Reported by Synced
Editor: Taozi
[Synced Intro] Vibe coding is mass-producing "AI nannies." A 15-year veteran developer, rushing to meet deadlines with AI programming, ended up with mountains of bugs and had to scrap the project, crying for half an hour. Now, a new profession, "Vibe Code Cleanup Specialist," is trending.
The trending "vibe coding" is now turning countless programmers into "AI nannies."
After 15 years in the industry, Carla Rover had to restart a project and cried for half an hour after using Vibe Coding.
It might sound absurd, but this is a true story.
The term "Vibe Coding" was first proposed by Karpathy and swept through the AI community overnight.
Whether it's Cursor, Copilot, Codex, or Replit, all these AI tools have become powerful assistants for developers in AI coding.
However, only those who have used them truly understand the actual experience.
Rover's experience has become a typical portrayal of many senior programmers using AI for coding today —
They themselves have become "AI nannies," constantly needing to rewrite and check the code generated by AI.
Recently, a report from Fastly showed that among nearly 800 developers, at least 95% needed extra time to fix AI-generated code.
And the bulk of the verification work falls on senior developers.
Even more interestingly, the rise of "vibe coding" has spawned a new profession: "Vibe Code Cleanup Specialist."
Some netizens joked that a Vibe Code Cleanup Specialist could earn at least $100,000 per year.
From "Vibe Coding" to Breakdown: 15-Year Programmer Cries
Carla Rover is a senior web developer, having been in the software industry for 15 years since 2010.
She is now co-founding a startup with her son, building customized ML models for various markets.
Just as AI tools were being hyped, Rover, in a rush to advance her startup, decided to embrace "vibe coding."
"I just wanted to achieve results quickly and took a shortcut, not even bothering to look at the automatically reviewed files," she recounted.
As a result, when Rover manually checked, the bugs were unbelievably numerous; a subsequent check with a third-party tool revealed even more issues.
In the end, they had no choice but to scrap the entire project and start over, which caused Rover to break down and cry for half an hour.
In an interview, Rover recalled, "I truly treated Copilot like an employee back then, dumping everything for AI to handle. I really learned my lesson from this experience."
Rover likens "vibe coding" to a cocktail napkin, allowing developers to sketch ideas as they please.
But she also complained that dealing with AI-generated code, expecting it to go live directly, was "more exhausting than raising children."
Because these AI models always mess up the work in various unexpected ways.
The creator of Linux once joked that VIBE stands for "very inefficient but entertaining."
In practice, some senior developers have found that AI-generated code is riddled with bugs —
Sometimes AI invents non-existent package names, sometimes it inexplicably deletes critical information, and it even secretly embeds security vulnerabilities.
If not carefully checked, the code written by AI could lead to products with more flaws than those written by humans.
In response, Rover drew a vivid analogy: using AI for programming is like asking a clever but 6-year-old child to serve coffee to the whole family, pouring a cup for everyone.
Can they do it? Maybe.
But will they mess it up? 100% yes, and the worst part is, they probably won't tell you after they've messed up.
However, Rover believes it's not that the child isn't smart, but rather that this kind of task simply cannot be completely handed over to them.
Explaining the Same Thing 15 Times, Harder to Manage Than a Child
Another developer, Feridoon Malekzadeh, also feels that using vibe coding is like raising a child.
He has worked in this field for over 20 years, holding positions in product development, software, and design.
Currently, Malekzadeh has also founded his own startup and often uses "vibe coding" platforms like Lovable.
Out of interest, he even created a small AI application specifically to generate "Gen Z" internet slang for people in their 60s and 70s.
Malekzadeh enjoys being able to complete projects independently, which saves both money and time, but he also found that "vibe coding" is completely different from hiring interns or junior programmers.
"It's more like hiring your stubborn, defiant teenager to help," he complained.
He griped, "You have to say the same thing fifteen times, and it only does a fraction of what you asked, then conveniently creates a bunch of things you didn't ask for, and messes up other stuff."
Some netizens have also compiled a collection of "failed commands," recreating hilarious scenes of programmers being driven crazy by AI in "vibe coding."
Malekzadeh roughly estimated his time allocation in "vibe coding" as follows —
50% writing requirements, 10-20% "vibe coding," and the remaining 30-40% entirely on "vibe fixing."
That is, most of the time is spent fixing bugs generated by AI code and a pile of unusable "scripts."
In reality, "vibe coding" does not possess advanced capabilities like "systemic thinking," which involves understanding how a complex problem affects the entire system.
In Malekzadeh's view, AI-generated code tends to solve more superficial problems.
He explained with an example: if a general function needs to be created, an excellent engineer would write it only once and call it whenever needed in all relevant places.
However, with "vibe coding," if five places require it, it will generate five different implementations. Not only will the user be confused, but the AI itself will also get disorganized.
On the other hand, Rover also discovered that as soon as real data conflicts with AI's built-in instructions, the AI simply "crashes."
"It might offer misleading advice, omit crucial elements, or even interfere with the ideas you're developing."
She also found that AI never admits mistakes; instead, it stubbornly argues or even fabricates results.
Once, when Rover directly questioned the AI's results, the AI began to ramble on, pretending its answers were based on data she had uploaded.
Only when Rover exposed it did the AI admit that it hadn't used the data at all.
Code Pitfalls, Fill Them Yourself
Even more troublesome are the AI coding security issues, which are enough to cause headaches.
Austin Spires, Director of Development at Fastly, has also been programming for over 20 years.
Based on his personal experience and customer feedback, he found that "vibe coding" tends to prioritize speed over stability, often leading to novice-level mistakes.
Typically, engineers need to review the code, correct AI's errors, and tell it, "You messed up here."
AI would immediately reply, "You're absolutely right," with such a good attitude that one couldn't stay angry.
Mike Arrowsmith, CTO of IT management company NinjaOne, has spent two decades in software and security.
He stated that "vibe coding" is creating a wave of new IT blind spots and security vulnerabilities, especially for startups, which are particularly susceptible to falling into them.
"Vibe coding" often bypasses strict review, but these processes are the foundation of traditional coding; without them, vulnerabilities cannot be hidden.
Painful, Yet Joyful
Almost everyone agrees that AI-generated code and "vibe coding" are useful in many scenarios.
However, they also unanimously agree that manual review is essential before relying on it to build commercial projects.
Despite widespread complaints, "vibe coding" has indeed fundamentally changed the coding workflow for many developers.
Rover stated that "vibe coding" helped her create better user interfaces; Malekzadeh frankly admitted that while fixing code takes time, with AI's help, they ultimately accomplish more work than before.
Fastly's survey also found that senior developers are twice as likely as junior developers to deploy AI-generated code to production environments.
This is because they believe the technology genuinely helps them accelerate their work.
Young engineer Elvis Kimara is experiencing these firsthand; after graduating with a master's degree, he is developing an AI trading platform.
Like many programmers, he feels that "vibe coding" has made work harder and particularly unrewarding.
"AI solves problems so easily that the dopamine from solving problems independently has vanished," he noted.
In his previous job, he mentioned that some senior developers at the company were no longer willing to mentor new hires—some hadn't fully grasped "vibe coding," while others simply delegated the mentoring task to AI.
Even so, Kimara firmly stated that even if he becomes a senior engineer, he will continue "vibe coding," considering it a true accelerator for himself.
The future of AI programming goes beyond writing code; it's about guiding AI, holding it accountable when issues arise, and acting as an AI consultant.
It seems that spending more time buried in the "weeds" of "vibe coding" to fix bugs has become a "tax" developers must pay for using this innovation.
Reference: