Carla Rover once had to redo a project she vibe coded, and she sobbed for thirty minutes.
During his fifteen years in the field, Rover has mostly worked as a web developer. She is currently working with her son to develop a firm that builds custom machine learning models for marketplaces.
She described vibe coding as a beautiful, unending cocktail napkin that allows one to constantly scribble ideas. However, because these AI models might alter work in unpredictable ways, dealing with AI-generated code that one intends to utilize in production can be “worse than babysitting,” she noted.
Because her business needed speed, she had resorted to AI coding, which is what AI tools offer.
“I took a shortcut and did not scan those files after the automated review because I needed to be quick and impressive,” she stated. When I completed it by hand, I discovered a lot of errors. I discovered more when I utilized a third-party program. I also took a lesson from it.
The emotions were caused by the fact that she and her kid had to redo their entire project. She said, “I passed it off as though the copilot was an employee.” “It isn’t.”
Like many seasoned programmers, Rover is using AI to assist with coding. But these programmers are also becoming AI babysitters, rewriting and fact-checking the code that the AI generates automatically.
Fastly, a content delivery platform, recently reported that at least 95% of the roughly 800 developers it polled stated they spend more time correcting AI-generated code, with senior devs bearing the brunt of this verification.
These experienced developers uncovered flaws with AI-generated code ranging from hallucinating package names to removing critical information and posing security threats. Left unchecked, AI code might result in a product that is considerably more buggy than what humans would create.
Due to the growing difficulty of working with AI-generated code, a new corporate coding position called “vibe code cleanup specialist” has emerged.
Experienced coders were interviewed by TechCrunch on their experiences with AI-generated code and what they believe will be the future of vibe coding. There were many opinions, but one thing was clear: There is still more work to be done on the technology.
Rover compared using a coding co-pilot to instructing a bright six-year-old to bring a coffee pot into the dining room and pour coffee for the family.
Are they able to pull it off? Perhaps. Could they not succeed? Indeed. And if they do fail, they probably won’t tell you about it. It doesn’t lessen the child’s intelligence,” she added. It simply indicates that you can’t assign [a task] in that way.
“You’re absolutely right!”
Additionally, Feridoon Malekzadeh compared vibe coding to a kid.
He has over 20 years of experience in the field, having held a variety of positions in software, design, and product development. According to him, he is primarily utilizing the vibe-coding platform Lovable while establishing his own firm. He also enjoys vibe coding tools like one that generates Gen Alpha slang for Boomers.
He acknowledges that employing a junior developer or intern is not the same as vibe coding, but he appreciates that he can work alone on projects, saving time and money. Rather, he told, vibe coding is like “hiring your stubborn, insolent teenager to help you do something.”
“You need to ask them to do something fifteen times,” he stated. “In the end, they break a lot of things along the way, do some of what you asked for, and do some things you didn’t ask for.”
About half of Malekzadeh’s work is spent defining requirements, 10% to 20% is spent on Vibe coding, and 30% to 40% is spent fixing Vibe, which involves addressing errors and “unnecessary script” produced by AI-written code.
Vibe coding, in his opinion, is also not the greatest at systems thinking, which is the process of determining how a complicated issue could affect a final outcome. According to him, AI-generated code aims to address relatively trivial issues.
Malekzadeh stated, “A good engineer would create a feature that should be widely available in your product once and make it available everywhere that it’s needed.” If something is required in five separate locations, vibe coding will produce it five times in five different ways. It causes a great deal of confusion for both the model and the user.
Rover discovers that AI “runs into a wall” when data contradicts its hard-coded functionality. “It can give you false advice, omit important details, or interfere with a thought process you’re going through,” she added. Also, she discovered that it will create outcomes instead of acknowledging mistakes.
She gave TechCrunch another example, in which she questioned the early outcomes that an AI model provided. Presuming to utilize the data she submitted, the model began to provide a thorough explanation. The AI model didn’t admit until she pointed it out.
She claimed, “It scared me because it sounded like a toxic coworker.”
In addition to this, there are security issues. Austin Spires, who has been coding since the early 2000s, is the senior director of developer enablement at Fastly.
He’s discovered via his personal experience and conversations with consumers that vibe code prefers to construct what’s quick rather than what’s “right.” This might expose flaws in the code that are common among rookie programmers, he noted.
Spires explained that it is common for the engineer to analyze the code, correct the agent, and inform the agent that they made a mistake. This tendency is why the ‘you’re absolutely right” stereotype has appeared on social media.
He’s alluding to how AI models, such as Anthropic Claude, answer with “you’re absolutely right” when corrected for mistakes.
Mike Arrowsmith, the chief technology officer at NinjaOne, an IT management software business, has been working in software engineering and security for almost 20 years. He claims that vibe coding is generating a new generation of IT and security blind spots, making nascent firms especially vulnerable.
According to him, Vibe code frequently avoids the rigorous review processes that are essential to traditional coding and critical to detecting vulnerabilities, he told.
NinjaOne, he added, fights this by promoting “safe vibe coding,” which includes access constraints for certified AI tools, as well as required peer review and, of course, security scanning.
The new normal
Although almost all of the people we contacted with concur that vibe-coding platforms and AI-generated code are helpful in many contexts, such as when generating ideas, they all believe that human scrutiny is crucial before launching a business on it.
Rover declared, “That cocktail napkin is not a business model.” “You need to strike a balance between ease and insight.” Despite all of the criticism of its shortcomings, vibe coding has transformed the job’s present and future.
Rover claimed that vibe coding greatly aided her in creating a more effective user interface. “Even though I spend a lot of time fixing code, I still get more done with AI coders than without them,” Malekzadeh stated.
Malekzadeh cited French theorist Paul Virilio, who talked about building the shipwreck alongside the ship, every technology has its own drawbacks that accompany technological advancement.
According to the Fastly poll, senior developers were twice as likely as junior developers to deploy AI-generated code into production, citing the technology’s ability to speed up work.
Another aspect of Spires’ coding process is vibe coding. For his own front-end and back-end projects, he leverages AI coding agents across many platforms. Though he described the technology as a mixed experience, he said it is useful for scaffolding out a test, building out boilerplate, and prototyping; it frees engineers from mundane jobs so they can concentrate on building, launching, and scaling products.
It appears that the additional hours spent sorting through the vibe weeds will only be accepted as a cost of utilizing the innovation.
A young engineer named Elvis Kimara is currently learning that. He recently earned his master’s degree in artificial intelligence and is developing a marketplace driven by AI.
He claimed that, like many other programmers, vibe coding has made his work more difficult and that it is frequently a depressing experience.
Solving a problem on my own causes no more dopamine. He stated, “The AI just works it out.” He said that skilled developers at one of his previous jobs weren’t as eager to assist young programmers – some of them were having trouble comprehending new vibe-coding models, while others assigned mentorship chores to the AI models. However, he stated that “the benefits greatly exceed the drawbacks,” and he is willing to pay the innovation tax.
He will continue to use it even when he advances into a senior post. For him, it has been a true accelerator. To learn even more quickly, he makes sure to go over each line of AI-generated code.
