Even the Inventor of 'Vibe Coding' Says Vibe Coding Can't Cut It (gizmodo.com)

OpenAI cofounder Andrej Karpathy this week open‑sourced nanochat, a “minimal, from‑scratch, full‑stack training/inference pipeline” that lets anyone spin up a ChatGPT‑style chatbot in hours for roughly $100. Contrary to his earlier celebration of “vibe coding,” Karpathy says he hand‑wrote the ~8,000 lines of “quite clean code” (with only tab autocomplete) because attempts to delegate the work to agents like Claude or Codex “didn’t work well enough.” Nanochat is pitched as a compact, end‑to‑end example of how to train, serve and interact with an LLM, but its creation underscores that sophisticated system engineering still often requires expert human labor. The release is a reminder to the AI/ML community that generative coding tools remain imperfect for complex, production‑grade systems: surveys from Fastly (95% of devs spend extra time fixing AI‑generated code) and research from METR showing slower task completion reinforce that AI can introduce more maintenance overhead than it saves. Nanochat both democratizes access to a working LLM pipeline and highlights why “vibe coding” is better suited to throwaway experiments than to reliable model training, inference, security and performance engineering—areas that still demand experienced engineers or improved tooling/hybrid workflows.