Devin: AI Software Engineer

Original Article ↗ Hacker News Discussion ↗

Hype, demos, and trust

  • Many are impressed by the polish of the demos and the SWE‑Bench numbers vs prior models.
  • Others find the Upwork demo misleading: Devin didn’t actually fulfill the posted requirements (e.g., AWS setup instructions), raising trust concerns about cherry‑picked examples.
  • Several want the generated codebases open‑sourced or more raw, unedited recordings to judge real quality.

Current capabilities vs real-world usefulness

  • SWE‑Bench score (~13.9% issues solved vs ~2–5% prior SOTA) is seen as real progress but far from “production ready engineer.”
  • Key limitation: you still must validate everything; if verifying Devin’s output takes nearly as long as writing it, net value is unclear.
  • People compare Devin to other agentic tools (Sweep, GPT Engineer, Pythagora, etc.) and note similar struggles: coherence over time, architecture, large codebases, subtle bugs.

LLMs as coding tools today

  • Many report strong gains from Copilot/Claude/GPT in:
    • Boilerplate and small module generation
    • Translating between languages/APIs
    • Writing or scaffolding tests
    • Documentation/search replacement and quick summaries
  • But they also report frequent hallucinated APIs, poor library knowledge, shallow reasoning, and sharp drop‑off on larger, nuanced tasks.

Code quality, maintenance, and long-term risk

  • Concerns that agent‑written code will increase technical debt and produce “unfixable spaghetti,” especially on mature codebases.
  • People fear “AI‑generated messes” that still require experienced humans to rescue and maintain.
  • Strong skepticism that complex, safety‑critical systems (e.g., autopilots, banking infrastructure) will be entrusted to current‑gen agents.

Jobs, juniors, and economic implications

  • Large thread on whether tools like Devin will:
    • Replace a big share of developers (especially juniors), or
    • Just raise productivity and shift humans to higher‑level design and “AI management.”
  • Many see juniors as most at risk and worry about how future seniors will be trained if the bottom of the ladder disappears.
  • Redistribution/UBI debates: some argue massive productivity gains could fund safety nets; others point to history and doubt any surplus will be shared.

Historical analogies and timelines

  • Comparisons to:
    • Printing press, tractors, industrial revolution → long‑term net benefit but harsh transitions.
    • Self‑driving car hype → big promises, slower actual deployment.
  • Split between “this is early but inevitable” and “we’re overestimating current AI; likely another AI winter or plateau first.”