Silicon Valley isn’t just coding anymore. It’s also “vibe coding.”
Using AI to write code has been gaining traction for years, but now a new buzzword coined by Andrej Karpathy, a computer scientist who cofounded OpenAI, is capturing the movement.
This month, he described what he sees as a new kind of coding in which “you fully give in to the vibes” and “forget the code even exists.”
It’s an approach that defies conventional wisdom in the tech industry, that developing software demands virtuosic skill from engineers.
“It’s not really coding — I just see stuff, say stuff, run stuff, and copy-paste stuff, and it mostly works,” Karpathy, who also led Tesla’s AI operations for five years, wrote on X.
AI’s ability to write code has come on leaps and bounds since ChatGPT’s release in late 2022. Less than two months after the chatbot’s release, Karpathy said on X, “The hottest new programming language is English” — an allusion to how smart prompting could generate good lines of code.
Software engineers have remained in hot demand since then, but the arrival of AI that can “vibe” code into existence has some industry leaders predicting big changes.
OpenAI CEO Sam Altman said during a visit to India in early February that he expected software engineering to be “very different by the end of 2025.” And last month, Mark Zuckerberg said on “The Joe Rogan Experience” that AI would soon do the work of midlevel Meta engineers.
How Silicon Valley is starting to vibe code
To give a sense of what vibe coding looks like in action, Karpathy shared a few ways in which he’d been using AI.
In one example, he said he’d been using a digital workspace tool called Composer — made by OpenAI and the Andreessen Horowitz-backed startup Cursor AI — alongside the Sonnet model from the AI lab Anthropic.
Cursor’s Composer tool is an AI coding assistant that it says can help users “explore code, write new features, and modify existing code.” When it’s used with Anthropic’s artificial intelligence, a popular choice for programmers looking to AI for help, making an app from scratch becomes easier. That’s because the AI just needs the user to guide it.
Meta
In another example, Karpathy said he could just talk to Composer by using Superwhisper — an AI-powered voice-to-text tool.
He said this meant he could do things with code in Composer without having to “barely even touch the keyboard.” If there’s a mistake, his approach appears to be just as simple: “When I get error messages, I just copy and paste them in with no comment; usually, that fixes it,” he said.
Others are doing similar things at a time when AI coding agents can be given simple instructions to do the heavy lifting that would have once required seasoned engineers to spend hours reading through reams of code — and beginners can bypass a seriously steep learning curve.
“For a total beginner who’s just getting a feel for how coding works, it can be incredibly satisfying to build something that works in the space of an hour,” Harry Law, an AI researcher at the University of Cambridge, told Business Insider.
Amjad Masad, the CEO of Replit, a software company backed by A16z and Y Combinator, addressed Karpathy’s original post, saying: “75% of Replit customers never write a single line of code.”
Replit provides software that coders can use to start projects, called an “online integrated development environment,” but aims to offer a version that brings AI into the mix so users can begin to build apps with simple prompts. “Vibe coding is already here,” Masad wrote on X.
Menlo Park Lab, a startup that builds generative AI consumer applications, is also all in on vibe coding. Its founder, Misbah Syed, is a big believer in the method.
Syed told BI he uses it for the startup’s products, including Brainy Docs, which lets users convert a PDF to an explainer video with slides. Syed said if it makes mistakes, he feeds back the errors, and it usually fixes them. For him, the approach means that “if you have an idea, you’re only a few prompts away from a product,” he said.
Vibe coding has its downsides
For all its potential, experts see some risks with vibe coding.
“Ease of use is a double-edged sword,” Law said. “Beginners can make fast progress, but it might prevent them from learning about system architecture or performance.”
Law added that overreliance on AI could also create technical debt, which means that coding could become unmanageable during the scaling or debugging process, which engineers routinely have to go through. “Security vulnerabilities may also slip through without proper code review,” he said.
A senior software engineer at Microsoft who spoke with BI on the condition of anonymity as he’s not authorized to speak with the media said the vibe coding concept was “a little overhyped.”
“LLMs are great for one-off tasks but not good at maintaining or extending projects,” he said, referring to large language models. “They get lost in the requirements and generate a lot of nonsense content.”
Andrew Chen, an A16z venture capitalist, said last week that while it was “brilliant” to be able to use “the latest AI codegen tools to do ‘vibe coding,'” he found it enormously “frustrating.”
Karpathy acknowledged some of these limitations in his post, saying that sometimes an AI model “can’t fix a bug.” Still, he said he’d found that he could “just work around it or ask for random changes” until the errors disappeared.
With AI already pushing the limits of what was previously possible for programmers, it may soon be time for an industry vibe check.
Correction: February 13, 2025 — An earlier version of this story misspelled the name of a Menlo Park Lab product. It’s Brainy Docs, not Brain Docs.