The world is not linear extrapolation: be the important variable in the game
If the world ran on logical deduction, today's biggest recommendation engine company would be Baidu. Xiao Hong says what a founder must do is not deduce the inevitable, but make themselves the variable that changes the outcome of the game.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
VC is expensive not when you're doing badly
Xiao Hong defines capital as a tool, and a very expensive means of raising money. The mechanism he gives: VC is not expensive when you're doing badly — under the terms, orthodox American financing doesn't have to be repaid — it's expensive when you're doing well, when you're killing it, the price you pay is bigger. From this he derives his judgment of a competitor: that company raised more money, so it must deliver a bigger story to its investors, and is therefore forced into doing the wrong things. His conclusion is that even if what you end up doing is smaller, it's far better than doing the wrong thing, and a CEO should have the courage to do the right but small thing.
— Xiao HongReading the platform: plug-ins will always get killed
When WeCom opened its interfaces, Xiao Hong judged that plug-ins would definitely be cleaned up by the platform, but nobody knew when. His reasoning: if you only think about customers, of course customers want to send as many ads as possible; but if you think about the platform, you know the WeChat ecosystem can't be allowed to fall into chaos, and Tencent will definitely restrict it. So the key move is not to wait until plug-ins get killed before developing, but to build ahead of time — because developing a SCRM tool takes at least a month, and by the time that day comes it's too late to start. He sums it up as a reusable principle: facing something you're almost certain will happen but don't know when, the earlier you prepare within the range of costs you can bear, the better.
— Xiao HongWhile Jasper celebrated its expansion, he saw the cracks
Xiao Hong studied Jasper, the most successful company running on the GPT-3 API, and saw the founder write an open letter celebrating the company growing from 9 people to over 200. His judgment was the opposite: management itself is anti-scale, the more people the more messes, and this company's execution could slow down. His second judgment was that Jasper was priced too high, and by the plain understanding of the computer industry, capability and API costs would fall sharply, so a market skewed toward individuals and professionals rather than enterprise procurement should emerge. Third, after ChatGPT came out he confirmed that conversational was the public's imagination of AI. All three together made him decide to start Monica before ChatGPT was even released, with an internal goal of just making 5,000 RMB a month.
— Xiao HongBought growth teaches the team nothing
After acquiring ChatGPT for Google, Xiao Hong didn't rename it Monica but built a new product instead. The reason wasn't branding but growth feedback: ChatGPT for Google had tens of thousands of new users a day from organic traffic, and if the team just sat on it, a colleague who found a way to bring in 200 new users a day wouldn't feel it was anything special, and the team's rate of learning would slow. By rebuilding Monica, everyone gets excited when they find even a little bit of new growth themselves, and thereby learns how to do growth and operations overseas. At the same time, a bought product is an asset, and funneling traffic to it costs nothing, whereas ad spend is money thrown into the water, which makes people hesitate.
— Xiao HongThe Andy-Bill law of the new era
Xiao Hong uses the semiconductor industry's Andy-Bill law to explain AI: no matter what Andy builds, Bill will eat it. Applied today, model capabilities keep evolving and getting cheaper, but the shell also needs to evolve. Cursor was founded very early, but it only became widely known after Cloud 3.5 Sonnet was released — the improvement in model capability brought an iteration in its product capability. Without Cursor, you could still write code with Cloud 3.5 Sonnet inside Claude, but it wouldn't be as smooth. So after model capability spills over, third-party vendors are needed to present the user-perceivable value, and that's the part founders can define.
— Xiao HongShowing the thinking process is an experience innovation
Xiao Hong believes DeepSeek's global explosion wasn't just about strong technology, but about presenting its technical innovations in a user-perceivable way. He quotes the CEO of Publicity: the two huge experience innovations of the AI era are, first, marking citation sources when answering with web access, and second, showing the LM's thinking process. OpenAI's o1 actually thinks too, but o1 charges money and only shows a simplified version, and many people simply don't know this exists, so it missed this experience innovation. DeepSeek is free and shows it in full, and with model quality genuinely pulled into the first tier, even friends back home can feel the generational difference.
— Xiao HongAn Agent should write its own code and call APIs
Xiao Hong describes the moment he was struck by lightning: using Windsurf's Yolo mode, the Agent said it was going to GitHub to pull code, do something, and then write. He realized it was actually using tools. From this he derives the Agent's architecture: a virtual server, a browser, able to write its own code and call APIs. Because much of human knowledge and many services are invoked not through APIs but through the web, it must have a browser; because it needs to execute command lines and install libraries, it needs a virtual machine in the cloud, and it doesn't matter if it breaks, it can be released when the task ends. He also points out that Yolo mode making users fill in yes is passing the buck — novice users have no idea what yes or no means.
— Xiao HongThink in terms of games, not logical deduction
Xiao Hong distinguishes two ways of thinking: logical deduction is ‘Baidu has the best algorithm engineers, so Baidu will definitely do recommendations’; game thinking is that because you appeared and other players appeared, the whole environment is different. He gives examples: without this DeepSeek wave, nobody would be thinking about open source; if the original ChatGPT had been built by a third-party company, OpenAI might have chosen to be a pure platform company rather than a consumer app company. His previous startup doing WeCom CRM was the same — because someone in the ecosystem was doing it very well, WeCom said it would leave it to the ecosystem. So founders should try as much as possible to make themselves that important variable.
— Xiao HongIn their own words · checked verbatim
But it's not expensive when you're doing badly. It's expensive when you're doing well. When you're killing it, the price you pay is bigger.
但是贵不体现在你不好的时候 他的贵体现在你好的时候 你牛逼了 你付出的代价是更大的
Xiao Hong33:35
You're facing something you're almost certain will happen — which is, plug-ins will get killed — you think it will definitely happen, but you don't know when. What can you do? Within the range of costs you can bear, the earlier you prepare for it, the better.
你面对一个 你几乎认为一定会发生的 which is 外挂会被干掉 你认为这一定会发生的 但是不知道什么时候会发生的事情 你能做的事情是什么呢 是在你能够承受得起的成本范围之内 越早为它做好准备越好
Xiao Hong44:45
In the AI era there are two huge experience innovations. One is that if you're connected to the web, it can answer with web access, and it marks out which webpage each sentence is cited from — that makes people trust it more, right, increases the credibility of the result, having a source. Yes, having a source. The second huge experience innovation he mentioned is showing the LM's thinking process.
在AI时代有两个巨大的体验创新 一个是如果你连网 能回答连网 他把这句话从引用哪个网页标准出来 这样会让人更相信对吧 增加结果的可信度 有source 对有source 第二个他说的巨大的体验创新 就是把LM思考的过程
Xiao Hong2:26:59
If the world ran on logical deduction, today's biggest recommendation engine company would be Baidu. So first, it is not logical deduction — knowing that, I think, is already very important for everyone. Second, try as much as possible to make yourself an important variable. Third, how do you try? There's no way — you can only be yourself.
如果世界是逻辑推导 今天最大的推荐引擎公司是百度 所以第一它不是逻辑推导的 首先知道这一点我觉得对大家来说就很重要 第二 尽量让自己成为很重要的变量 第三 如何尽量成为呢 没有办法 就只能be yourself
Xiao Hong3:13:36
Figures
| Monica's internal revenue goal at project launch | 5,000 RMB a month | 1:12:08 |
| Monica's user scale | several million users | 1:30:24 |
| ChatGPT for Google's user scale | several million users | 1:30:24 |
| Revenue at the sale of Xiao Hong's first startup | close to 10 million RMB | 38:41 |
| Funding raised by Xiao Hong's first startup | 1 million RMB SPA | 19:25 |
| Xiao Hong's age | 32 years old, born in 1992 | 3:04 |
| Registration date of Xiao Hong's first company | January 20, 2015 | 6:10 |
| Number of languages Monica supports | several dozen | 1:33:25 |
| When Xiao Hong first went to the US | last September | 1:35:27 |
Glossary
- Andy-Bill law
- No matter what Andy (Intel) builds, Bill (Microsoft) will eat it — hardware capability is ultimately consumed by software.
- Agentic
- A model's ability to plan over long horizons, solve problems step by step, and call tools, as distinct from solving things in a single round of conversation.
- Yolo mode
- You only look once — the Agent automatically executes operations like error fixes instead of asking the user to confirm step by step.
- PMF
- Product-Market Fit; Xiao Hong says it's like mathematics, if it's not there, it's not there.
- GAIA
- The complex-task benchmark mentioned when Deep Research was released, requiring multi-step reasoning and tool calls.
How to listen
AI application founders, heads of product going overseas, investors watching the Agent space; especially suited to teams currently agonizing over whether to build a model, and whether to build an Agent.
The rapid-fire Q&A starting at 3:03 and the first startup retrospective from 5:15-33:16 can be fast-forwarded; the second half is worth more.