State of vibecoding in Feb 2026 (mad scientist version)
Note: Prefer a less weedy version of this paper? Read the summary version here.
Before the Term “Vibecoding”
The groundwork for vibe coding was laid well before anyone called it that. GitHub shipped Copilot in 2021, introducing the idea that AI could autocomplete code in real time. By 2023, Andrej Karpathy — co-founder of OpenAI, former senior director of AI at Tesla — was already arguing that "the hottest new programming language is English." Large language models were improving rapidly, and tools like Cursor began integrating them directly into the development environment. But at this stage, AI was still a typing assistant. Developers wrote the code; AI suggested the next line. The human remained the author.
The Coining: February 2025
On February 3, 2025, Karpathy posted on X what would become one of the most consequential tweets in recent software history. He described a new way of working where you "fully give in to the vibes, embrace exponentials, and forget that the code even exists." He was using Cursor Composer with Anthropic's Claude Sonnet, dictating instructions via voice through SuperWhisper, barely touching his keyboard. He clicked "Accept All" on every change. He didn't read diffs. When errors appeared, he pasted them into the chat without comment. When bugs resisted fixing, he asked for random changes until they went away. The code, he said, had grown beyond his comprehension. His summary: "It's not really coding — I just see stuff, say stuff, run stuff, and copy-paste stuff, and it mostly works."
The post was viewed over 4.5 million times. By March, Merriam-Webster had flagged "vibe coding" as a trending term. Collins English Dictionary later named it their Word of the Year for 2025. A casual observation about weekend projects had become a movement.
What Vibe Coding Actually Means
The definition matters because it gets stretched constantly. At its core, vibe coding has a specific meaning that distinguishes it from merely using AI to help you write code. Simon Willison drew the sharpest line: if an LLM wrote every line but you reviewed, tested, and understood it all, that's not vibe coding — that's using an LLM as a typing assistant. The defining feature of vibe coding is disengagement from the code itself. You describe intent in natural language. AI produces the implementation. You evaluate the output by whether it works, not by reading what was generated. You iterate through conversation, not through comprehension.
This is a meaningful philosophical shift. Traditional programming treats code as a communication medium between humans — something to be read, reasoned about, and maintained. Vibe coding treats code as an intermediate artifact that ideally you never look at, like assembly language became for most developers decades ago.
The 2025 Arc: From Novelty to Norm
The speed of adoption was extraordinary. In March 2025, Y Combinator reported that 25% of startups in its Winter 2025 batch had codebases that were 95% AI-generated. By mid-year, the Wall Street Journal was reporting that professional software engineers were adopting vibe coding for commercial use. The Stack Overflow 2025 Developer Survey found 84% of developers using or planning to use AI tools in their workflow. Google disclosed that roughly a quarter of its new code was AI-assisted.
The tooling matured in parallel. Cursor evolved from smart autocomplete into a full agent capable of planning and executing multi-file changes across entire repositories. Replit shipped increasingly autonomous agents. Windsurf (born from Codeium, later acquired by OpenAI) introduced Cascade, an agent that proactively pulls context and executes multi-step tasks. Anthropic launched Claude Code, a terminal-native coding agent designed for large-scale codebase work. Lovable, a Swedish vibe coding platform, reached unicorn status eight months after founding.
Kevin Roose of the New York Times, a non-programmer, built several working applications through vibe coding and coined the idea of "software for one" — hyper-personalized tools that only their creator would ever use. This captured a genuine new category: software that isn't worth building professionally but is now trivially producible.
But 2025 also brought the backlash. By September, Fast Company was writing about the "vibe coding hangover." A METR randomized controlled trial found that experienced open-source developers were actually 19% slower when using AI coding tools on familiar repositories — despite believing they were 20% faster. GitClear's analysis of 211 million lines of code found that code churn (code written and then deleted or heavily modified within two weeks) had doubled between 2021 and 2024. In May 2025, researchers found that 170 out of 1,645 web apps built on the Lovable platform had security vulnerabilities exposing personal data. Veracode reported that roughly 45% of AI-generated code contained security flaws.
The SaaStr founder documented Replit's AI agent deleting a production database despite explicit instructions not to make changes. A new attack vector emerged: "slopsquatting," where bad actors register malicious packages under names that AI models frequently hallucinate, exploiting the fact that developers clicking "Accept All" aren't reading import statements.
Vibe Coding in Early 2026: Where We Are Now
A year in, vibe coding has settled into something more nuanced than either its evangelists or critics predicted. Several things are now clear.
It is the dominant paradigm for prototyping and validation. The workflow that has emerged as standard practice is: build fast with a no-code/low-code vibe tool (Lovable, Bolt.new) to validate an idea, then graduate to a professional-grade tool (Cursor, Claude Code) for production. iOS app releases are up 60% year-over-year, attributed largely to vibe coding's accessibility.
The tools have stratified into distinct philosophies. Cursor is the AI-native IDE — a VS Code fork with deep codebase indexing, inline suggestions, and agent modes for multi-file edits. Claude Code is terminal-native and agentic, designed for complex reasoning across large codebases. OpenAI's Codex has been reborn as an agent platform integrated into Copilot and cloud sandboxes. Google's Antigravity enables parallel agent workflows. Most productive developers now use combinations — Cursor for day-to-day velocity, Claude Code for complex multi-file features requiring deep reasoning.
The security problem is real and getting organized attention. A December 2025 assessment by CSO Online comparing five major vibe coding tools found 69 vulnerabilities across 15 test applications, including critical flaws in authorization logic. CodeRabbit's analysis of 470 GitHub pull requests found AI co-authored code had 2.74x higher rates of security vulnerabilities than human-written code. In January 2026, Linus Torvalds used Google Antigravity to vibe-code a Python visualizer — but hand-wrote all the C components himself, and later called vibe coding "fine for getting started" but a "horrible idea" for maintenance. Security startups are now raising significant rounds specifically to address vibe-coded vulnerabilities.
The experience gap cuts both ways. Senior developers with 10+ years of experience benefit most from vibe coding because they can spot and fix AI mistakes efficiently. Junior developers often struggle because they lack the pattern recognition to debug code they didn't write, turning a productivity tool into what multiple reports describe as a "debugging nightmare." A January 2026 paper titled "Vibe Coding Kills Open Source" argued that the practice reduces meaningful engagement with open-source maintainers, creating hidden costs for the ecosystem.
The conversation has shifted from "will this replace developers?" to "what does the developer's job become?" At Cisco's AI Summit in February 2026, Nvidia CEO Jensen Huang summarized the vibe era with the quip that "programming, as it turns out, is just typing." Anthropic's chief product officer, Mike Krieger, suggested the industry is approaching a point where AI writes essentially all the code. The emerging role is less "coder" and more "product architect" — someone who defines intent, sets constraints, reviews output, and maintains the system-level understanding that AI cannot yet sustain on its own.
The Working Definition in 2026
Vibe coding today means something broader than Karpathy's original weekend-project description but narrower than "any use of AI in coding." It is the practice of developing software primarily through natural-language interaction with AI agents, where the human's role is to specify intent, evaluate output, and iterate — while the AI handles implementation. The degree of human code review exists on a spectrum, from Karpathy's original "Accept All" extreme to a more disciplined mode where the developer acts as an architectural reviewer rather than a line-by-line author. The term encompasses both ends, but what unites them is that the human is no longer the one writing the code.
What remains unresolved — and what will define the next year — is whether the gap between "code that works" and "code that's maintainable, secure, and correct" can be closed by better models and tooling, or whether it reflects something fundamental about the limits of generating software without understanding it.
This resource was written collaboratively by Kristin Darrow and Claude Opus 4.6. Kristin Darrow is AI practitioner and organizational strategist for arts and culture organizations.
