in the seed round stage, with only about five or six people. We had defined a strategic business model called "outcome-based pricing." At that time, this term hadn't caught on in Silicon Valley yet—it probably became hot there only last year. But our definition might not have been that clear—maybe called delivery, doing landing. This is completely different from selling shovels, selling tools, or the sexy business model investors understand as SaaS with decreasing marginal costs as it scales. So actually, when we explained this to investors, it was very resisted. But fortunately, we were able to get revenue from clients from a very early stage, and in the process continuously improved our product capabilities and scaling middle-platform capabilities. The hottest word recently is FTE. Essentially, in Chinese terms, isn't it just on-site staffing? It's just that during the on-site process, you distill customized client service into standardized SOPs, then use those to serve other clients. Back to how we make money—what hasn't changed is that we're providing outcome value for clients. For marketing, the biggest outcome is helping clients make money. You can use commission sharing, earning commission—we currently also have some revenue from this. More importantly, accompanying some promising brands to grow together will become a very important revenue line for us in the future. As for another model, actually in the minds of many brand-side marketing directors today, CPC (cost per click), CPM (cost per impression), CPS and so on have already formed a series of metrics. I think to date, no pure tool can fully achieve this, but we're still persisting in moving in this direction. Over the past two years, AI's participation in completing delivery has changed. Initially, AI participation might have been only 50%, with the remaining 50% relying on humans; but today we find AI participation has reached 95%, with maybe only 5% relying on humans. From the company vision perspective, we firmly believe that in half a year or a year, we can achieve 100% AI completion of a CPM or CPS delivery method—this is what's changing for us.
Wang Chaochao CC: Let me briefly summarize—through highly AI-concentrated standardized marketing, earning revenue through commission. Although Mr. Kun didn't say what the revenue scale is or what stage we're at, I can reveal one thing: when I met Mr. Kun a year and a half ago, his physique wasn't like this. At that time, you could directly associate him with special forces. Everyone can understand that commercialization degree and body weight are directly proportional for Mr. Kun.
Yin Tianming: We were originally in content, right? Our core business is game publishing and short drama production and distribution. The essence of entertainment content distribution is paid user acquisition. Last year we made hundreds of real dramas, and when AI took off in Q4, we started vigorously doing AI content. Whether real dramas or AIGC dramas, we've also made many hits. In April this year, we successfully integrated with the overseas TikTok short drama box mini-program, and next we'll do short drama going-global business. In integrating the entire chain end to end, we've completed ecosystem construction. Because in the process of doing content, to make better content, better distribution business and better ROI, we internally developed the Senluo Zhihui AI creation platform. With this tool, we now make a premium short drama in only about three to four days. We've integrated all mainstream domestic and international models, optimizing through our own technical algorithms. Everyone talks about how sampling wastes resources and slots—others on the market might sample five times to get a satisfactory result; we average four times, with resource savings of at least 20%. Our marketing ROI, after using the AI creation platform, improved by at least 30% to 50%. Our company's entire commercial revenue is currently still in game publishing and short drama distribution, but with the AI creation platform, overall ROI improved by 30% to 50%. Now we're also letting some friend companies in gaming or short drama use our tools, and basically everyone has varying degrees of ROI improvement empowerment. So I think in terms of the company's future positioning, we'll continue to focus on internal content creation plus AI-related ecosystem empowerment, continuously deepening.
Wang Chaochao CC: Because Mr. Tianming comes from gaming, and gaming is a heavily commercialized product. So the entire Senluo is doing heavily commercialized content, selling content—which forces you to produce very high-quality content, because if the content is ordinary it's very hard to compete now. It forces you to reduce costs and increase efficiency, needing handy tools and platforms, but the core is still to do content and sell content, right?
Yin Tianming: We might be different from others making AI creation platforms or SaaS services. Others might just make a tool to sell under the AI wave. We're here because we understand the business—we understand game publishing, we understand what gives short dramas their thrill, we understand the rules of going global, and we know what kind of value and tool embodiment is needed for business empowerment. So we're different from other tool providers—we're a team that relatively understands content in the industry.
Wang Chaochao CC: In other words, you've seen big money, you've seen large-scale user acquisition and distribution, large-scale money-making and monetization.
Phil: Here's where we are now. Before answering this question, I suddenly realized I have a lot in common with Mr. Kun. Not only do we do similar things, but I also served in the military from 2019 to 2021—I also have a military career. I also very much identify with this outcome-based pricing approach, but overall, outcome-based pricing is too demanding for us. Who produces your outcomes is actually unclear—it could be an Agent, AI, human plus tools, or human plus AIGC. We're more like "selling employees," selling ten digital employees. Your marketing department has a data analyst, strategist, operations person, content producer, Amazon e-commerce store optimizer, website builder. Each part can be understood as having a corresponding AI Agent—you can directly understand it as hiring an employee. The FTE concept right now means: I onboard this employee, how to make him understand the company's workflow logic, understand what products are being sold, what the brand positioning is. At the very beginning, we tell our Agents the company background framework, and they'll act like an employee. For example, right now on my phone, my OpenClaw (小龙虾)—every day I don't even need to check what it's doing; it just produces what it should produce. I check the website and see how many articles it posted for me on Reddit, how many followers it got, what it posted on Twitter and Instagram, KPIs completed—and I can pay for it. This is also why we want to introduce digital employees to bonded sellers.
Wang Chaochao CC: Here's the question—are you selling headcount, or selling outcomes and result sharing?
Phil: Right now, selling headcount. When we have enough confidence that every digital headcount is top-quality, then we'll sell outcomes.
Wang Chaochao CC: Selling your special forces digital employees.
Phil: You can understand it that way. When you first hire an employee, you pay a salary, right? But when they become exceptionally good and become my CTO, I can give them profit sharing. When they become even better and become a partner, I go and sell outcomes for them.
Wang Chaochao CC: You should have joined the previous roundtable about "digital employees."
Phil: I was arranged by Mr. Wu to be in this one.
Angel Lam: Actually, I also want to add something. Just now hearing everyone's sharing, I think Google has different cooperation methods in different scenarios. For example, for AI Agents we can provide technical support; if there are R&D problems we can support. Like Mr. Tianming, when you're doing distribution business, maybe distribution is on the Play Store, and we have a Play Store team that can help everyone. Like Mr. Liu, if you have needs in ROI data insights, our advertising team can also share some industry analysis. So in different scenarios, we're like partners growing together with different startups.
Wang Chaochao CC: Let me interject—actually, Unique Capital's conferences are somewhat public-welfare in nature. On one hand, we invite guests to share insights and Know-how with everyone; on the other hand, it also lets guests do some on-site large-scale product pitching, right? OK, I plan to combine the second and third questions. Now everyone has a relatively good understanding of what AI large models can do. Next, let's talk about how much better AI can do than humans. Everyone is making their own products—what is the core technical moat? Why do clients choose you and not others? What advantages does such a moat bring at the commercial level? I hope guests can use cases around them to talk about this stickiness, moat and payment issue. OK, still starting with Angel?
Angel Lam: Let me share an example. On the cloud, we usually divide into three layers. The bottom layer is TPU and GPU computing power, supporting different enterprises in product R&D and model testing. Then there's the platform layer, providing access to different models to train their own AI Agents. Finally, the application layer. I recently collaborated with a YouTube platform strong in music to help them make a virtual singer MV. Because making an MV has very high quality requirements, especially lip sync, character consistency, etc.—very high quality requirements. In the process, how our technology continuously R&Ds with the MV production platform to make high-quality MVs requires very long communication. What I want to say is: if the future goes the high-quality route, because eventually everyone can generate images and materials, but how to achieve high quality requires taste and very detailed workflows. We're here to help these different enterprises and platforms sit down and see how to make workflows more detailed and achieve higher quality.
Wang Chaochao CC: Mr. Kun, about the moat.
Liu Kun: Let me start from the moat. This is quite emotional—I think the word "moat" might only be useful for big companies like Google and ByteDance. While listening beside them, I kept thinking that whatever I say can't compare to theirs. Entrepreneurs probably all have similar feelings. We've been startups in the AI track for over three years. For us startups, our moat is actually a dynamic moat. At each stage, we do things a little differently or a little more than the industry big brothers, using that slight lead to win clients and stickiness, then using those to accumulate the next moat, forming this flywheel. In this process, in early 2023 when we were doing early open-source project construction like Stable Diffusion, ComfyUI, WebUI, we focused on some minor updates. For example, solving image consistency, character material consistency, frame-to-frame continuity and other detail issues. We achieved some small breakthroughs, attracting client and capital attention. When we moved toward product optimization and user experience, we found that original technical lead was quickly run over by big players, so we had to find new technical directions. Our current conclusion is: focus on doing things beyond the big players' firing range—things they're too lazy to do or don't deign to do. For example, in the beauty and FMCG track, in the past 48 hours or a week on TikTok Southeast Asia (like Thailand or Indonesia), what kind of content has higher ROI? Under the same budget, what tactics should be used, how many influencers, KOLs, KOCs of what follower scale should be deployed? This series of capabilities—go ask a big player, and they'll say they can do it too, but you might have to wait a month or two. So we've gradually accumulated some more vertical, closer-to-result moat capabilities, providing more direct value for clients. For example, in a large-scale Campaign, previously a 1 million budget got 50 million impressions and 100,000 to 200,000 conversions; under the same budget, we can achieve conversion rates 1.2 to 1.5 times or even higher than before. This is a dynamic process for a startup. If I have the opportunity to stand here again next year, I might propose a new moat.
Wang Chaochao CC: Just now I had an experience, let me summarize—it's about finding a model suitable for yourself in the process of verification and practice, verifying your mature methodology, then rapidly iterating, replicating and expanding capacity, using the leverage principle to unlock infinitely amplified returns.
Yin Tianming: For me, the core moat of our AI creation platform is not making a special image or cool video—it should be the deep combination of product productivity and AI Know-how. You need to know how to build this kind of ecosystem. Our three core veterans are basically all 20-year practitioners in the gaming industry. Our operations partner was formerly the distribution head at Tencent IEG, and our marketing user-acquisition head was an early employee at 37 Interactive Entertainment. We've always touched the market through content, understanding how to make short dramas and the rules of going global. This is our moat. Why do users choose us? The core is that we ourselves are people who make ROI, and when we sell to others, they also want to make money. Many people make tools, but few go deep into the market—just like many people write recipes, but not everyone can cook.
Wang Chaochao CC: That is, integrating 20 years of heavily commercialized content expertise into one platform. I've repeatedly verified that I want to make money—do you want to try?
Yin Tianming: So we've integrated our own ad-delivery system logic. We ourselves need a large amount of advertising materials every day, knowing how to use AIGC optimization to make more usable content. Our team itself is the first demand side, using the demand side's logic to iterate tools. On the ROI logic, repeatedly iterating our own content and tools is more valuable. In the past, a good art original painter might produce two or three a day at best. Our team now produces over a hundred high-quality materials per day. Let me give another simple example: I applied the old "film-game linkage" logic to "drama-game linkage." For example, when I launch a Journey to the West-themed game, I make a Journey to the West-themed AI short drama for linkage. On the platform, I'll cut the short drama into countless clips, which in turn provide user-acquisition materials for the game. Same for xianxia games—pair with cultivation-themed AI short dramas. AI saves a lot of live-action filming costs and production cycles, with very high cost-performance. I've enjoyed dividends in drama-game linkage, so I'm deepening the iteration of our tool system in this area.
Wang Chaochao CC: That is, linkage between multiple contents. AI has lowered the cost of things previously wanted to be done enough that now strong linkage can be achieved. The system Mr. Tianming's team makes is a product that has been repeatedly verified and continuously iterated, right?
Yin Tianming: There are similar systems on the market, but I think others may not optimize as well as us, because they only make tools and don't produce dramas or distribute short dramas themselves. In our optimization process, we've thought about two different experiences for professional users and novice users. Novice users can generate scripts, sample, storyboard and video through one sentence. After ChatGPT's latest release, it helped us optimize one step: at any time, you can modify any intermediate place at extremely low cost. We have algorithms to optimize, and with slight prompt modifications, we can achieve extremely cost-effective content output. For senior AI users, efficiency savings are even higher. Essentially, it's still about giving users a better experience.
Wang Chaochao CC: That is, integrating ROI into the capillaries—every step emphasizes ROI.
Yin Tianming: That's always how we've thought about it.
Phil: That was a lot of information just now. Let me summarize briefly. First, I agree that for startup teams, frankly, there's no moat at all.Just grab two or three people, take a few million in funding, and the probability is that tomorrow they can put together something similar to yours. So our moat is dynamic—it lies in how far we are from the large model vendors and from Google's "firing range." Google is very friendly right now, but every model iteration actually knocks down a whole bunch of startup teams, which means they're too close to the firing range.
Wang Chaochao CC: Can't blame the environment.
Phil: Exactly, can't blame the environment. Startup teams need to judge what large model iteration won't do. For example, my Image Model is precise to every detail—even facial hair can be adjusted—but what large models generate is always a Normal Distribution. Large models first need humans to input prompts. As long as I can achieve that you don't need prompts to generate a good image, that's a bit farther from the large models' firing range. If the boss gets involved, the probability is they'll only say "help me generate a premium image," but in a large model, "premium" is mediocre. What kind of images truly go viral on Instagram and Pinterest? They have many characteristics. So the boss doesn't need to tell me what a premium image is—you tell me the brand and website, and I'll automatically generate a recognized premium image for you. This is our moat as a startup hiding from the large models' firing range. Additionally, Google's moat is having massive data, and your (Liu Kun/Mr. Tianming) moat is the same—your data isn't on Google, so Google can never train what you have. So in the end, it still comes down to whether we have data and Know-how, and whether we're hiding outside the large models' firing range.
Wang Chaochao CC: Data and Know-how—whether it's on the cloud, whether it's been taken by others, whether it will be shared, right? Because 40 minutes is actually very short, and the discussion was quite vivid. Whether big, medium or small companies, everyone is in different fields and scales—no superiority or inferiority, each making their own money, right? Startups don't need to charge into the aircraft carrier's lane—the farther away, the better.