GPTs and Feeling Left Behind

Original Article ↗ Hacker News Discussion ↗

Overall split in experiences

  • Thread shows a sharp divide: some report transformative productivity and enjoyment; others find LLMs mostly useless or net‑negative.
  • Many say results are highly variable: “sometimes amazing, sometimes nonsense,” leading to very different narratives depending on which experiences people emphasize.

Where LLMs tend to work well

  • Boilerplate and scaffolding: setting up build systems, configs, CRUD backends, admin panels, unit/e2e tests, simple refactors, repetitive syntax, and frameworks they already understand.
  • IDE‑integrated agents (e.g., Claude Code / Cursor‑style tools) with repo/context access are seen as much more useful than pasting snippets into generic chat.
  • Helpful as a “rubber duck” or junior pair‑programmer: explaining error messages, suggesting approaches, drafting documentation/wikis, and translating requirements into code in familiar stacks.
  • Particularly valuable for:
    • Less‑experienced devs or those returning after years away.
    • Hobby/side projects and “throwaway” or low‑criticality code.
    • Frontend/UI polish and copy, when the user is not a design specialist.

Where LLMs fail or cause harm

  • Complex, niche, or safety‑critical domains (GPU drivers, compilers, robotics, intricate business logic, large legacy systems) often get hallucinated APIs, wrong algorithms, or fragile designs.
  • Introduce subtle bugs (e.g., undocumented params that “kind of work” but corrupt behavior), weak tests, and inconsistent patterns that are hard to spot and maintain.
  • Some open‑source maintainers and enterprise devs report being “drowned in AI slop” and spending more time correcting than coding.

Productivity, skills, and quality

  • Debate over whether they actually speed up experienced devs; one cited study suggests decreased productivity for seniors, prompting arguments over “you’re using it wrong” vs. “perceived gains only.”
  • Concern that reliance on LLMs erodes foundational skills and deep understanding, analogous to skipping “scales” in music; counter‑arguments reference historical shifts to higher‑level languages and GC.
  • Strong disagreement on how much code quality matters outside mission‑critical systems.

Tooling, models, and workflows

  • Outcomes depend heavily on model choice, integration, and workflow: structured prompts, AGENTS/CLAUDE.md files, small targeted edits, multi‑model cross‑review, and frequent context resets are common “success patterns.”
  • Others deride this as “spellcasting” and folk wisdom lacking hard evidence.

FOMO, hype, and psychology

  • Several comments frame LLM coding as slot‑machine‑like: intermittent “jackpots” encourage overuse and magical thinking.
  • FOMO and marketing are seen as driving a lot of blog posts and “gaslighting” experts into doubting their own negative experiences.
  • Some advise ignoring hype, experimenting playfully, and focusing on enduring skills; if/when tools stabilize, they can be learned quickly.