1.1 From Chat Windows to Action: Why We Must Redesign AI Products
Today, I want to share how we view building products in the AI era.
My own experience is relatively rich. I am the oldest member of our management team; my other partners were born in 1992 and 1993, while I was born in 1986 and will turn 40 next month. Seeing a nearly 40-year-old person among students might make you think this is a typical middle-aged person.
I entered this industry early, starting work right after my undergraduate degree without pursuing graduate studies.
Before joining the team, I always wanted to build a product that could run continuously 7×24 hours: one that doesn't require constant human input but can continuously reason and produce output. I hadn't even figured out the specific form at the time, but I really wanted to do it because I believed the upper limit of model intelligence had already surpassed humans. We needed a reasonable product form to proactively and automatically stimulate the model, letting it continuously generate output. Regardless of the final application scenario, I believe this is the ultimate form an AI product should have. Many thoughts came from exploring the current boundaries of model capabilities.
This was also an important reason for me deciding to join. At the time, Red approached me saying he wanted to build an AI browser, and intuitively I felt this direction was right.
Later, I summarized the core problem encountered in our research into one point: The Chatbot product form itself limits the capability expression of AI models.
This means you built a product, but the upper limit of its value is constrained by users: Users must know how to ask questions AND be willing to ask them to unlock the product's value. What if users don't know how to ask? Then the product's value becomes minimal.
Browsers are a typical scenario: You don't need to continuously input prompts; writing emails, searching, browsing web pages—these behaviors themselves are expressions of intent and can become model inputs, generating massive reasoning needs in the background. Another reason our Chief Scientist and I were willing to join was that Red is not only a very impressive product manager but also deeply understands company operations.
1.2 Stopping the AI Browser: Time Matters More Than Sunk Costs
By October 2024, our AI browser was actually ready. Internal testing always revealed one or two particularly bright features.
But when recommending it to family and friends, serious problems emerged: Everyone is too familiar with browsers. The first question they asked during trials was often, "Where did that feature I used in Safari or Chrome go?"
This is the biggest issue when grafting new experiences onto old ones: Users will always benchmark against their old experience baseline. But that isn't where your highlights lie. You want them to focus on the AI part, but everyone cares most about how well the basic browser experience is done.
As a startup team, aligning basic experiences with giants like Apple and Google on a product with over 20 years of mature benchmarks is nearly impossible. Forcing it would consume excessive extra energy, causing us to neglect AI development. At the time, I thought the product had real highlights and hardcore users would love it upon release, potentially catching fire after a few iterations. But Red was extremely conflicted in September and October. On one hand, he felt logically unsound to hang new experiences on old ones; on the other, he worried that once released, the entire team might get stuck in it for 3 to 6 months. Founders often hesitate to give up; if the first version performs poorly, they think "maybe it's just not good enough yet, let's iterate," wasting half a year.
In school, you might not grasp the concept of time, but for startups, money is actually the easiest resource to solve—as long as the direction is right, you can always raise funds. The most precious and non-wasteful resource is time. Wasting time doesn't just drain personal energy; more terrifyingly, it causes you to miss phased market windows and trends.
At that time, in October 2024, our entire company had only 30 people. With ZhenFund's support, we had only raised Angel and Pre-A rounds, with little cash on hand. After investing 7 months in a new project that was already polished and likely to stun the market upon release, Red ultimately decided to kill the project. Looking back now, that was a crucial decision, because if we hadn't stopped the browser, there would have been no Manus later.
When we first shared this story last year, critics said we shouldn't disparage the AI browser track since we didn't build one ourselves. But today we can confidently say the AI browser track indeed has no truly successful products. The essential reason is: To create new experiences, best to create entirely new ones rather than trying to do "AI + Old Experience."
1.3 Seeing General Agent Opportunities in Coding Tools
After stopping the browser in October 2024, the team began thinking about what the next product should be. The Beijing team was very small then, occupying just one small office in a co-working space with five people total. As CPO, I handled product; the Chief Scientist handled models; and a full-stack developer colleague named Abin. Although our roles differed, all three of us had coding backgrounds. I started writing code in third grade.
The most phenomenal product of 2024 was Cursor, which exploded in popularity in July. We used it daily but developed different feelings during use. Especially as a non-professional programmer, I started using Cursor for many other things: handling reimbursements, converting MP4 to MP3 for my wife, assisting with novel writing, etc.
In this process, we increasingly felt the power of the Claude 3.5 Sonnet model behind Cursor. Especially after subsequent versions, both Cursor and Windsurf saw significant capability improvements. Many domestic and international startups were imitating Cursor, leading to a surge in coding tools, though they weren't quite called "Coding Agents" yet.
The true opportunity we saw was this: Sonnet 3.5 v2 unlocked far more than simple coding ability—it unlocked autonomous planning and action capabilities.
Autonomous planning and action capabilities mainly manifest in two aspects:
First, Long-horizon Planning.
Now this might seem common sense, but it wasn't at the end of 2024. At the time, OpenAI's models struggled with this, preferring ChatGPT-style turn-by-turn dialogue, solving problems directly within one interaction. Starting with Claude 3.5 Sonnet and greatly enhanced in later versions, the ability to perform long-horizon task planning emerged: It knows not to rush completion in one turn, but actively calls tools, asks users questions, gathers more information, and executes only when sufficiently confident.
Second, Tool Use.
Cursor builds tools enabling more people to become engineers. But our judgment was that the world doesn't need more engineers, nor do engineers need more fancy tools—they already have enough. If any of you code regularly, I believe your computer has many tools installed, constantly swapped and evaluated. That's how engineers operate.
We believed that once this capability was unlocked, it effectively unlocked general abilities for ordinary people. While the whole world bet on Coding Agents, we saw the opportunity to use coding as a medium to solve general scenarios for ordinary users, so our entry perspective was completely different.
When we launched the project in December 2024, I was actually very panicked because I felt surely others saw this future too. We were tense, rallying all available company resources to develop quickly, fearing someone else would beat us. After two months of development, by the March 5, 2025 launch, we discovered that in our observation, almost only we were working in this direction; others were still building coding tools.
We spotted the inflection point at the end of 2024, keenly identifying long-horizon planning and tool-use capabilities at the model's boundary. After multiple explorations, we kept trying to find the right product form.
1.4 Targeting Non-Technical Users: Making Intelligence Visible, Building Trust
Early Manus looked very much like Cursor, resembling a coding tool. Once, ZhenFund investor Yusen visited our office to see the early prototype. He said, "Does this three-column layout look too professional? It feels off-putting." Our judgment when planning the product was to target non-technical audiences. Since targeting non-technical people, having a first impression that looks too professional is definitely problematic. So we changed the interface from three columns to two.
During this process, we made many innovations for the target market. Because Agents were too new, ordinary people found it hard to believe they could actually deliver. During those two and a half months of R&D, my voice was hoarse every day when I got home.
After getting the first prototype running on January 5, 2025, I posted on Moments asking friends: If you had an intern with bachelor's-level education, what tasks would you assign them? I collected over 400 tasks from 100+ friends, forming Manus's earliest evaluation dataset.
From the first prototype onward, as the Agent execution framework (later called Harness) improved, those tasks brought daily surprises. Excluding 100+ tasks requiring embodied intelligence (like buying milk tea), the remaining 300+ tasks were completed in ways beyond imagination. I couldn't help shouting in the office every day, coming home with a hoarse throat. My wife asked why work made me so exhausted. But the feeling of intelligence bursting through the screen was irresistible—you knew this was the future.
However, this impact was so intense that if we just gave results to users, they might think it was fake. So we made a judgment: Transparency itself must become a product feature.
In Manus's first release, two critical product designs became design paradigms for almost all AI Agents:
First, Step-by-step execution display. Letting users see how the Agent progresses tasks, calls tools, visits web pages, writes notes, and writes code.
Second, Session Replay. When sharing sessions, others can review the complete execution process.
Product intelligence performance might account for 80%, but these two growth-oriented UI designs were crucial for initial explosion. Many initially didn't believe in AI Agents, but seeing step-by-step how it progressed tasks, called tools, visited pages, wrote notes/code, they realized it worked like a human, building trust. With Session Replay, users didn't need invite codes or payments; just viewing shared sessions allowed them to feel the surprising moments. These designs stemmed from clearly defining our target users as non-technical: We couldn't assume technical background, so we had to use product innovation to help them understand what was happening and build trust.
1.5 The Real Reason Behind One Launch: Product Innovation and Founder Expression
Regarding the March launch, last November at Tsinghua University, a student asked if many thought we bought lots of KOLs and staged a marketing event in March.
A friend in marketing told me, "Even with a budget of 20-30 million RMB, if you asked me to replicate Manus's marketing, I couldn't achieve this effect." Young people often stigmatize marketing, thinking it's wrong. If you want to start a business but think marketing is wrong, that's terrible cognition. How to tell a good story is precisely a skill Chinese founders lack, something I experienced deeply in Silicon Valley over the past two years.
People underestimate marketing difficulty, assuming it's just throwing money. If it were that simple, startups would never have chances because you can't outspend big companies.
I used to say the March launch cost 0 RMB, but later realized this wasn't rigorous:
On February 28, 2025, the website was ready. CEO Red reviewed it and said, "Tao, isn't this site a bit plain? Should we add a video?" Only days remained before the March 5 launch. I contacted two or three video production companies in Beijing; all said minimum two weeks lead time.
Finally, I shot the video myself using a Sony A7C II camera. Due to poor lenses, I borrowed a 50mm f/1.2 GM lens from ZhenFund's Yusen. I edited for two nights. When exporting to YouTube, I discovered 4K export required payment. So I spent about $12-15 for a one-month CapCut membership—the only extra money spent on that launch, canceled immediately after renewal.
After the video went viral, people assumed we were great marketers who made an amazing video. I repeatedly tell this story to warn young entrepreneurs: If your takeaway from a phenomenal product is "they're good at marketing," you learned the wrong lesson. Things are never that simple.
First, Manus appeared at the right time, using the right product form to show the world where model capability boundaries lay. Now frameworks are called Harness; back then, the term wasn't popular. I believe we were among the earliest teams doing such frameworks, pushing model boundaries to the limit. After launch, I met Anthropic's team in San Francisco in March. They didn't believe we only used their models because their internal tests didn't match our strength. I said I genuinely loved their models. Later, Claude Code leaked some code, and we saw echoes of our early architecture designs. Thus, I believe the core foundation of launch success was significant product innovation.
Without good product innovation, relying solely on videos is putting the cart before the horse. Later, I felt helpless seeing countless startups adopting the same style: founder sitting on a sofa with plants behind. Everyone acted like East Asian kids copying exam templates without knowing what to do otherwise. This is funny because we chose that angle purely because it was the only clean background spot in our co-working park, quiet enough for filming. If you watch closely, there's a frame jump because our Chief Scientist spoke while a cleaning lady mopped behind him. What truly matters is never the video itself, but your product. Yet looking back, that video remained important because I shot it and our Chief Scientist presented it.
After Manus blew up, many teams faced boss/investor questioning: "Why can't you build Manus?" The safest excuse often became: "We're a solid tech team, unlike them who are good at marketing." This sounds grounded, blaming failure on "not being good at marketing."
People ignore the essence: I'm a product person; the other handles models as Chief Scientist. Neither of us came from marketing. We simply built a product we were proud of and satisfied with, then expressed it to the world in our own words.
Early-stage startup marketing must involve founders personally. Outsourcing to agencies or hiring "marketing experts" risks severe information distortion. Only founders truly know how to tell the story. So looking back, the video mattered not for its format, but because we did it ourselves.
1.6 Designing Products for Model Capabilities Six Months Ahead
For AI application innovation, design products for model capabilities six months ahead. I believe next year, 2027, will be the big year for AI applications.
What does "six months ahead" mean? Don't wait until you read an article this morning about a model update adding capabilities, then think about building apps based on it. By then, it might be too late.
In our team, the Chief Scientist provided crucial input. Having researched NLP for over a decade, he could keenly perceive frontier lab evolution directions from current model performance and public/non-public info. For example, we saw AGI (Artificial General Intelligence) capabilities emerging from Claude 3.5 Sonnet, stepping up again in V2. We judged they were investing heavily in this direction. If it stepped up another level in 3-4 months, wouldn't our product be incredibly powerful?
So we decided to start immediately. Originally expecting a jump from 3.5 to 4.0, planning to launch with 4.0, but 3.7 came out first. The night 3.7 launched, I posted on Jike: "Night of faith top-up." Why? That night, our Harness code didn't change a single line; we just switched the model interface from 3.5 V2 to 3.7. On our eval set, performance surged, solving roughly 20-30% of previously unresolved issues instantly. We knew we'd bet correctly.
Our Chief Scientist always emphasized: "Less structure, more intelligence." After launch, when I shared this in Silicon Valley and domestically, many dismissed it. Common practice then involved humans designing workflows, teaching AI steps: writing mandatory rules or conditional branches wherever errors occurred. But our first Agent execution framework didn't use fixed workflows dictating each step; instead, the model decided next actions autonomously.
Whether doing research, slides, websites, or data analysis, the model decided next steps rather than rigid processes chaining everything. Multi-agent systems were hot then, but we persisted with single-agent routes.
Why stick to this? Less preset structure maximizes benefit from rising model capabilities. Writing rigid workflows prevents enjoying dividends from model updates. Before 3.7, we didn't know if this judgment held. But on launch night, we confirmed: Without changing execution frameworks, just improving model capabilities boosted product performance. This meant not just this time, but potentially continuous benefits from future model progress. So for AI entrepreneurs, crucially, don't patch current model weaknesses endlessly, as those will eventually be solved. Instead, ask: What will models look like in 6 months? If so, what suitable product form matches that future model?
Start now. In 3 months, products might seem poor, but maintain strong conviction: It's just because models haven't caught up. Once they do, product form advantages manifest immediately. When models truly reach that stage, you'll be the only one positioned there, fully prepared.
1.6.1 A Forward-Looking Example: Generative Experience with One-Second Latency
You might find this abstract. What can actually be done? Beyond what we've done, are there other opportunities? I can share one opportunity (which I won't pursue myself):
I've been waiting for a moment: Video generation models like Seedance reducing latency from prompt input to visual output to under one second. I believe that will be a critical turning point for this round of AI app entrepreneurship. Because in my view, a one-second-latency video generation model isn't just a tool anymore—it's a new-generation operating system.
When this future arrives, what should software look like? How should users interact with one-second-latency generative products? My judgment is these changes will happen within six months. Preparing around this direction is a great entrepreneurial opportunity. This is just one example. Many models progress in various fields. What will they look like in 3 or 6 months? How should ordinary people harness these capabilities? This is the core opportunity for AI app entrepreneurship.
1.7 Lowering Input Barriers, Using Payment to Identify True Value
Next, the second judgment: Why don't we build Chatbots? Because their ceiling is easily limited by user prompts; only highly knowledgeable users extract full value, making products picky. Future new product attempts aim to minimize user input costs, letting users continuously feel value from model intelligence in daily life.
In the AI era, the most important product value is letting users feel intelligence. Many don't realize this, thinking value lies in being free. It's never about free; free subsidies are the biggest trap in the AI era.
Looking back at China's AI app entrepreneurship over the past 2-3 years, the biggest trap was being misled by previous mobile internet free-subsidy thinking, assuming this generation could do the same. Completely wrong:
First, if intelligence performance is good enough, people will pay. For instance, some core platform users spend $4,000-$5,000 monthly. High amounts, but willingness indicates found value.
Second, free subsidies trap teams in invalid feedback. Of course listen to feedback, but free users demand "I want this, I want that." These demands aren't worthless but may distract entrepreneurs. Looking at irrelevant demands, thinking about wrong directions essentially wastes time.
Third, sticking to paid models ensures feedback comes from users who've felt value and voted with money/actions. Building around their real needs better attracts other paying users.
Over the past 2-3 years, domestic attempts to grow via early free subsidies (whether Coding Agents or others) mostly failed. People blamed weak branding, but the core reason was going free. This is especially prominent in big companies: Rich resources allow spending tens of millions on subsidies day one, leading down a twisted path. I can't do free. If you hit 100k DAU in month one, it becomes your KPI. Next month you only want higher numbers; no one dares switch to charging. You carry this free DAU burden, impairing perception of real user markets. Crucial point: Letting users feel true intelligence performance is the most important thing in AI-era product building.
We didn't launch paid features initially, starting charges late March 2025. By April's first month, ARR exceeded 40 million, approaching 50 million. Why? Because the product truly delivered overwhelming intelligence. That's key.
1.8 Why Persist with General-Purpose: User Mindset and Long-Tail Needs
The third judgment is persisting with general-purpose. When launching Manus, we claimed to be the world's first general AI Agent. Why emphasize general-purpose?
Early on, we heard many investor opinions, like modifying UI, but refused one suggestion. Investors asked, "What exactly can this product do? Every visit shows weird long-tail tasks." I showed internal/external eval results; they said, "Can you name one practical, value-demonstrating scenario?" I admitted I couldn't.
They suggested focusing, e.g., website generator or slide generator. Those were possible, but we didn't do them.
Retrospectively, sounding calculated, we actually didn't think that much then. Simply, during those two and a half months developing Manus, I was repeatedly shocked by intelligence, extremely excited daily. Knowing it was incredible, performing all long-tail tasks like a human, I wanted to share that joy/surprise with the world.
It was pure product mindset, not pre-planned general strategy. But looking back, insisting on general-purpose was a critical strategic judgment in Manus's journey.
Reason: Brand mindset is hard to buy with money.
Your cognitive position in users' minds is difficult to change even with billions. Since we started as "General Agent," subsequently launching website dev, slides, video commands felt natural. Users thought, "Manus does everything; this is normal." Meanwhile, vertical competitors expanding to general tasks confused users: "Aren't you a website tool? Why data analysis now?" Don't underestimate this.
A product's cognitive occupation in users' minds usually happens once in a lifetime. Changing it is immensely hard, possibly taking years of major strategic/product adjustments after survival. Insisting on general-purpose let us enjoy dividends in every AI wave, continuing to today.
Another important point: Internally, we have a matrix: X-axis = money paid to us; Y-axis = user satisfaction. Who occupies the top-right corner (highest payment, highest satisfaction)?
Not coders. We target non-technical users; coder usage exists but is lower. Most satisfied are corporate managers: CEOs, small bosses, department heads. Why? Bosses have most miscellaneous tasks, thinking about this today, that tomorrow. They find Manus incredibly satisfying. Before Manus, no AI tool solved everything/discussed anything for them. Previous vertical AIs (e.g., Marketing Agents) were limited; only Manus did everything. So CEO groups used it happily, becoming loyal high-end users. Insisting on general-purpose secured high-end clients and built mindshare barriers, surviving 1-2 years of tech iteration without elimination.
Of course, not everyone should do general-purpose. We succeeded because no one else did at that time, seizing mindshare first. Trying later is much harder.
1.9 Design Around User Life, Not Technical Habits
The fourth judgment is understanding users.
As mentioned, we defined users as non-technical from the start. We aimed to empower all ordinary people with explosive model capabilities, not just programmers.
Thus, we did extensive UX work:
First, building trust through transparent execution processes.
Second, asynchronous cloud execution. Programmers using Cursor click confirmations after commands, finding it cool. But ordinary people don't stare at computers all day; they might use phones or do other things. So we equipped each task with a cloud sandbox, running tasks in the cloud. Common scenario: Pull out phone, assign task, put phone back pocket, no need to keep screen on, check results when done.
Third, making the process itself shareable. We designed Session Replays narratively. Early on, many couldn't use products due to invite codes, but understood value via replays.
1.9.1 Behind Invite Codes: Funding and Inference Capacity Constraints
Incidentally, why strictly use invite codes early?
Simple reasons. Like shooting sofa videos: We did it, others copied; we used invite codes, others copied. But knowing underlying reasons makes imitators' actions comical.
Invite codes were purely because we had ~$8-9 million left. Just raised a round, but still a poor company. Before Manus, many domestic VCs didn't favor us; compared to flashy AI peers, our cash was scarce.
Even with invite codes, peak daily burn hit $500k. Fully open, I feared the company might collapse within a week of March 5 launch. Lacking big-company resources, unexpectedly huge launch response forced us to control pace via invite codes, then launch paid features. During unpaid period, invite codes managed consumption.
One reason above; second reason unknown outside industry: Today, buying APIs/tokens from AWS/GCP is usually fine, unlimited supply (mostly because products don't scale that big). Our challenge? Manus was the world's first consumer-facing general Agent. Previously, cloud platforms served Chatbot/companion apps. In early tests comparing Manus (Agent) vs Monica (Chatbot, our colleague's product), single DAU token consumption differed by 200-1000x.
How big is this number? Procuring model call throughput from clouds uses TPM (Tokens Per Minute). M = Million. In Asia-Pacific, large chat/companion apps needing 10M-20M TPM were top-tier clients. Three days post-launch, fully opening invite-code holders, demand hit ~800M TPM.
Concept? Calling a cloud vendor (unnamed here), I said we needed 500M TPM. Accustomed to startups exaggerating for discounts, they asked, "Is this projected volume for next year?"
I said no, I need 500M now.
They couldn't provide it. All cloud vendors lacked sufficient compute reserves; capacity grew incrementally watching each other. Early on, we couldn't afford burning that much, and cloud services weren't prepared for AI Agent explosions.
Fortunately, we secured hundreds of M TPM from one cloud platform, surviving the first month. Afterwards, we continuously communicated with vendors, telling them Agents would explode, urging capacity prep.
By May when invite codes ended, many platforms offered 30-50M TPM to entrepreneurs due to prior increases.
So when I shared this in May/June, people disbelieved, calling us braggarts: "We never faced this; asking for 10M anytime works."
I thought: First, two months ago you might not get it; second, 10M vs 1000M differ by 100x. People didn't realize how much supply chain trouble Manus's launch demand caused. Cloud resources, infrastructure, payments—all required pioneers to solve. Later Agent entrepreneurs might not face identical issues.
1.10 Advice for Young Entrepreneurs
Finally, many here might become AI-native entrepreneurs. Here are suggestions.
1.10.1 See Global from Day One, Not Just West
First, global vision.
What is global vision? Last June, VivaTech, Europe's major tech summit, invited us. Happy to speak at a global phenomenon product's summit. Three days before departure, organizers promised a surprise. That evening, Gmail received email from elysee.fr (French Presidential Palace domain): President Macron invited me to dinner.
First reaction: Scam. Normal people think scam, right? Organizers confirmed it was the surprise. After summit, I dined with Macron at Elysée Palace. Recently posting on Xiaohongshu riding Jensen Huang's traffic happened because Macron sat to my right, Huang opposite. Huang gave brilliant speech; I recorded/posted it, trending recently.
Why tell this? Product launched only 3+ months prior, yet invited to such high-profile events. Dozens of tables filled with European founders/investors. Post-dinner chats revealed flawed startup ideas.
European startup ecosystem lags China significantly, defensive mindset: "Build in Europe, for Europe." Many advocated strengthening local AI so European firms/govs use European AI, fearing US-China dominance.
I felt uneasy. Nearby American founders agreed it was unbelievable. Formal setting prevented blunt speech, but internally: If we'd adopted "Build in China, only for China," I wouldn't sit here today.
We believe LLMs' greatest value includes drastically lowering language barriers. Among our three leaders, Red and I had no overseas study/living experience pre-startup; Red hadn't visited US. Only Chief Scientist lived abroad briefly as child. Precisely because language barriers dropped, we marketed/served/thought globally, creating such products.
Yet many young European founders tried locking themselves in Europe. Wrong approach. I mentioned this during exchanges.
Today, encouraging domestic young AI-native founders: Consider global expansion/user service from day one. With language barriers removed, going overseas is absolutely not a problem. Our internal systems/customer service are multilingual because barriers no longer exist.
AI is heaven's gift to Chinese entrepreneurs. Previously competing with US tech exports was tough due to English dominance/cultural disadvantages. But in AI, we possess strong ability to win global markets via product/dimensional innovation. So global vision is paramount for AI entrepreneurship.
Emphasizing: Global ≠ only West. World is vast; Global South offers opportunities. Largest revenue market is US, but most users reside in Brazil. Middle East, Europe, Japan/Korea, SE Asia, Africa—all have users.
Never say "abroad" when globalizing. I correct team members internally banning "foreign countries" because it's vague. "Foreign" implies binary opposition with "domestic," oversimplifying complex global distribution across regions/languages/cultures. True global vision spotting regional opportunities generates new insights.
1.10.2 Use Products Personally, Don't Replace Judgment with Emotional Narratives
Second advice for product folks: Read less self-media.
Clarification: Not targeting quality creators here. But product builders reading only self-media articles fall into nihilism: Today "nuke-level explosion," tomorrow "industry dead." Since Manus launch, countless times declared "disrupted"/"unused," yet business consistently grew.
Most self-media lacks industry experience; traffic drives writing. They cater to emotions, stimulating dopamine via happiness/anger/confusion, manipulating you. Perceiving AI development this way creates massive disconnect from reality.
E.g., 2-3 months post-launch, TPM hit 800-900M. Globally, except Cursor/Windsurf etc., few startups burned more tokens. Yet many self-media still taught "how to build Agents."
Language is power/tool. Written views appear equal, but bases differ. Some lack market contact/user research/online validation; views are imaginary. Daily consumption drags you emotionally.
Real product sense lies in details. Hearing "OpenAI released X, industry dead," try it yourself: Many OpenAI products suck. Ten launches yield one success due to manpower/funding/base strength; nine fail. Watching self-media concluding "big players did it, track closed" loses opportunities. Everything requires hands-on touching/using/practicing.
Interviewing AI PMs lately felt awful. Campus recruits claiming AI PM interest listed one tool; asked to show recent chats, phone lacked app or had 2-3 tasks—clearly cramming for interview.
AI disrupting industries for 3-4 years, yet AI PM candidates don't use AI. Shows resume screening flaws (now requiring actual usage proof), but fundamentally scary: People don't realize prerequisite for AI work is deep personal AI usage. Only firsthand experience reveals current model/product boundaries/new opportunities. Consuming digested content never conveys reality.
You can write articles/self-media influencing others (team members do; I occasionally); but never rely solely on self-media for nourishment, else industry cognition becomes deformed, unrepresentative of truth.
1.10.3 Discover Problems, Embrace Unknowns, Correct Through Action
Third advice, not just for AI PMs, given to aspiring excellent PMs for over a decade:
Current models/tech are abundant; scarcity isn't tech/exec, but problem discovery. Ability to discover/define problems, clearly describing resolved states, is rare. Not acquired merely listening to startup shares. Being young/in campus, pressure low, observe society/life, classmates/intern colleagues, identify real problems awaiting solutions.
How does this social mechanism operate? Where can AI improve efficiency? All stem from life observation. Hope for greater curiosity toward life itself: Curiosity drives understanding/refinement. But many currently lack curiosity.
Nearly two years post-Manus launch, young entrepreneurs frequently asked "How did you succeed?" Seeking deterministic answer sheets, believing following ten steps guarantees success. Never true. Entrepreneurship is highly uncertain. Most decisions made without full clarity, acting anyway, dynamically adjusting.
This distinguishes entrepreneurs from VCs. VCs conduct lengthy research because decisions are heavy/unreversible; invested money stays invested. They seek high accuracy via pre-research, asking about endgames/pathways. Entrepreneurs cannot think this way. Pre-defining everything/best paths means it's NOT the best path.
Entrepreneurs live dynamically in battlefields, observing daily changes, understanding world/market/tech shifts, making dynamic daily decisions. Excellent entrepreneurs operate thus.
Past 2-3 years, strong feeling: Increasingly, young entrepreneurs fear embracing unknowns, demanding permanent answers: How to shoot this gun? How to handle marketing/product/tech/globalization? Craving certainty. When unable to provide, they suspect hidden secrets; hearing "no secrets, act first," they deem unwillingness to share.
Thus, stay curious about world. Unknowns are most beautiful. Big opportunities lie in unknowns, not certainties. If certainty existed, we'd have built AI browsers instead.
1.11 Conclusion: Let Minds Grow Hands
Finally, many may not know Manus's name origin. Derived from MIT motto Mens et Manus, Latin for Mind and Hand. Why choose this? Because all large model vendors build Minds, but at end-2024, why couldn't such strong minds exert huge influence? Because they had brains but no hands. Manus represents large models' hands, enabling brain thoughts to affect real world, solving genuine problems.
This applies to today's theme: All thoughts require practice. Entrepreneurship's essence is practice, not mere PPT simulations in contests or flashy fundraising decks.
Daring to practice, facing real society, staying curious, solving unsolved world problems—that's true entrepreneurship spirit, founders' duty—not playing house pretending to be entrepreneurs. Thank you.





