Vibe Coding: easy to say, difficult to survive (gioorgi.com)

"Vibe coding" — an Andrej Karpathy-popularized term for AI-assisted software development — is rapidly spawning startups and features from major model providers, but faces technical and business limits that mirror early Kubernetes-era fragmentation. Practically, tools like OpenAI Codex and GitHub Copilot excel at localized tasks (refactoring, small conversions, CI wiring) and can feel like "having StackOverflow on demand." Yet they stumble on higher-level outputs (accurate READMEs, correct license detection), require senior developer review, and deliver variable latency: productivity gains depend heavily on faster, reliable responses. The author’s hands-on notes show Codex handles concrete edits well but makes predictable semantic mistakes, and Copilot is similar but often needs human vetting. Commercially, standalone "vibe coding" platforms risk thin margins and unsustainable monetization: many resell big-model APIs (OpenAI/Anthropic/Mistral), and most end users still use free tiers (over 90% for ChatGPT), shrinking revenue opportunity. Replit’s recent pricing backlash after Agent 3 and $250M funding is cited as an early warning. Security and IP concerns in sending code to cloud models, plus the modest real-world speedups today, mean startups must either add clear proprietary value or be absorbed by larger players. A likely outcome is consolidation or absorption by model providers unless these services demonstrably boost developer throughput and address privacy/compliance.