The Strange State of AI Startups: Six Months of Watching from the Sidelines

"Justification by Faith"

This weekend at a Zhihu event, a.k.a. AI Construction Site Comedy Night, I did a stand-up set. Good news: I memorized the script and performed it without notes. Bad news: I forgot two whole sections, and forgot to shout out Cherry Studio and Toki 😭

Following the principle of maximum extraction, here's the full script, plus a video of the actual performance.

Hi everyone, I'm Xianyu.

My title is "actual controller" of Zangai Xianyu — that's "Buried Love Xianyu" for you. I registered a company in Beijing called Zangai Xianyu Technology Co., Ltd.

From a procedural standpoint, I'm fully entitled to call myself CEO. But since I'm the only person at this company, and I'm an honest, decent sort, it's better to stick with "actual controller."

For the past few months, I've been running a WeChat public account called Zang AI — "zang" as in "bury," AI as in AI. Burying AI.

I've noticed that AI projects today share three common characteristics.

First, the stories are always massive. "I'm building the next-generation Agent platform, the next-generation content platform, the next-generation whatever platform." And the founders all have massive vision, massive ambition to replace someone. They want to replace, including but not limited to: Douyin, WeChat, Taobao, Word, PowerPoint, Excel, Google, Meta, Amazon. (Trying for that "reciting the menu" effect here.)

Second, everyone has gotten incredibly creative with qualifiers. Today someone says, "I'm the world's first general-purpose Agent." A couple days later it's "I'm the world's first mobile general-purpose Agent." The qualifiers keep getting swapped in: world's first design Agent, L4 Agent, Office Agent, Web3 trading Agent, personal Agent — you name it.

Add enough qualifiers and you can always be first at something. For instance, I'm the world's first actual controller to do stand-up at a Zhihu event. (Forgot this whole qualifier bit 😭)

Then I realized, the third characteristic of AI projects: none of them are actually useful. I don't use a single Agent, and I'm doing just fine, living my best life.

Actually, I use AI more than most people. I write with AI every day. Any hot AI product that crosses my feed, I'll try it out.

But whenever someone asks me, "What's your most-used AI product?" I can't answer.

Because what I use most is just the base models themselves. Gemini, ChatGPT, Qwen 3, K2 — there are already so many models I can't even keep up, and the base models alone solve the vast majority of my needs.

Beyond the base models, I only regularly use two small tools.

One is Cherry Studio, a third-party client that can call LLM APIs, with pretty good version management, so I do most of my model chatting through Cherry Studio.

The other is a scheduling tool called Toki. I can send it a screenshot of a WeChat conversation where I'm making plans, or just send a voice message saying "tomorrow afternoon, this time, this place, meeting this person." It sets a calendar reminder and pings me through phone notifications when the time comes.

(Forgot to shout them out 🙂‍↕️)

Beyond that, I really don't need any other Agents.

I'm perfectly happy spending my days scrolling Zhihu and WeChat Channels. None of these Agents are as fun as Zhihu. (I genuinely scroll this every day.)

So I think the biggest problem with AI applications right now is too many Agents, not enough user demand.

With user demand this scarce, how's everyone trying to stand out?

I've observed that AI founders are in a personal-brand arms race.

Phase one: title inflation. A lot of people add me on WeChat with self-introductions that end in some random English word — orange, banana, head, foot — followed by @CEO. Fantastic. AI has achieved CEO for all, the complete decentralization of the chief executive.

Phase two: founders start reverse-inflating their credentials. MIT and Stanford pedigrees are boring now. The AI industry has moved on to reverse-competing on dropout credentials.

In the PC internet era, the legend was Bill Gates dropping out of college to build a great company.

By mobile internet, the story became: "Fuck, I love startups so much, I have to do internet startups, I dropped out of high school to start my company."

After the mobile boom collapsed, the story inflated to: founders who couldn't even read. Like the founder of that famous tea chain who tells the story: "I couldn't read, but I wanted to build a world-class beverage brand."

The AI industry has currently updated to roughly the "late mobile boom collapse" patch. We already have founders who dropped out of middle school to do AI. At this rate, I believe we'll soon have illiterate founders building AI apps. Could make a literacy Agent.

This whole scene reminds me of another former hot industry.

I've realized that the biggest, most certain trend in AI is Web3-ification.

The logic of these two industries is identical.

Web3 also produces its share of miracle figures. I follow a bunch of crypto miracle figures, see them in my feed every day.

The crypto miracle story goes: "I dropped out of middle school, got kicked out by my parents, made tens of millions trading crypto." Or the simpler version: just state your age — "I'm a 19, 18, 17, 16, 15-year-old genius trader."

How is this any different from the AI founder story? "I dropped out of high school, closed a famous investor in ten minutes, and our product hit eight-figure ARR within two weeks of launch."

And with current AI projects, you can't evaluate them on retention metrics or anything concrete. The value of an AI project comes entirely from story and belief. You see it because you believe it. This is exactly the crypto logic: "I believe this coin has value, therefore it has value."

No difference, friends.

If you studied humanities, you probably get what I'm saying immediately. This is exactly what the 16th-century Protestant Reformation called justification by faith. Both AI and Web3 are having their justification-by-faith moments. I hope a new Martin Luther emerges soon. (Improvised this part 😎)

But there is an essential difference between AI and Web3.

In Web3, a small group of people actually made money.

In AI, aside from course-sellers and KOLs, basically no one's making money.

Thank you, everyone. I'm Xianyu.

(Illustrations generated by ChatGPT, writing assisted by Gemini 2.5 Pro. Live video below.)