Earlier this year, former OpenAI exec Andrej Karpathy coined a new term — “vibe coding” — for using artificial intelligence to rapidly develop software using natural language prompts.
But the approach comes with some glaring shortcomings that have gradually come to light, from major cybersecurity problems leading to mass leaking of sensitive personal information to rampant hallucinations that turn vibe-coded projects into a buggy mess that has to be painstakingly fixed by human programmers.
Even Karpathy himself has seemingly fallen out of love with his own creation. His latest project, dubbed Nanochat, is a “minimal, from scratch” interface that strips down a ChatGPT-like experience to its very basics.
“You boot up a cloud [graphics processing unit] box, run a single script and in as little as four hours later you can talk to your own [large language model] in a ChatGPT-like web UI,” he boasted in a recent tweet.
But as it turns out, the project wasn’t the result of AI vibe coding — it was Karpathy himself.
“It’s basically entirely hand-written,” Karpathy wrote in a followup. “I tried to use Claude/Codex agents a few times but they just didn’t work well enough at all and net unhelpful, possibly the repo is too far off the data distribution.”
In other words, even the godfather of vibe coding doesn’t trust the tech enough to use it on his own project.
To be fair, even Karpathy himself never intended for “vibe coding” to replace human developers in the long run.
“Sometimes the LLMs can’t fix a bug so I just work around it or ask for random changes until it goes away,” he wrote in the February tweet in which he first coined the term. “It’s not too bad for throwaway weekend projects, but still quite amusing.”
But overrelying on the technique can have disastrous consequences as companies continue to cut costs in favor of investing in AI — regardless of Karpathy’s original intentions.
As 404 Media reported last month, a growing number of coders are being tasked with fixing AI-hallucinated code. At best, projects never reach a satisfying level of polish. At worst, the shoddily-put-together lines of code can wipe out entire databases.
Researchers have also found that AI-assisted coding can actually slow down human developers, instead of boosting their productivity.
In a recent report, management consultants Bain & Company found that despite being “one of the first areas to deploy generative AI,” the “savings have been unremarkable” in programming.
“Generative AI arrived on the scene with sky-high expectations, and many companies rushed into pilot projects,” the consultants wrote. “Yet the results haven’t lived up to the hype.”
Content delivery platform Fastly similarly found that at least 95 percent of 800 developers it surveyed had to spend extra time fixing AI-generated code.
Experts have also warned that the trend could result in human coders never learning the ropes properly. Leaning on AI coding too much may be “a bit of an impending disaster” as MIT computer scientist Daniel Jackson told Wired earlier this year.
“Not only will we have masses of broken code, full of security vulnerabilities, but we’ll have a new generation of programmers incapable of dealing with those vulnerabilities,” he added.
More on vibe coding: Amateurs Using AI to “Vibe Code” Are Now Begging Real Programmers to Fix Their Botched Software
The post Inventor of Vibe Coding Admits He Hand-Coded His New Project appeared first on Futurism.