So where are all the AI apps?

Original Article ↗ Hacker News Discussion ↗

Where the “AI apps” actually are

  • Many commenters say the big change is in personal and internal software: tiny utilities, dashboards, data wranglers, and one-off tools built for a single user, team, or company.
  • Examples: custom grocery and budgeting apps, label/printing tools, photo-cleanup utilities, hyper‑personal IDE/dashboards, niche music or language‑learning tools, internal admin panels, observability setups, and workflow automations.
  • These are often hacked together in minutes or hours with coding agents and never published, so they don’t show up on PyPI or in public repos.

Why PyPI stats may mislead

  • Many argue PyPI packages are a poor proxy:
    • Most AI‑assisted code is in private repos, corporate codebases, or unpublished scripts.
    • Libraries are harder and more “design heavy” than apps; AI is better at app-level glue than careful API design.
    • AI can reduce need for small helper libraries because it can generate ad‑hoc code on demand.
  • Alternative signals mentioned: rising iOS app submissions, more Steam games (some disclosing gen‑AI use), and GitHub Octoverse data showing increased private contributions.

Productivity gains and their limits

  • Broad agreement: AI makes getting to a working prototype dramatically faster (often 5–10x).
  • The “last 10–20%” — robustness, edge cases, security, performance, deployment, UX, and ongoing maintenance — remains hard and still requires traditional engineering skill.
  • Several report that AI speeds infra-as-code, debugging, and repetitive tasks, but whole‑product timelines aren’t 100x shorter.

Quality, maintainability, and “vibe-coded slop”

  • Many describe “vibe‑coded” apps as buggy, insecure, and architecturally fragile; fun demos, but not production‑grade.
  • Experienced devs say AI can amplify bad patterns and create “comprehension debt” — large codebases no one fully understands.
  • Open‑source maintainers are wary of low‑effort AI PRs and “AI slop,” which may discourage publishing AI-assisted code.

Business, distribution, and attention bottlenecks

  • Coding was never the main bottleneck for successful products; marketing, sales, data access, app store discoverability, and support still dominate.
  • App stores are reportedly flooded with mediocre AI‑generated apps; very few gain users or revenue.
  • Some predict more “build vs buy” swinging toward build (cheap custom tools), others note SaaS still valuable for reliability and operations.

Shifts in how software is created and shared

  • Several claim we’re entering a “personal/disposable software” era: lots of throwaway, hyper‑specific tools that exist only on one machine.
  • Some say they now use fewer third‑party libraries and open‑source tools, preferring to have an LLM generate bespoke code instead.
  • Skeptics argue the overall economic/productivity impact is still unproven; enthusiasts counter that we’re early and effects are mostly invisible in traditional metrics.

Open questions and disagreements

  • Is imagination/ideation now the true bottleneck, with AI mainly removing execution cost?
  • Will agents ever handle the “last mile” (requirements, design trade‑offs, human communication, and production readiness)?
  • How much of the current wave is durable productivity vs hype-driven “slopware”? Opinions diverge sharply in the thread.