I Am Not Convinced by Vibe Coding (notashelf.dev)

This piece critiques the recent enthusiasm for "vibe coding"—the practice of using LLMs via prompts, voice, or vague intent to generate and iteratively "feel" code into existence—arguing it’s a seductive but fragile shortcut. The author likens it to a machine that spews millions of car sketches until one looks good: occasionally useful for demos or quick prototypes, but fundamentally noise with sporadic signals. Vibe coding masks hidden complexity (edge cases, side effects, security, data access, performance) behind surface-level success ("it runs, the UI clicks") and encourages aesthetic selection over rigorous engineering. Tools like Copilot, Claude and Gemini can accelerate work but also embed anti-patterns and known security mistakes present in their training corpora. For the AI/ML community the piece is a cautionary note: generative models are assistants, not replacements. Key technical implications include increased technical debt, hallucinations, and fragile systems if generated code isn’t reviewed, tested, and integrated into CI, linting, and observability pipelines. The author recommends restricting vibe coding to throwaway scripts, demos, or controlled internal tooling while preserving human ownership, code legibility, tests, and security audits for production. In short: embrace AI for productivity gains, but don’t abdicate responsibility—understand, review, and instrument the artifacts before shipping.