Manchi | 晚点聊 LateTalk Summary

1. Key Themes

Survival Over Profit: The Deliberate Low-Margin Strategy for AI Applications

Chen Mian (陈冕) argues that aggressive early-stage margins are a mistake for application-layer companies. She targets a gross margin ceiling of 30%, viewing high margins as demand suppression. Her logic: if you don't own the model, your most valuable asset is users.

"I don't think application-layer players in the early stage, especially within the first year of launch, should have high gross margins. Storage has 80-90% margins. Chips have 80-90% margins. SOTA models also have 70-80% margins. I don't think applications should. So I've always set our target: don't have very high gross margins in the early stage." [00:08:08]

"Because you're chasing high gross margins, you suppress demand. But as an application-layer player, you don't control the model. Isn't your greatest asset users? Isn't your greatest pursuit to have more users? Otherwise, what's left of the application layer?" [00:09:07]

The Gym Membership Business Model for AI Tools

The entire pricing architecture of LibTV and Lavart is based on a gym membership assumption: users buy credits they will never fully consume. The insight is not novel — it describes every SaaS and tool subscription — but Chen Mian is unusually transparent about the mechanics.

"It's essentially a gym membership model. It bets that after you buy an annual membership, these people won't come every month or every day. But this needs to correct a logical point — this is how tool software's business model has always worked. In the previous era it was the same, except there was no token. But this era added something: you can neither completely abandon subscription pricing to base purely on token usage, nor fully embrace subscription, so people feel this contradiction." [00:19:15]

"If you price assuming every user will consume all their credits, the price users pay would be extremely high. If you look at it, in the previous era a user with 365 days of an active product year was only active about 20-30 days. So you can understand that within a year, users are generally active for four or five months, and within those months, active maybe five or six days each." [00:18:17]

Timing Is the Supreme Competitive Variable

Chen Mian identifies timing as the single most important lever available to a startup, and explicitly prioritizes speed over quality in certain windows. Her willingness to ship an imperfect product fast is a conscious strategic choice, not recklessness.

"You need to understand one thing: when you have nothing, using time to exchange for space is your strongest weapon." [00:01:16:50]

"In this era, all engineering innovations have a shelf life of about one week. A few days, one week. Model competition advantages disappear within a month. How long did you expect your engineering innovation to survive?" [00:42:17]

"I realized: whatever advantage you build in one month of aggressive launch determines how much constitutional advantage you carry forward." [00:01:17:50]

Saturation Attack as the Core Go-to-Market Tactic

For LibTV's launch, Evoken deployed what Chen Mian calls "saturated attack" marketing — flooding all possible channels in a compressed window. The actual spend ($1M over one month) is remarkably modest relative to the revenue outcome; the ROI came from differentiated content, not sheer volume of paid ads.

"You can understand that whether it was LibTV or our previous products, all went through a period of saturated attack. LibTV's window was about one month. How much did we spend? $1 million? $1 million isn't even much money. Given the revenue outcome at this scale, I called it an extremely high-ROI strategy." [00:43:46]

"The core logic of high ROI is that you need to have content. Because you're not doing paid ad placements. So the essence of high ROI is you need differentiated content. I believe in this era, if you cannot provide a new experience based on models, or provide a new model, you have no meaning." [00:44:43]

The PMF Discovery Gap: Finding PMF ≠ Winning the Market

Chen Mian makes a pointed and controversial argument — that being first to PMF confers far less durable advantage than the startup community believes. This is used to justify pursuing markets others have already validated.

"We must understand one thing: being the first to discover PMF, in the history of human technological competition, being first to discover PMF and ultimately winning the market are worlds apart." [00:37:56]

"The commercial world also doesn't let you own a market just because you were first to PMF. I'm not a giant saying 'I can do this or not, I'll crush innovation.' That's not what this is." [00:38:52]

Application Companies Must Build Two-Sided Networks to Escape Gravity

Chen Mian's long-term survival thesis hinges on building network effects through user-generated value — not just accumulating users. Without a two-sided network, application companies remain permanently subject to "gravity" (i.e., model companies pulling value upward).

"I believe that user scale is the only way you escape gravitational pull — and today that gravitational pull is the model companies. And it may not only be user scale. I think application companies need to try to leverage users to create wealth... The core is: can your users create irreplaceable value for you? This is essentially a two-sided network." [00:01:36:21]

"Liblib is a two-sided network. LibTV has not yet clearly formed value creation through users to retain users." [00:01:36:49]

The "Unify Jiangdong" Strategy: Surviving in a Giant's Blind Spot

Chen Mian's most elaborate strategic framing is a historical analogy to the Three Kingdoms period — occupy a defensible region while the dominant powers fight for the central plains, use the time to build your own "Yangtze River" defensive moat.

"When he's fighting for the central plains, you can be unifying Jiangdong — it has nothing to do with him conquering the central plains. He also knows Jiangdong is being unified; he may send some troops, but whether you can hold on is your own fate. His core will always be conquering the central plains." [00:01:32:25]

"How do we use time to exchange for space, then use space to exchange for resources, then use resources to build our moat? This is the most important strategic question every startup must think about if it wants to survive independently." [00:01:33:23]

The Judgment Scarcity Thesis: Ideas Are Now Cheap, Discernment Is Rare

As vibe coding and AI tool access democratize execution, Chen Mian argues the scarce resource shifts from ideas to judgment — knowing which idea is the right one at the right moment.

"In this era, idea is cheap, because vibe coding makes implementation of any idea extremely easy. So in the sea of ideas, finding the correct one is the greatest value of a human. Because judgment has become scarcer — this is what this era has brought us." [00:01:22:13]

"How do you distinguish idea from judgment? Sometimes they're the same thing. A very imaginative person has many ideas every day. But which idea is the best choice at which timing — that is judgment. This is the most important lever for a startup." [00:01:23:13]

Recognition Over Management: The Startup That Ignored Org-Building

Chen Mian openly admits that management was the most systematically neglected function across Evoken's three years, framing it as a deliberate trade-off rather than an oversight — though she now acknowledges the costs in talent retention and team health.

"I believe management is the relatively least important thing in this era. But that doesn't mean it's not needed — because you need people to stay. In our past few years we neglected management. This is also my recent reflection." [00:51:25]

"The only purpose of management is to allow AI-native, high talent density people who fit your business characteristics and requirements to remain on your platform. That is management's only meaning in this era." [00:49:32]

2. Contrarian Perspectives

"First to PMF" Is Overrated and Possibly a Disadvantage

The startup world treats PMF discovery as a defensible moat. Chen Mian flatly disagrees — and built a $2B company by deliberately entering markets others had already validated.

"We must understand one thing: being the first to discover PMF, in the history of human technological competition, being first to discover PMF and ultimately winning the market are worlds apart." [00:37:56]

She argues those who established PMF can be overrun if they don't aggressively capture incremental growth at the moment of market expansion — and that aggressive fast-followers who correctly read the market size have a legitimate claim to winning.

Application Companies Should Actively Maintain Low Gross Margins

Conventional wisdom pushes SaaS companies to maximize gross margins. Chen Mian argues this is wrong for early-stage AI applications — high margins suppress demand, and demand is the only asset that matters when you don't control the model.

"I don't think application-layer players in the early stage should have high gross margins. I've always targeted: don't exceed 30%. Because chasing high gross margins suppresses demand. But as an application-layer player, you don't control the model. Your greatest asset is users." [00:08:08]

She adds that the "gym membership" model — where most users never fully consume their purchased credits — is simply how all tool software has always worked, and is not a sign of financial manipulation.

Interface Innovation Is Not Real Innovation and Should Not Be Valued As Such

The industry rewards "first UI" with prestige. Chen Mian says the optimal UI for a given era tends to converge rapidly and that copying interface patterns is not copying innovation.

"I don't believe interface is the most important form of innovation. Interface is honestly hard to have absolutely zero-to-one original innovation, because you're always evolving and borrowing. Short video was always scroll-up before TikTok did it. All phones now look like this." [00:36:26]

Innovation Companies Do Not Need to Always Innovate

This is directly counter to how innovative startups brand themselves. Chen Mian argues that innovation has costs, and that executing on validated ideas is as legitimate a strategy as discovering new ones.

"Everyone has a misconception: innovative companies must constantly innovate. Please — innovation has costs and consequences. If you've innovated something but can't hold on to it, can't build it well, what's the point of innovating it?" [00:58:42]

In the AI Era, Management Is the Least Important Function

This will be deeply unpopular with organizational theorists and most investors, but Chen Mian argues that in a fast-moving AI application environment, the cost of management overhead exceeds its benefit in the early stage.

"I can say that in this era, management has become unprecedentedly less important. Management's only purpose is to allow the right people to stay on your platform. But we've neglected management over the past few years. This is my recent reflection." [00:49:33]

She nuances this by separating the talent-identification function (which she says Evoken does well) from the talent-retention and emotional care function (which she admits they've done poorly).

3. Companies Identified

Evoken (言语科技) / Liblib

Parent company behind Liblib, Lavart (Love Art), and LibTV. Three AI-native creative tools. Raised $300M at a $2B valuation — the highest valuation of any Chinese AI application company at the time of the interview. Cash flow positive since May. LibTV now contributes more than half of total revenue (implying >$150M ARR from LibTV alone).

"Starting from May, our cash flow has been positive. So you can understand: our first magical experience was that May and June both showed positive cash flow, and the company's account balance was actually growing — while people outside were saying we had exploded." [00:05:14]

Lavart (Love Art)

Evoken's AI image generation product, described as a globally unique creative tool oriented toward professional designers. First major monetized product; described as the company's near-death and survival story.

"Lavart is essentially an innovation wrung out in extreme pain. It is a precedent." [00:01:03:33]

LibTV

AI video creation canvas product with node-based interface; launched in early 2025. Described as the fastest-growing product, reaching $1M daily revenue peak within one month of launch. Currently the majority revenue contributor for Evoken.

"Within one month of launch, the peak hit $1 million in revenue. That's real." [00:01:26:37]

Liblib

AI model sharing and designer community — described as a "one-of-a-kind product" globally: a model asset-sharing community oriented specifically to professional designers. Still growing and cash flow positive.

"Liblib is a unique product globally. You won't find a second designer model asset sharing community. That time, people said we looked like a foreign website, but that foreign website was very unhealthy. So Liblib's designer community is a one-of-a-kind." [00:01:03:33]

Adobe

Repeatedly referenced as the proof point that high-value production tools capture the most revenue with the fewest users — the model Evoken consciously chose to emulate.

"The globally most profitable creative tool product in the previous era was Adobe. But Adobe had the fewest users. Something wrong — why does it have the fewest users but earn the most? Because in the productivity track, it's not whoever has the most users who wins, but whoever controls the highest-value production." [00:26:34]

Manus

Agent-based AI product; described as having launched before LibTV's agent capabilities, which accelerated Chen Mian's urgency to ship. Credited with establishing the "agent" paradigm faster due to its stronger technical foundation-first orientation.

"Manus had already launched. After Manus launched, I realized the agentic concept would quickly become consensus. And after Manus launched, one very significant impact on us was: it took the tech brand crown that startup companies fight over." [00:01:13:51]

Tusi (土司)

Run by Shen Zhenyu; the main early competitor to Liblib in the AI image community space. Better capitalized (latest round of $70-80M vs. Evoken's $4M angel). Ultimately lost the competitive battle; Tusi focused on entertainment/casual content while Evoken focused on professional design.

"At that time Tusi. Shen Zhenyu made it; he's already moved to his next project, no issue. But they really gave us — at that time they had already raised $70-80M in their latest round, while we were still a $4M angel round company, and they were also very aggressive and capable." [00:25:08] "When I looked at the tool and saw their content was mostly entertainment-oriented while almost all design content was on our side, I already felt game is over." [00:27:04]

Caidu (Coffy / ComfyUI)

Referenced as the originator of node-based UI in AI image generation, making the point that LibTV's node-based interface was not copying any single competitor but following an established paradigm.

"In the multi-modal field, ComfyUI was the earliest node-based UI product." [00:35:55]

Flair (Flora)

Mentioned as the earliest video product to adopt node-based canvas in the video domain, predating LibTV's approach.

"In Silicon Valley there should be a product called Flair — the earliest node-based product in the video domain I saw was theirs." [00:36:05]

OpenAI

GPT Image generation (Image One / GPT-4o image capabilities) cited as the product that made Evoken's planned intermediate workflow product obsolete overnight, triggering Chen Mian's crisis moment and pivot to agent-based LibTV.

"After Image One came out, I realized that the node-based stronger-control workflow solution we'd been exploring for a year was no longer needed. The workflow had already been internalized by the model." [00:01:08:55]

Meitu (美图)

Referenced as a company Chen Mian worked at, where she observed firsthand that high-value productive tools (vs. high-volume consumer tools) are more monetizable — experience that shaped Evoken's product philosophy.

"I worked at Meitu. After working at Meitu, you discover something: in the previous era globally, the most profitable creative tool was Adobe. But Adobe had the fewest users." [00:26:26]

Mobike (摩拜)

Chen Mian worked at Mobike and initially believed shared bikes were a business with strong forward momentum — later found the assumption was incorrect. One of her early lessons in observing vs. actually doing.

"When I worked at Mobike, I thought shared bikes were a business with great forward momentum. Later I found the answer is not." [00:29:32]

Meituan (每日优先 / 美团 ecosystem)

Referenced indirectly via Meiriyouyi (每日优鲜), another company where Chen Mian worked, observing retail strategy and later concluding she had initially misread the dynamics.

ByteDance / TikTok (字节跳动 / 抖音)

Chen Mian worked at ByteDance and observed TikTok's rise; credits this with giving her competitive combat instincts (knowing you must fight or die) but explicitly says this gave her "combat memory" not "business muscle memory."

"I was at ByteDance, then observed TikTok — it's not the first short video, so why did TikTok ultimately win? This kind of thing." [00:30:29]

Sea Dance (即梦 / Jimeng, ByteDance's video model)

The underlying model for LibTV. Chen Mian confirms they are among the largest customers of this model (ByteDance's Jimeng), though she pushes back on the "biggest ByteDance reseller" framing.

"We are of course leveraged on the model as an application layer — we are inherently selling the model. Because your value is built on top of the model." [00:13:54]

Kimi (Moonshot AI)

Referenced briefly in context of Xiaohong's (Manus founder's) background, implying comparison between product-led vs. research-led founding approaches in Chinese AI startups.

Jasper Park (JasperAI)

One of the "wrapper three brothers" — Chen Mian, Manus, and Jasper Park's founder were once grouped together. She notes Jasper Park chose general-purpose AI while she chose vertical.

"I think we have some fundamental differences. For example, they chose General while I chose vertical." [00:01:44:10]

4. People Identified

Chen Mian (陈冕)

Founder and CEO of Evoken (言语科技), parent company of Liblib, Lavart, and LibTV. Former employee of Meitu, Mobike, Meiriyouyi, and ByteDance. Self-described product manager by identity, not a technologist. Built China's highest-valued AI application company ($2B) from a $4M angel round, with the company account reaching as low as 4,000 RMB before a wire arrived. Known for extreme urgency, high competitive aggression, and deeply personal product intuition.

"Anxiety, speed, and aggressiveness are Evoken's survival strategies over the past three years. 'Survive' is the word she says most in this interview." [00:02:21] "I am a very anxious founder. My soul is that of a product manager. Being called a great businessman totally confused me." [00:58:13]

Shen Zhenyu (沈振宇)

Founder of Tusi (土司), Evoken's early key competitor in the AI image community space. Already moved on to his next project at time of interview. Credited with building a well-funded, aggressive competitor, but ultimately chose the wrong market positioning (entertainment vs. professional design).

"Tusi. Shen Zhenyu made it; he's already moved to his next project, no issue. But they really gave us — at that time they had already raised $70-80M in their latest round, while we were still a $4M angel round company." [00:25:08]

Xiaohong (小红, Manus co-founder/CEO)

Described as a personal friend of Chen Mian; also a product manager background. Credited as an example of a founder who was "technology-first, then matched to users" rather than "user-first, then found technology" — which is why Manus was faster to ship agentic capabilities.

"Manus is technology-out, matching to users. I am user-in, finding cutting-edge technology. They will naturally be faster than me. So they are better suited for general-purpose." [00:01:14:20]

Tang Jie (唐杰, Professor)

Referenced approvingly as someone who articulated a nuanced version of Chen Mian's own management philosophy: management is less important, but organization and culture still matter.

"A few days ago Professor Tang Jie and I exchanged messages. He had a hidden condition: management is less important, but organization and culture are very important." [00:49:32]

Zhang Yiming (张一鸣)

Referenced regarding his famous quote about not watching competitors when driving. Chen Mian partially disagrees — arguing you must watch competitors on tactical matters (survival) even if you ignore them on strategic direction.

"Wasn't Zhang Yiming's famous saying 'don't stare at your competitor while driving'? I think this is incorrect — or rather, you need to look at this in two phases." [00:24:38]

Luo Yonghao (罗永浩, "Lao Luo")

Referenced as someone Chen Mian gave a public interview to, and felt uncertain about whether she deserved the platform. Cited as a moment of self-questioning about whether her success is real.

"When I went to do the interview with Luo Yonghao, I looked at the previous guests and asked myself: do I belong here?" [00:02:04:00]

5. Operating Insights

The Differentiated Content Requirement for High-ROI Marketing

The key lesson from LibTV's launch is that "saturation attack" marketing only generates extreme ROI if the content being amplified is genuinely differentiated. Without differentiated content, you're just buying paid traffic. Chen Mian spent $1M in one month and went from zero to $1M/day revenue peak — specifically because the content (first human+AI simultaneous video canvas) was novel at that precise moment.

"The core logic of high ROI is: you need to have content. Because you're not doing paid ad placements. So the essence of high ROI is you need differentiated content. I believe in this era if you cannot provide a new model-based experience or a new model, you have no meaning." [00:44:43]

The operating implication: identify the narrow timing window when your product has a genuine differentiator, then compress your entire marketing spend into that window rather than spreading it. The window may be only days to weeks.

The "Assumed Consumption Rate" Pricing Framework

Chen Mian describes a sophisticated pricing architecture that treats subscription credits as a function of predicted consumption rate × predicted renewal rate. Rather than pricing based on full utilization, price based on statistically expected utilization.

"This pricing is fundamentally a balance between your assumptions about renewal rate and assumptions about user consumption rate. You choose an appropriate consumption rate and renewal rate based on experience and set a value. The reason it looks aggressive is that 3.9x is an outcome — not an input." [00:11:59]

The practical playbook: (1) calculate your per-token cost at expected model quality, (2) model realistic user consumption patterns (LT30 ~5 days active per month), (3) set credits generously assuming most users consume 20-30% of them, (4) price the package based on full consumption but capture margin on idle credits.

Build Attention Before Ability — But Have the Ability Ready

Chen Mian's sequencing is: ability (ship product) → timing window opens → saturated attention attack. The error most competitors made was having ability (product) but not building attention at the moment of market explosion, thereby losing first-mover scale advantage even if they had first-mover PMF.

"Building attention, for an application company, is itself a very important thing. I don't believe that once I find PMF, the game is over. Even if you happen to hit a PMF, you should be very vigilant. Because if the market's incremental growth will explode at a certain moment, and you don't capture enough of that incremental growth during the explosion, your existing base means nothing." [00:45:41]

6. Overlooked Insights

DeepSeek's Impact on Revenue Was the Catalyst for Evoken's Pivotal Fundraising — and Chen Mian Was Simultaneously Panicking About Existential Obsolescence

This was mentioned extremely briefly but is hugely significant. At the exact moment Evoken was trying to close its funding round, DeepSeek's release caused a massive spike in Chinese AI adoption — doubling Evoken's revenue and making the fundraise look easy. But Chen Mian was simultaneously in private despair because GPT Image capabilities had just made her planned intermediate product (Xinliu/心流 workflow) obsolete. She was presenting a bullish revenue story to investors while internally convinced the business model had broken.

"At that time it was bittersweet. DeepSeek drove a massive increase in China's AI adoption rate, so our revenue doubled. The team was very happy. But I knew: it's over. If a shareholder hadn't discovered that other thing, I'd just say revenue grew. If they discovered it — telling them the truth wouldn't help anyway. I said: complete disaster." [00:01:15:21]

The implication for investors: revenue momentum at a single point in time can be entirely coincidental and temporally disconnected from the strategic health of the company. Chen Mian used the DeepSeek revenue tailwind to close a round while pivoting away from the product that drove it. The round was justified not by that product's future but by the bet on LibTV — which the investor almost certainly did not fully understand at close.

The Actual Moat Chen Mian Is Building Is a "Human Creativity Network" Bet Against AI Centralization — and This Is a Deliberate Philosophical Counterposition to "Accelerationist" AI

This was stated only briefly but is the most consequential long-term thesis in the entire interview. Chen Mian explicitly says her company's survival thesis depends on humans retaining creative value for the next 10 years. She is consciously building on the opposite side of the AI-replaces-humans bet. This means her product architecture, talent choices, and network effect strategy are all downstream of a single philosophical wager.

"The last thing I bet on is whether humans have value. I'm not betting on whether humans have value forever, because ASI may happen. I'm betting that within ten years, humans have value. So within ten years, if humans have value, can we be together with users, using humanity's final glory to build an application that has that Yangtze River defensive advantage?" [00:01:39:47]

"For AI to create wealth, you either replace human production — or you empower human production. These are two different AI eras. Everyone is now noisily pursuing AI that replaces human production to increase efficiency. But our question is: if all human production efficiency has been improved by AI, what are humans doing with their time? Why can't we use AI to empower humans to create more?" [00:01:40:45]

The investor implication: Evoken is not simply an AI tool company. It is a calculated bet that the consumer market will bifurcate between AI-generated commodity content and human-AI collaborative premium creative content — and that the latter cohort will form a defensible, monetizable community. If this bet is wrong (ASI accelerates faster than expected), the entire moat thesis collapses simultaneously across all three products.