The Trap of Vibe Coding: You're Just Making Garbage Faster

When Writing Code Is No Longer the Barrier

Recently, one term has gotten a lot of people excited—Vibe Coding. Put simply, it means you don't need to seriously write code, and you don't need to be particularly tech-savvy. As long as you can converse with AI, it can help you build websites, make apps, create tools, set up backends, and build databases.

It's worth noting that the term Vibe Coding didn't come out of nowhere. It was formally proposed and named by Andrej Karpathy—OpenAI co-founder and former Tesla AI director—in early 2025, and quickly became a buzzword in the tech world. Its core idea is: developers no longer write and review code line by line, but instead describe their intent to AI in natural language, and the large model automatically handles generating, modifying, and debugging the code. Karpathy described this state as "fully giving in to the vibes, almost forgetting that the code even exists." Behind this is a leap in the code-generation capabilities of large language models, along with the maturation of a whole suite of AI programming tools.

In the past, if you wanted to build a product, you had to know frontend, backend, and deployment. Now you just say "help me build an app like so-and-so," and the AI gets to work. Pages, buttons, and logos appear instantly, the official website is there too, login and registration are figured out, and even those fancy-looking pricing pages, user reviews, product introductions, and FAQs are all set up for you.

You look at the screen and suddenly feel like you've leveled up. But Bilibili content creator Haozi said something not-so-pleasant in his latest video: Many people, after mastering Vibe Coding, haven't become creators—they've become more efficient garbage producers.

Efficiency Up 67x, but Value Is Zero

When there was no AI, an ordinary person trying to build an app was actually pretty inefficient. Spending half a weekend cooped up might yield just a half-finished product—ugly interface, lousy features, lots of bugs, no logo, no website, no users, and certainly no revenue.

This stuff was of course useless, but it had one hidden benefit: because it was slow to build, it actually blocked a lot of garbage ideas. Building a bad project still cost you time, energy, and money, and you'd get a forced reality check—does anyone actually want this thing?

And now AI has arrived, and efficiency has exploded. In one weekend you might build sixty or seventy apps, each with a name, a logo, a beautiful official website, and a slogan that sounds impressive. But the problem is—every one of these apps has zero users.

Haozi offered a clever analogy: in the past you could only print one sheet of waste paper a day; now you can print sixty-seven sheets of colorful waste paper a day. The paper is more elegant, the layout is more sophisticated, and the titles look more like startup projects, but still nobody needs them. Efficiency really has gone up, but value hasn't gone up at all in step.

"Built It" Doesn't Mean "Made It"

The biggest threat of Vibe Coding is that it makes people mistake "not having made it" for "having succeeded."

Many people now open Cursor, Claude Code, or Codex, whip up a website in a few hours, and get excited: "I can develop too, I can build products too—are programmers doomed?"

The three tools mentioned here represent the typical forms of current AI-assisted programming tools. Cursor is an AI-native code editor deeply reworked from VS Code that can understand an entire project's context and generate whole blocks of code in natural language. Claude Code is a command-line AI programming assistant from Anthropic, good at handling complex multi-file engineering tasks. Codex comes from OpenAI and is one of the technical foundations behind GitHub Copilot. These tools have boosted the efficiency of "intent to code" conversion tenfold, making it possible for a single person to independently complete development work that used to require an entire team. But the more powerful the tools, the more they test the user's judgment.

Calm down a bit—you've only let AI turn an idea into an interface, and an interface is absolutely not a product.

The truly hard part of a product has never been writing code. It's: Who needs it? Why do they need it? How do users currently solve the problem? Are they willing to change their habits for your solution? Are they willing to pay? Will they come back after using it? How will you sell it?

AI won't solve any of these for you. AI can help you generate code, but it can't make business judgments for you; it can help you build pages, but it can't find real demand for you; it can help you design a beautiful website, but it can't bring you your first paying user.

"Demo Machines" That All Look the Same

Have you noticed that things built with Vibe Coding now all look remarkably similar?

The homepage has a big headline "Transform the Way You Work with AI," followed by three selling points "More Efficient, Smarter, Simpler," then feature cards "AI Generation, Automatic Analysis, One-Click Export, Smart Recommendations," and finally a few laughably fake user reviews "This tool completely transformed my workflow, I save three or four hours every day."

As soon as you click in you'll discover: the features are cobbled together, the flow is fake, the data is fabricated, the scenarios are empty, and there are no users. The biggest characteristic of such a product is that every part looks like a product, and every part isn't a product. It's more like a beautifully decorated show home that nobody actually lives in.

A Brutal Divergence Is Coming

Haozi predicts that a brutal divergence will emerge in the future.

The first type of person will use AI to quickly build demos: today an AI accounting app, tomorrow a resume tool, the day after an AI travel planner. Every project gets posted to their social feeds. In the end they have a pile of links, a pile of screenshots, and a pile of project names—but no users, no revenue, no repeat purchases, no assets.

The second type of person only builds some very small tools, so small they're not sexy at all. For example, automatically organizing customer inquiries, automatically generating quotes, automatically turning orders into travel reminders. They're not cool, not sexy, and don't look like they'll change the world, but they solve a real problem—someone is willing to use it, someone is willing to pay for it, and once someone uses it they can't do without it.

This is the real gap: it's not about who can build sixty or seventy apps a day, but who can turn a small problem into something real, usable, verifiable, and monetizable.

AI Amplifies Ability, and Also Amplifies Garbage

AI amplifies your abilities, and it also amplifies your garbage ideas.

• If you have industry experience, AI will help you launch quickly; if you have user insight, AI will help you validate quickly; if you have good taste, AI will help you produce designs faster.
• The reverse is equally true: if you have no judgment about demand, AI will just help you build things nobody wants faster; if you have no product ability, AI will just help you pile on features faster; if you have no distribution ability, AI will just help you build isolated-island projects faster.

Many people, after adopting AI, appear busier on the surface—building new things every day, trying new tools, posting screenshots, saying "I just launched another new project." But ask them one question: how many real users do you have? How many people pay? The answer is often silence. This isn't entrepreneurship, this is arts and crafts class.

The Right Order: Demand First, Code Later

So how should ordinary people actually use Vibe Coding?

It of course has value, and its greatest value is driving down the cost of validation. In the past, if you had an idea you had to find someone to develop it, spend money, and wait through a development cycle; now you can quickly build a small version and put it in front of users to try.

This actually echoes the core methodology of the Lean Startup. The MVP (Minimum Viable Product) concept proposed by Eric Ries in The Lean Startup advocates building, at minimal cost, a version containing only core features, releasing it to real users, and rapidly validating hypotheses through the "Build—Measure—Learn" loop, avoiding over-investment in the wrong direction. Traditionally, the biggest obstacle to an MVP was precisely the high cost and long cycle of development. The real significance of Vibe Coding lies in compressing the cost of the "Build" step to near zero. But note that it only solves the building part—"Measure" and "Learn" still depend on you.

So the order must never be wrong. The correct order should be:

1. First find a real problem
2. Then find real users
3. Then do the minimum validation
4. Finally let AI help you accelerate development

Many people get this exactly backwards: first let AI write code, then make the logo, build the website, write the marketing copy, and only at the very end remember to ask, "Who exactly am I selling this to?" Once this order is wrong, it's basically doomed. You can't build a car first and then go around asking who wants to go in this direction; you should first know where someone wants to go, then decide what vehicle to use to take them there.

Three Questions to Judge the Value of a Vibe Coding Project

To judge whether a Vibe Coding project has value, you can't look at how pretty it is—you have to look at three things:

• Is anyone willing to use it—are there users who genuinely want to open it, use it, and hand their problem over to it.
• Is anyone willing to use it repeatedly—using it once may just be trying it out; repeated use is what shows it has truly entered a workflow.
• Is anyone willing to pay for it—money is the most real feedback. A user saying "I need this" and actually pulling out their wallet are two completely different things.

If a product has no usage, no repeat use, and no payment, then no matter how pretty it is, it's just an exquisite ornament.

Creation Gets Cheaper, Judgment Gets More Expensive

So don't be fooled by the speed of Vibe Coding. Building sixty or seventy apps in one weekend sounds terrifying, but if none of them have users, it's essentially you using AI to automate ineffective labor. In the past you were producing garbage inefficiently; now you're producing garbage efficiently.

In the future, AI programming will make "creation" cheaper and cheaper, but precisely because creation is getting cheaper, judgment will become more and more expensive. More and more people can write code, more and more people can build demos, and more and more people can generate logos and websites—what's truly scarce are the people who can see real demand clearly.

There's nothing to be proud of in mastering Vibe Coding. Only when what you build has users, gets paid for, and becomes indispensable is it worth being proud of. Otherwise, those so-called efficiency gains just turn one useless app into sixty or seventy even more useless apps. This isn't awkward—it's "a productivity upgrade for garbage."

Key Takeaways