“Vibe Coding” vs. Reality

Original Article ↗ Hacker News Discussion ↗

Productivity Gains vs. Replacement

  • Many report meaningful speedups (2–10x) from tools like Cursor/Claude for boilerplate, small bugfixes, tests, and CRUD-style code.
  • Others say LLMs slow them down because every line must be checked, so they only trust them for tiny, isolated tasks.
  • Consensus: refusal to use AI will disadvantage developers, but current tools do not fully replace a competent engineer, especially on non‑trivial systems.

Team Size, Jobs, and Economics

  • One commenter claims a product team was reduced from 9 to 2 devs with AI, triggering strong skepticism, accusations of exaggeration, and debates over prior over‑hiring.
  • Some argue faster devs → fewer jobs; others invoke expanding demand and Jevons paradox: cheaper software often creates more software and thus more work.
  • Concern remains that non‑tech companies will use AI mainly to cut headcount rather than expand scope.

What “Vibe Coding” Is (and Isn’t)

  • Original meaning: deliberately not reading AI‑generated code, just iterating on natural‑language instructions and “fix this bug” until it runs—intended for toy or throwaway projects.
  • Several complain the term has been hijacked by influencers and marketers to imply serious, production‑grade development without human review.

Quality, Maintainability, and the 80/20 Problem

  • Strong theme: LLMs can get you “80% of a prototype,” but the last 20% (edge cases, security, correctness, performance) is ~80–90% of the real work.
  • People liken LLM output to messy outsourced code or Excel macro systems that later become expensive technical debt to untangle.
  • Multiple anecdotes of junior/outsourced devs over‑using LLMs, shipping unread, duplicated, brittle code, and being let go.

Roles, Skills, and Future Trajectory

  • LLMs are compared to a fast, tireless but context‑ignorant junior dev or consultant: great on generic patterns, poor on domain‑specific constraints.
  • Some foresee engineers evolving into architects/product or domain experts who orchestrate AI; others doubt LLMs can ever truly “understand” systems.
  • Broad agreement that hype (10–100x, “everyone can code now”) far exceeds observed reality, but many expect continued, possibly rapid, improvement.