The Eternal Sloptember

Original Article ↗ Hacker News Discussion ↗

What LLMs and agents can do today

  • Many commenters say current models can program: they produce compilable, idiomatic code, handle boilerplate, migrations, tests, refactors, and are especially strong in mainstream stacks.
  • Others report success building full applications with agents plus custom “harness engineering,” treating them as powerful but constrained tools.
  • A minority claim they still get better and faster results by hand, especially on highly novel, niche, or low-level work (e.g., unusual SDKs, complex netcode, EMR systems).

Limits, slop, and long‑term maintainability

  • Strong concern that agents overproduce “slop”: large, overcomplicated PRs, hacks, bad abstractions, and subtle bugs that are hard to detect.
  • LLMs are seen as good at syntax and local correctness but weak at architecture, modelling, and choosing the “right problem” to solve.
  • People worry about legacy codebases degenerating into unreadable AI output and about future maintainability once context windows or models change.

Human‑in‑the‑loop, harnesses, and process

  • Repeated theme: tools are valuable only with strict human review, tight scopes, tests, guardrails, and rollback/canary practices.
  • “Harness engineering” (agents, memory, review skills, staged plans, TDD, integration tests) is presented as key to scale their use safely.
  • Critics note that without disciplined review, organizations will rubber‑stamp huge AI PRs and ship regressions.

Impact on developers and organizations

  • Seniors often find LLMs a major force multiplier; some claim ~5–10× individual productivity for standard business work.
  • Others report skill atrophy, burnout, and feeling like they’re starting with a legacy codebase from day one.
  • Debate over whether agentic coding will net‑improve quality or simply accelerate production of mediocre software, especially by low performers.

Broader analogies and societal concerns

  • Comparisons to crypto, industrial automation, Luddites, Eternal September, digital cameras, and cars vs horses.
  • Some argue AI is clearly more useful than crypto; others stress that social and economic choices, not raw capability, determine real impact.
  • Energy use, GPU constraints, climate implications, and concentration of power are recurring worries.