Vibe computing is a generalization of vibe coding. The two terms came in that order, and the second describes what happens when the pattern of the first leaves the code editor and starts applying to everything else a person does at a computer.
The vibe coding inheritance
In early 2024, Andrej Karpathy coined vibe coding on X. He was naming a habit that had quietly become common: rather than authoring each line of code, a developer states the desired behavior in plain English, accepts the AI assistant’s output, runs it, sees what happens, and iterates. The work of translating intent into a programming language gets delegated to the model. The developer’s job becomes describing and validating, not typing.
The term spread fast because it named something practitioners were already doing. Tools like Cursor, Claude Code, GitHub Copilot, and a handful of agentic coding services had been shipping the pattern in production for some time. “Vibe coding” just gave it a label.
The generalization
The property that drew attention to vibe coding was not specific to code. What practitioners noticed was that the user no longer translated their intent into the machine’s formal syntax. Software development is only one kind of work where that translation is the bottleneck. Writing an email, reorganizing files, navigating a web page, comparing prices across tabs, scheduling a meeting, filling forms; each follows the same shape. The user knows what they want. The machine knows how to do it. Until recently, the user also had to know how, because they were the one operating the controls.
Vibe computing is the term proposed for the broader pattern by Philip Alm, in conversation on the Swedish AI podcast Veckans AI with host Magnus Paues. The proposition was simple: if vibe coding is “describe the thing instead of writing it” applied to software, vibe computing is the same pattern applied to everything else. Paues followed up with a public post quoting the framing, which is where the term first circulated in writing.
What it looks like in practice
A vibe computing system operates the real computer. It clicks, types, drags, scrolls, and uses applications the way a person would. The user states intent, by voice or by text, and watches the system carry it out, intervening when it goes off course. The interaction is realtime and collaborative, not asynchronous and opaque.
The current generation of these systems is built on large language models with vision and tool-use, so they can see what is on screen and act on it. Products in this shape include Anthropic’s Claude Computer Use, OpenAI’s Codex desktop, Microsoft Copilot, Manus, Incredible, and Clicky, among others. The directory covers the rest. Different products take different positions on voice vs text, screen-reading vs accessibility API, full desktop vs browser-only scope.
What it is not
Vibe computing is a claim about the interface, not the intelligence. A system can be addressable in natural language for a narrow set of tasks and still count. A system can be very capable in some sense and still not count if it can’t act on the user’s computer. The term names a shape of interaction, not a level of capability, though, as discussed in a separate piece, the two converge in practice.
It is descriptive rather than prescriptive. It names a pattern that exists; it does not assert the pattern is preferable to direct operation, or that delegating control to a probabilistic system is wise. People will continue to argue about that. The point of the term is to have something more specific to argue about than “AI.”