Eighteen months ago, "vibe coding" didn't exist. Today it's in Merriam-Webster, it's Collins Dictionary's Word of the Year, and it's already been blamed for an rsync security scare, a wave of leaky startups, and — depending who you ask — either democratizing software development or quietly wrecking it. Here's how one tweet turned into an entire era of how software gets built.
This isn't a definitional piece — we already have one of those (see: What is vibe coding?). This is the timeline: who coined the term, how it spread, what got adopted, and what got broken along the way.
The tweet: February 2, 2025
The term traces back to one person, one date, and one tweet. On February 2, 2025, Andrej Karpathy — a founding member of OpenAI and Tesla's former AI director — posted this on X:
"There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists... I 'Accept All' always, I don't read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension... It's not too bad for throwaway weekend projects, but still quite amusing. I'm building a project or webapp, but it's not really coding — I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works."
This wasn't Karpathy's first attempt at the idea. Back in 2023, he'd already argued that "the hottest new programming language is English" — the observation that large language models had gotten good enough that natural language, not formal syntax, could be the interface to a computer. Vibe coding was that same idea, taken to its logical, slightly chaotic conclusion: not just writing code in English, but not reading the code at all.
From tweet to trend, in about two weeks
Business Insider ran a piece on February 13 calling it "Silicon Valley's next act." Two weeks after the tweet, the New York Times' Kevin Roose published his own experiment — building small apps despite not being a professional coder, and coining a term of his own: "software for one." One of his creations, an app called LunchBox Buddy, fabricated fake reviews for an e-commerce testing scenario — an early preview of the reliability problems that would define the following year of coverage. Cognitive scientist Gary Marcus pushed back, arguing that Roose's success said more about the LLM reproducing existing training data than about any new capability.
By March 5, Ars Technica was already asking whether "the future of software development" would "run on vibes" — quoting developer Simon Willison's now widely-cited line: "Vibe coding your way to a production codebase is clearly risky."
The word gets official: dictionaries, valuations, and even Linus Torvalds
Merriam-Webster added "vibe coding" to its slang & trending list on March 8, 2025 — five weeks after the original tweet. That same month, Y Combinator reported that 25% of startups in its Winter 2025 batch had codebases that were 95% AI-generated. The Economist, riffing on the term, coined "vibe valuation" to describe VC-backed AI startups being priced without reference to normal metrics like annual revenue.
Not everyone embraced the framing. In June 2025, AI researcher Andrew Ng pushed back publicly, arguing the term misleads people into thinking engineers just "go with the vibes" when using AI tools — when in practice, plenty of serious engineering still happens underneath. By July, the Wall Street Journal was reporting that vibe coding had moved past hobbyists and was being adopted by professional engineers for commercial work.
The term's biggest mainstream stamp came on November 6, 2025, when Collins Dictionary named "vibe coding" its Word of the Year — beating out competition from across the language. And in a nice bit of irony, even Linus Torvalds — the person most associated with hand-written, no-compromises code (Linux, git) — was reported in January 2026 to have vibe-coded a visualization tool for one of his side projects using Google Antigravity, writing in the project's README that "the Python visualizer tool has been basically written by vibe-coding."
Then came the reckoning
The same properties that made vibe coding fast — not reading the diff, trusting the model, accepting every suggestion — turned out to have a cost, and by the second half of 2025 the receipts were piling up.
- Security: In May 2025, researchers found that 170 of 1,645 apps generated by Swedish vibe-coding platform Lovable had a vulnerability exposing personal data to anyone. In October, security firm Veracode published a three-year study showing that while LLMs had gotten dramatically better at generating functional code, the security of that code hadn't improved at the same pace — and bigger models weren't meaningfully safer than smaller ones. In December, a security researcher demonstrated a live flaw in the Orchids vibe-coding platform to a BBC reporter.
- Maintainability and quality: A December 2025 CodeRabbit analysis of 470 open-source pull requests found AI-co-authored code carried about 1.7x more "major" issues than human-written code, with security vulnerabilities specifically 2.74x more common. GitClear's longitudinal analysis of 211 million lines of code found refactoring had dropped from 25% of changed lines in 2021 to under 10% by 2024, while copy-pasted code — for the first time in two decades — overtook code that was properly moved and reused.
- The productivity paradox: perhaps the strangest data point of all: in a July 2025 randomized controlled trial, AI research org METR found that experienced developers using AI coding tools were actually 19% slower at real tasks — despite believing beforehand that the tools would make them 24% faster, and still believing afterward that they had been made 20% faster. The vibes, in other words, were not reliable narrators.
- Open source: by January 2026, a multi-university paper titled "Vibe Coding Kills Open Source" argued that the practice was quietly undermining maintainer incentives — since AI tools reduce the direct user engagement that many maintainers rely on for recognition and reputation. The following month, GitHub itself publicly acknowledged the strain, describing an "Eternal September" of low-quality AI-generated contributions overwhelming maintainers — citing curl's decision to end its bug bounty program after a flood of AI-generated (and mostly bogus) security reports, and Ghostty's move to invitation-only contributions.
The clearest flashpoint came in May 2026, when an rsync update broke incremental backups for a number of users. Digging in, they found the maintainer had been committing AI-assisted changes under the byline "tridge and claude" — prompting a GitHub issue bluntly titled "Please Do Not Vibe Fuck Up This Software" that spread to Reddit and beyond. Around the same time, Mario Zechner and Armin Ronacher — engineers behind the Pi coding harness inside the OpenClaw agent system — told the Wall Street Journal they saw a "vibe slop" crisis coming: "You have infrastructure that's falling apart, and you have software that's now very, very buggy compared to before."
What 18 months of history actually tells developers
Karpathy never claimed vibe coding was for production software — his original tweet called it fine "for throwaway weekend projects." Almost every failure mode that made headlines through 2025 and 2026 — the Lovable data exposure, the rsync mess, the open-source strain — happened when that boundary got ignored: when a prototype workflow got pointed at a real codebase, a real user base, or a real production dependency.
That's the exact line we've been writing about on this blog for months: what a "frontend-only" tool hides, why context erodes as a project grows, and why nobody sees the token bill until it's too late. The history of the term backs up the argument better than we could have written it ourselves — the industry ran the experiment in public, at scale, and got a fairly clear answer back.
Vibe coding, as Karpathy described it, is still a genuinely good way to build something in a weekend. Building something that survives past that weekend is a different problem — one Vibe Coder was built to solve.