Vibe CodingVibe Coding

In February 2025, Andrej Karpathy, OpenAI co-founder and former Tesla AI director, described in a post on X a style of programming he called vibe coding: fully giving in to the vibes, embracing exponentials, and forgetting that the code even exists. You describe what you want in plain language, the AI writes the code, and when something breaks you paste the error back and ask it to fix it.

The confession that landed hardest was that he barely read the diffs. He would accept changes wholesale, and the code would grow beyond his usual comprehension of it. For a programmer of Karpathy's stature to admit this, half sheepishly and half enthusiastically, gave everyone else permission to say they worked the same way.

The name filled a vacuum. AI coding assistants had evolved from autocomplete into agents that could scaffold and iterate on entire applications, and a generation of users, many with no formal programming background, were already shipping working software this way. The term spread through the industry within weeks, and Collins Dictionary named 'vibe coding' its word of the year for 2025.

The backlash was immediate and often fair. Critics pointed out that code nobody reads is code nobody can secure, debug, or maintain, and stories circulated of vibe-coded apps leaking credentials or collapsing under real users. A distinction emerged between vibe coding as throwaway prototyping, which almost everyone endorsed, and vibe coding as production engineering, which almost nobody did.

The deeper shift outlasted the discourse. The programmer's scarce skill began migrating from writing code to specifying intent, reviewing output, and designing the checks, tests, evals, and guardrails, that let you trust software you did not write line by line. Karpathy had earlier quipped that the hottest new programming language is English; vibe coding was that idea arriving on schedule.