The man who coined the term 'vibe coding' says code written by AI can still be 'awkward' and 'gross'

Andrej Karpathy, the former head of AI at Tesla and a founding member of OpenAI, recently shared insights on the quality of code generated by artificial intelligence during a talk at Sequoia Capital. Despite having coined the term "vibe coding," he expressed reservations about the current state of AI-generated code, highlighting that it often requires human oversight. Karpathy described AI agents as akin to interns, emphasizing the necessity for human developers to maintain control over aesthetics, judgment, and overall quality. He candidly admitted to feeling anxious about the imperfections in AI-generated code, noting that it can be overly complex, contain excessive repetition, and feature awkward abstractions that, while functional, can be quite unappealing. Currently leading the AI-driven educational platform Eureka Labs, Karpathy introduced the concept of vibe coding in a February post on X, characterizing it as a development approach where developers minimally engage with the code themselves. The term gained significant recognition when Collins Dictionary named it the word of the year for 2025, reflecting its impact across both tech and other sectors. During his recent address, he underscored that while AI can assist in various coding tasks, humans must retain authority over major development decisions. He noted that there's no fundamental barrier preventing AI from producing cleaner code; rather, the focus during model training has not adequately addressed this issue. The rise of vibe coding has transformed hiring and reward practices within tech companies and influenced software stock valuations. This trend has catalyzed a surge in startups creating tools for developers, with companies like Lovable, Cursor, and Replit achieving billion-dollar valuations. However, the enthusiasm for AI in coding comes with caution, as professional developers warn against overreliance on AI due to its propensity to generate messy and untested code. Recent security incidents have further fueled concerns, including a situation involving Lovable, which reported a security flaw that exposed user data. The company acknowledged the error and promptly reverted the changes that had inadvertently made chats on public projects accessible. This incident underscores the ongoing challenges associated with integrating AI into software development processes.