What Is Vibe Coding, Really?
A new wave in software creation and why it’s as misunderstood as it is promising
In early 2025, a strange phrase began trending in tech circles: vibe coding. Coined by Andrej Karpathy, the former AI chief at Tesla and co-founder of OpenAI, it sounded like a joke at first something about “embracing the vibes” and “forgetting the code even exists”. Yet just weeks later, it was added to Merriam-Webster’s list of trending terms, and 25% of Y Combinator’s startup cohort were claiming their codebases were 95% AI-generated. Something had shifted.
But what exactly is vibe coding? Is it just another name for using AI tools like GitHub Copilot or ChatGPT? Is it a Silicon Valley gimmick? Or is it the first glimpse of a new programming paradigm?
Let’s set the record straight.
Origins: Forgetting the code exists
Vibe coding refers to a style of programming in which the human describes a task in natural language and an AI model generates the code. That alone isn’t new, AI-assisted coding has been around for years. What’s distinctive about vibe coding is the attitude: it’s less about careful engineering and more about flow.
Karpathy described it as “just see stuff, say stuff, run stuff, copy-paste stuff and it mostly works.” It’s not just AI writing code for you, it’s trusting it enough not to check every line. A surrender to speed and intuition over structure and scrutiny.
It’s made possible by the rapid evolution of large language models (LLMs) that understand and generate code fluently. Tools like Replit’s Ghostwriter, Cursor and Claude in Composer allow users to “talk” their way to working applications. What used to take months can now take a weekend. For hobbyists, tinkerers and even startup founders, that’s a compelling offer.
What vibe coding is not
As the buzz grew, so did confusion. Many people now label any AI-assisted coding as vibe coding but that misses the point. As developer Simon Willison puts it, if you use an LLM to help write code, but still review, test and understand it all, you’re not vibe coding. You’re just using a tool wisely.
Vibe coding, at least in its purist form, often skips those checks. You describe your intent, the AI delivers something usable and you run with it. It’s liberating and risky.
This misinterpretation matters. If we think vibe coding is simply Copilot with enthusiasm, we miss both its creative potential and its dangers.
From hobby hacks to million-dollar games
Despite (or because of) its chaotic energy, vibe coding has delivered real results. In just weeks, it’s been used to:
Build working MVPs in hours instead of weeks
Power apps by non-coders, like Kevin Roose’s “LunchBox Buddy”
Enable a solo developer to launch a flight simulator that hit $1M ARR in 17 days
Let people prototype AI models, games, and tools with simple voice or text prompts
Some teams report doubling productivity. Replit’s CEO says 75% of users never write traditional code. These aren’t minor stats. Something is changing.
Still, there’s a catch. Several failed startups have also emerged projects built entirely via vibe coding, launched fast, and then crumbled due to bugs, security issues, or incomprehensible code. One founder famously shut his app down with a tweet: “Cursor keeps breaking the other parts of the code. I give up.”
Speed without discipline can backfire.
Why critics are worried
There are three main criticisms of vibe coding:
Quality and maintainability: AI-generated code can be messy. If no one understands it, it’s hard to debug, improve, or secure.
Developer deskilling: Relying too heavily on AI could weaken programming fundamentals, especially for newcomers.
Ethical and security risks: Models trained on vast codebases might reproduce flawed or biased patterns. Accepting outputs blindly can lead to real-world harm.
Even Karpathy cautions against using it for critical systems. His own experiments like building an iOS app in an hour without reading docs were more proof-of-concept than production-ready.
So, what is it good for?
Vibe coding works best in three scenarios:
Rapid prototyping: Testing ideas fast without building full architectures.
Personal projects: “Software for one” tools that solve a user’s own problem, quickly.
Creative coding: Games, generative art, and playful experiments where iteration is more important than robustness.
It also shines in the hands of experienced developers who know how to guide AI, troubleshoot issues, and refine outputs. Ironically, it works best for those who don’t need it.
The future: paradigm or passing trend?
Is vibe coding a gimmick, or the future?
The answer might be both. It’s not going to replace traditional programming wholesale. We still need human understanding, structured design and secure code especially for complex systems.
But vibe coding points to a different kind of programming: one that’s conversational, intuitive, and more about specifying intent than writing syntax. It’s a glimpse into what software development might become when natural language becomes the interface, and code is the by-product.
Some compare it to the shift from assembly to high-level languages. Others say it’s the spiritual cousin of the no-code movement, but with real code behind the curtain. Whatever the analogy, the shift is cultural as much as technical.
If we’re heading towards a world where “English is the new programming language”, vibe coding may be the first step into that world with all its promise and all its pitfalls.
Final vibe check
The hype will fade, but the tools are here to stay. The real challenge ahead isn’t whether AI can write our code. It’s whether we, as developers, designers and thinkers, can steer that process wisely.
Vibe coding isn’t just about moving faster. It’s about rethinking what it means to build software at all. And in that sense, it’s already changed the game.
nice