Vibe Coding: Meaning, Origin, and Where It Breaks

2026-07-23 · Micro Learner AI team

Vibe coding is building software by describing what you want in natural language and letting an AI write the code — steering by vibe rather than by reading the code itself. The term was coined by Andrej Karpathy in early 2025, and within months it went from a throwaway social media post to the default word for how a growing share of software gets started.

This is our attempt at a definitive explainer: what the term means, where it came from, why it caught on, and the specific places it breaks down.

The vibe coding definition

Strip away the hype and the definition has three parts. First, the input is intent, not syntax: you say "build me a pomodoro timer with a sound at the end" instead of writing the timer yourself. Second, the AI writes the code — all of it, not just autocompleting your lines. Third, and this is the part that makes it "vibe," you steer by outcome rather than by review: you look at what the thing does, react ("make the breaks five minutes"), and iterate, often without ever reading the underlying code.

That third part is what separates vibe coding from ordinary AI-assisted programming. A developer using an AI assistant still reads diffs and owns the code. A vibe coder treats the code as an implementation detail — which is exactly why the practice spread so fast beyond professional developers, to founders, students, and operators building small tools that nobody was ever going to write for them.

Where the term came from

Andrej Karpathy — co-founder of OpenAI, former head of AI at Tesla — introduced the phrase in a February 2025 post describing his own weekend workflow: talking to an AI, accepting its code without much review, copy-pasting error messages back until things worked. His framing was half confession, half observation: this was a new way to program, and it was oddly effective. The name stuck because it was honest. It captured both the freedom (you don't need to know the syntax) and the slight recklessness (you don't actually know what you shipped) of the practice.

Why it works

Vibe coding works because it attacks the right bottleneck. For most small software, writing the code was never the scarce resource — knowing what to build and being able to express it was. Modern models are fluent enough in every mainstream language that a clear description of a small tool usually produces working or nearly-working code. Add the iteration loop — paste the error, get a fix — and a determined non-programmer can get surprisingly far on personal tools, prototypes, and one-off automations. The demand was always there; vibe coding is the first interface that met it.

Where vibe coding breaks

Three failure modes show up again and again. The first is the run-it-yourself gap: most AI chats hand you code and leave the setup to you, and as we covered in our piece on whether ChatGPT can build an app, the copy-paste-run loop is where non-developers give up. The second is maintenance: code you never read is code you can't fix, so a vibe-coded project works until the first change the AI can't cleanly make, and then it collapses all at once. The third is security: code nobody reviewed tends to contain flaws nobody looked for. One widely cited 2026 scan of 5,600 live vibe-coded apps (by security firm Escape) found over 2,000 vulnerabilities, including more than 400 exposed secrets — the predictable result of shipping software by vibe alone.

None of this makes vibe coding a fad. It makes it a starting point with known limits.

Vibe coding vs traditional coding

Traditional coding trades time for control: slower to start, but every line is someone's deliberate decision, which is why it still owns anything where correctness, security, or scale genuinely matter. Vibe coding trades control for speed: a working prototype in an afternoon, at the cost of not fully knowing what's inside. The sensible reading isn't that one replaces the other — it's that they've split the world of software into "things worth engineering" and "things worth just having," and the second category is much larger than anyone assumed. The vibe coding tools competing right now are really competing over who serves that second category best.

What comes after vibe coding

Vibe coding's first wave proved people want to describe software and receive it. The wave forming now is about closing the gaps this article listed: tools that run the code instead of just writing it, that hand you the working artifact instead of a project, and that make the AI's decisions explicit enough to review. That's the direction we're building Micro Learner AI in — an AI that runs code in the conversation, sandboxed — but the shift is bigger than any one product. Vibe coding was the question. Execution is the answer.