Developer explains why vibe coding never clicked: cost aversion, experience with Brooks’ accidental vs. essential complexity, and distrust of LLM metacognition on messy data.
Key Takeaways
LLMs address accidental complexity (boilerplate, CLI flags, refactoring) but essential complexity – system design, abstraction choices, edge cases – remains human work.
Fred Brooks’ “No Silver Bullet” framing: tooling improvements reduce accidental complexity, but essential complexity doesn’t compress away.
LLMs can’t interrogate their own training data the way a data journalist interrogates a dataset; the model is its own reality.
Every abstraction is also an occlusion – simplified models miss the messy factors behind the numbers, a risk that compounds when outputs are trusted uncritically.
Friction in coding is a feature: solving hard problems builds developer skill and judgment that can’t be outsourced cheaply.
Hacker News Comment Review
Commenters split sharply: critics note the author tried a non-frontier model on a free trial and quit, making the conclusion weaker than the argument; others say frontier models like Copilot also degrade at month’s edge on cheap tiers.
A subset of commenters resonated with the “narrow window” framing: LLMs are useful for tedious, well-scoped tasks but the interesting, hard work is exactly what developers want to own.
One commenter reported shipping non-trivial features (syntax highlighting for 40 languages, real-time payment API integration) quickly with AI, directly countering the author’s skepticism with concrete project examples.
Notable Comments
@wvenable: Ran cheap fallback models when Copilot quota ran out mid-month – calls them “dumb as rocks,” suggesting many negative opinions stem from tier-limited exposure.
@thangalin: Lists specific shipped projects with AI assistance including a raw Git reader with cloning in ~5 days, countering the article’s skepticism with measurable output.
