Why I don’t vibe code
As AI coding assistants and “vibecoding” spread through software development, programmers are sharply divided over whether these tools are a force multiplier or a threat to craft, clarity, and long‑term maintainability. Some see LLMs as indispensable tractors in an industry that once farmed by hand, enabling small teams and startups to ship more with less effort, even if the models are fallible and costly. Others reject or strictly limit their use, arguing that dependence on paid, unreliable services deskills developers, encourages bloated and fragile code, and turns a creative, educational process into prompt‑tuning and review of machine‑generated output.
Reactions to the article’s anti-LLM stance
- Many readers think the critique overgeneralizes from limited experience with weak or free models.
- Others resonate with the discomfort at “paying to think” and the desire to avoid SaaS lock‑in.
- Some appreciate the writing and the focus on process over product, even while disagreeing with the conclusion.
Productivity vs. craft and “hard problems”
- One camp argues LLMs automate “lower-tier” mechanical coding, freeing humans for higher‑level design and more complex systems.
- Another camp feels core, enjoyable parts of engineering are being offloaded, weakening skills and understanding.
- Disagreement over whether recent typical dev work was truly “hard” or mostly framework/config glue.
Spectrum of LLM use (beyond vibecoding)
- Several commenters reject the binary of “no LLMs” vs “agent writes everything.”
- Common “middle ground” uses: autocomplete, one-off snippets, boilerplate, tests, integration glue, while humans review every line.
- Others report using agentic tools heavily but still steering architecture and reviewing output.
Costs, access, and “cheapskate” ethos
- Strong current of people who avoid recurring SaaS fees and prefer FOSS and local tools; LLM subscriptions feel culturally wrong, not just expensive.
- Counterpoint: $20–$100/month is seen as trivial relative to productivity gains, especially for startups.
- Concern that rising and opaque token costs could make experimentation and hobby work less viable.
Code quality, maintainability, and complexity
- Some see LLMs enabling faster delivery of working systems and personal projects that would otherwise be infeasible.
- Others report AI‑written codebases as sprawling, incoherent, and harder to reason about than hand‑written code.
- Fear of becoming dependent on tools to maintain code they generated; worry about “deskilling” and bloated, low‑quality output.
Agentic environments and local models
- Enthusiasts emphasize that results depend heavily on the “harness”: sandboxing, tooling, context strategies, and multi‑agent workflows.
- Local/open‑weight models are seen as a path to reduce cost and lock‑in, though performance and hardware demands are debated.
Analogy and culture wars
- Recurrent analogies compare LLM refusal to refusing cars or tractors; critics call this a “luxury belief,” supporters note external costs.
- Some frame coding-without-LLMs as “trad coding” or a kind of identity/virtue choice, for better or worse.