The future of software development is software developers

What’s actually hard about software development

• Many agree: the hardest part is turning vague, contradictory human requirements into precise, testable specifications and architectures, not writing syntax.
• Others argue the truly hard part is understanding and evolving large existing codebases and capturing the “why” behind decisions—something code alone (and LLMs) don’t encode well.
• Several note that LLMs help with “what” and “how” but still struggle with “should we do this at all?” and “is this the right abstraction?”

Current capabilities of LLMs for coding

• Positive reports:
  • Strong at boilerplate, CRUD apps, UI ports, glue code, small utilities, and reading/annotating unfamiliar code.
  • Some claim large productivity gains (up to “one person doing work of many” on well-structured, testable tasks).
• Negative reports:
  • Frequent hallucinations, outdated APIs, fragile project plans, and architectural nonsense for novel or intricate domains (fintech, low-level, cryptography, complex simulations).
  • “Vibe-coded” projects often become unmaintainable and require large cleanups.
• Widely observed: they behave like tireless but inconsistent junior developers—sometimes brilliant, sometimes bafflingly wrong.

Trust, safety, and agentic systems

• Concerns parallel self-driving cars: tools work impressively until they fail in ways users can’t predict or quickly recover from.
• Some treat LLMs as another “Swiss cheese” safety layer (linting, test generation, review), not a replacement for human judgment.
• Advocates of modern agentic setups (tool use, compiling, tests, web search) say these sharply reduce hallucinations for many coding tasks; skeptics say variance is still too high for critical systems.

Jobs, skills, and industry structure

• Strong anxiety from some devs, especially newer ones, about being replaced or down-skilled to “AI conductor.”
• Others emphasize that requirements discovery, system design, trade-offs, risk ownership, and talking to stakeholders remain human bottlenecks.
• Expectation that low-skill / repetitive coding and some offshore work are at greatest risk; higher-level problem solving may grow in value.
• Worries that juniors raised on LLMs will never develop deep debugging and design skills, leading to brittle systems.

Long‑term outlook and analogies

• Historical parallels cited: 4GLs, VB/Delphi, low-code, open source, industrial looms, cars vs horses, and crypto.
  • In each case, productivity jumped, more software/things were built, and specialists remained, but many lower-skill roles vanished.
• Debate over whether this wave is “just another tool” or a genuinely different inflection; the thread remains deeply split.