Skip to content

The shift from operating computers to directing them.

Vibe computing.

Updated 2026-05-29

Vibe computing is when an AI does the work on your computer the way you would. It clicks, writes, and uses your apps in real time, together with you. You only have to instruct it. The phrase generalizes Andrej Karpathy's vibe coding,[1] a term he introduced in February 2024 to describe writing software by talking to a model in loose prose, from software development to digital work in general.

The change matters because it relaxes a contract that personal computing has held since the 1960s. For sixty years the user had to translate intent into the exact syntax the machine could parse. That translation was the price of using a computer and the source of most expertise around it. Vibe computing names the moment that price drops, because language models competent enough to read intent from loose prose have made the translation step the machine's job. This is the third such relaxation in the history of personal computing, after the move from the command line to the graphical interface and the move from the desktop to the network.[2]

Three eras of personal computing

The history of personal computing can be told several ways. The story below organizes it by one variable: how much precision the machine demanded from the user. Each era lowered the precision tax and expanded the population for whom the machine became usable.

The command line (1960s through 1980s)

The first interactive systems descended from Project MAC at MIT and the Compatible Time-Sharing System of the early 1960s. Ken Thompson and Dennis Ritchie wrote the first version of Unix at Bell Labs between 1969 and 1971; the design fixed the shape that command-line computing kept for the next twenty years.[3] The user typed at a prompt. A short program parsed the typed string as a command in some specific language. The program executed and printed a result. Composition of small tools through pipes, the durable design lesson of the period, came directly from this constraint.

The precision tax was high. Every command had a name; every argument had a position; every flag had a letter whose meaning lived in a manual page. The user who knew the grammar was extraordinarily productive. The user who did not was helpless. The split was visible at every workstation in every research lab of the era, and it produced the first sociological feature of personal computing: the computer specialist, whose value to an organization was that they had memorized the relevant grammars.

Computing in this era was for the people who would pay the precision tax. The population was small. It included the engineers who built the machines, the programmers who wrote the software, and a thin layer of professional users (system administrators, scientists, financial analysts, librarians) whose work justified the learning curve. Everyone else was nominally able to use computers and in practice did not.

The graphical interface (1980s through today)

The second era started, by convention, with the Xerox PARC research of the 1970s and the commercial products that followed. The 1968 demonstration by Douglas Engelbart at the Fall Joint Computer Conference, later called "the mother of all demos," put nearly every later GUI idea on stage at once: a mouse, hypertext, video conferencing, collaborative editing.[4] Alan Kay's group at PARC built the working system that translated those ideas into the Smalltalk environment in the 1970s. Apple's Lisa shipped in 1983; the Macintosh followed in 1984; Microsoft Windows reached broad use through the late 1980s and 1990s.

The precision tax dropped but did not vanish. The user no longer had to remember the command for "list the files in this directory"; they could open the directory and see the files. They no longer had to know the syntax for "move this file there"; they could drag it. Donald Norman's design-of-everyday-things vocabulary (signifiers, mappings, feedback, constraints) entered the design community in this period because it described what the graphical interface had figured out about lowering the precision tax.[5]

But the tax was still there. The user had to know which menu held which command. The grammar had not disappeared; it had spread across menus, dialogs, and toolbars. The expertise that the command line rewarded as memorization, the graphical interface rewarded as familiarity with each particular application. The web inherited this model and refined it through the 1990s and 2000s. Mobile pushed it further from 2007 onward, lowering the precision tax again at the cost of a smaller screen and a single-task focus. By the late 2010s, the graphical era's design vocabulary was the lingua franca of every consumer product, including the ones that ran on a phone. The population expanded enormously. Application use became a general adult skill.

Vibe computing (2024 onward)

The third era is still in its first decade. It began somewhere between the public release of GPT-3 in 2020[6] and the public availability of ChatGPT in November 2022,[7] depending on which signal counted as the start. The interaction model is not yet fully settled. The structural change is that the precision tax can drop again, because the system can read intent from loose, unstructured input.

The user describes what they want. The system, sitting on top of a competent language model, interprets the request and produces the precise output the user would otherwise have assembled by hand. When the system is uncertain, it asks. When the user wants something more specific than the system guessed, they say so and the system adjusts. The grammar of the system is the user's grammar, and the work of mapping between them moves into the machine. The first widely cited examples were in software development; the pattern has since extended to writing, scheduling, file management, customer support, design, research, and most other digital work.

Each previous transition produced a category of people who had got good at the previous era's grammar and found the new era's lower precision tax unsettling. The Unix wizards of the 1980s viewed the GUI with suspicion. The application power users of the 2000s viewed early voice interfaces with the same skepticism. Some professional designers in the 2020s are voicing the same concern about vibe computing. They are usually wrong in the same way the previous skeptics were wrong: they overestimate the cost of the transition and underestimate the population the new era opens computing to.

· · ·

Origin of the term 'vibe computing'

The word "vibe" entered this conversation through Andrej Karpathy, the former director of AI at Tesla and a founding member of OpenAI. In a short post on February 2nd, 2024, Karpathy described a way of writing software he called vibe coding: he stopped reading the generated code line by line and instead inspected the running result, iterating by talking.[1] The phrase caught on quickly. It named a posture that was already spreading and did not yet have a name.

Later that year, on the Swedish AI podcast Veckans AI, the developer Philip Alm proposed extending Karpathy's framing. In conversation with host Magnus Paues in November 2024, Alm argued that the feature people found striking about vibe coding was not specific to software at all: the user stating intent in plain language rather than translating into the machine's formal syntax was the same thing happening across writing, scheduling, file work, and the rest of digital life. The broader phenomenon needed a broader name. Alm suggested vibe computing. The framing has circulated since in both Swedish and English discussions of AI interaction paradigms.

The lineage matters for two reasons. The first is attribution: any reference to the term should credit Karpathy for the root coinage and Alm for the generalization. The second is that the term has a short, locatable history. It was named twice, by named people, in identifiable moments. That is rare for paradigm vocabulary and makes the concept easier to argue about precisely.

· · ·

Why vibe computing is happening now

A new era of computing needs a capability that was not available in the previous one. The capability behind vibe computing is the modern large language model: a neural network trained on enormous quantities of text to predict the next token, capable enough at scale to read loose input and produce useful output. The trajectory is short and well documented.

The transformer architecture was published in June 2017 by a team at Google Brain.[8] OpenAI's GPT-3, released in May 2020, demonstrated that scaling the same architecture to 175 billion parameters produced surprising in-context behavior, including following instructions it had never been shown.[6] Jared Kaplan and colleagues showed in early 2020 that capabilities scaled predictably with compute and data, a result that drove the next half-decade of investment.[9] By 2022 and 2023, models from Anthropic, Google, Meta, and Mistral joined OpenAI in shipping production-grade systems.

Three more pieces had to land for vibe computing to become a product category and not just a research curiosity. Tool use let models call external functions during generation.[10][11] Agentic loops gave models the structure to plan, act, observe, and revise across many steps.[12][13] Standardized tool interfaces, most notably Anthropic's Model Context Protocol, made the surrounding ecosystem reusable across products.[14] The combination is what makes vibe computing real in 2026: models that read intent, tools that act on the world, loops that compose the two, and protocols that let the whole stack be built once and reused.

The capability is not unbounded. Hallucination, the model's tendency to produce confident output that is not supported by its inputs, remains the defining failure mode.[15] Long contexts have a documented position bias: information in the middle of a long input is often overlooked.[16] And the relationship between model capability and product reliability is not direct; the engineering around the model is doing a great deal of work that is easy to miss.

· · ·

What vibe computing looks like in practice

The clearest way to see what vibe computing actually is in 2026 is to look at the products that exhibit it, grouped roughly by where the activity sits.

Software development

The first and most mature instances. The developer states an intention; the tool reads the codebase, edits files, runs commands, and iterates.

  • Claude Code (Anthropic, 2024) is an agentic coding environment built around the Claude model family. Reads repositories, edits files, runs tests, and reports back through a streaming tool log.[17]
  • Cursor (Anysphere, 2023) is a fork of VS Code with model-driven editing, chat, and agentic mode. The default editor of the vibe-coded era for many developers.[18]
  • Codex (OpenAI, 2025) is the cloud-hosted agentic coding service from OpenAI, succeeding the earlier Codex API of 2021.[19]
  • V0 (Vercel, 2023) turns a natural-language description into a working React component, with the user iterating by talking.[20]
  • Lovable (Lovable AB, 2024) is a browser-based vibe-coding environment aimed at full-stack web apps from a single prompt.

Voice and desktop assistants

Products that take spoken or typed natural-language input and act across whatever applications the user has open.

  • Wispr Flow (Wispr, 2024) is desktop dictation that replaces typing with speech across any application.
  • Incredible (Incredible, 2024) is a voice-first desktop assistant that reads what is on screen and acts across the user's applications.[21]
  • Apple Intelligence (Apple, 2024) is the redesigned Siri and writing tools, built on a combination of on-device and cloud models.[22]

General-purpose assistants

The chat-style products that exposed the underlying capability to a mass audience.

  • ChatGPT (OpenAI, November 2022) is the product that turned the language-model era from a research curiosity into a consumer category.[7]
  • Claude (Anthropic, March 2023) is Anthropic's competing model and product family. Projects, Artifacts, and a strong tool-use story define its 2025 and 2026 generation.[23]
  • Gemini (Google, 2023) is Google's model and product line, with strong multimodal capabilities and live voice.[24]

The list is not exhaustive. It is meant to show what the pattern looks like in practice: products where the user describes intent in loose language, the system supplies the procedure, and the result is precise enough to act on. The common factor is not the model in the middle. It is the willingness of the surrounding product to take loose input and do real work in response.

· · ·

Sources

  1. Andrej Karpathy, vibe coding post on X, February 2024. ↩
  2. Graphical user interface, Wikipedia. ↩
  3. Dennis Ritchie, "The Evolution of the Unix Time-sharing System", Bell Labs. ↩
  4. Douglas Engelbart, "The Mother of All Demos" (1968 demonstration), Doug Engelbart Institute. ↩
  5. Donald Norman, The Design of Everyday Things, 1988. ↩
  6. Tom Brown et al., "Language Models are Few-Shot Learners" (GPT-3), arXiv, 2020. ↩
  7. OpenAI, "Introducing ChatGPT", November 2022. ↩
  8. Ashish Vaswani et al., "Attention is All You Need", arXiv, 2017. ↩
  9. Jared Kaplan et al., "Scaling Laws for Neural Language Models", arXiv, 2020. ↩
  10. Timo Schick et al., "Toolformer", arXiv, 2023. ↩
  11. OpenAI, Function calling guide, 2023. ↩
  12. Shunyu Yao et al., "ReAct: Synergizing Reasoning and Acting in Language Models", arXiv, 2022. ↩
  13. Guanzhi Wang et al., "Voyager: An Open-Ended Embodied Agent", arXiv, 2023. ↩
  14. Anthropic, "Introducing the Model Context Protocol", November 2024. ↩
  15. Ziwei Ji et al., "Survey of Hallucination in Natural Language Generation", arXiv, 2022. ↩
  16. Nelson Liu et al., "Lost in the Middle: How Language Models Use Long Contexts", arXiv, 2023. ↩
  17. Anthropic, "Claude 3.5 Sonnet on SWE-bench", 2024. ↩
  18. Cursor, Anysphere. ↩
  19. OpenAI, "Introducing Codex", 2025. ↩
  20. V0, Vercel. ↩
  21. Incredible. ↩
  22. Apple, "Introducing Apple Intelligence", June 2024. ↩
  23. Claude, Anthropic. ↩
  24. Gemini, Google. ↩

Further reading

Foundational and adjacent writing on interface paradigms, not directly cited above.

© 2026 vibe-computing.aiContact