Thoughts on the Future of Software Development

Original Article ↗ Hacker News Discussion ↗

Creativity and Capabilities of AI

  • Disagreement over whether AI is meaningfully “creative.”
    • Pro side: “Creativity” = producing interesting or surprising outputs; LLMs clearly do this more than a few years ago.
    • Con side: models remix statistics over prior data, lack true reasoning or invention; surprise can come from randomness too.
  • Some argue current systems can already ask for more info or follow‑up, others say they still fail badly on nonsensical or underspecified questions.

How AI Fits Into Software Development

  • Many see LLMs as powerful assistants: writing boilerplate, tests, refactors, glue code, and documentation; replacing Google for quick lookups and serving as a “rubber duck.”
  • Strong skepticism about full “AI devs”: experiments show they struggle with non‑trivial integration, debugging, edge cases, and evolving large codebases.
  • Domain knowledge, architecture, performance, and trade‑off decisions are repeatedly cited as things current LLMs can’t do reliably.
  • Some imagine multi‑agent systems, tool use, formal methods, and better feedback loops could significantly extend capabilities.

Job Impact and Labor Market

  • Split views:
    • Optimistic: like IDEs and cloud, AI boosts productivity, expands where software is used, and ultimately increases demand for developers.
    • Pessimistic: junior and rote CRUD work get automated first; fewer people hired, lower salaries, especially as big platforms consolidate.
  • Several report huge personal productivity gains, raising concerns about fewer new hires; others say AI‑generated PRs rarely pass serious review.
  • Anxiety that remaining roles become high‑level “prompting, integration, and oversight,” removing the enjoyable parts of coding.

Quality, Maintainability, and Correctness

  • Many worry AI‑written code “works” short‑term but is brittle, incoherent, or architecturally erosive, making long‑term maintenance and refactors harder.
  • LLM tests and proofs are often called low‑value or misleading; correctness is hard to judge and can’t be RLHF’d away.
  • Counterpoint: lots of human code is already bad; some businesses may accept AI‑level quality if it’s cheap and fast.

Liability, Ethics, and Regulation

  • Legal discussions around software liability and “safe harbor” suggest humans will need to review AI output in high‑risk domains.
  • Concern about using AGI as a de facto slave; debates over when machine sentience or copyable minds would deserve rights.
  • Businesses may still push AI despite risk, offloading blame to “the system” until regulation or lawsuits catch up.

Changing Nature of Software and UIs

  • Some expect AI agents to replace many bespoke apps and UIs with natural‑language interactions, potentially even replacing much backend logic.
  • Others doubt prompting will ever be easier or more reliable than writing code for complex, precise systems.