A coder considers the waning days of the craft
Programmers are grappling with what AI coding tools like GPT‑4 mean for their craft, careers, and the future of software development. Many see LLMs as powerful accelerators for boilerplate and “glue code” that lower the barrier to entry and raise productivity, but argue they still fail on complex, novel, or deeply contextual problems and require experienced engineers to supervise and debug their output. A recurring concern is less about technical capability and more about economics: the risk that management will use perceived efficiency gains to shrink teams and wages, widening inequality even as the overall demand for software—and for people who can reason about systems—continues to grow.
Perceived capabilities of LLMs today
- Widely seen as strong at: boilerplate, “CRUD + glue” code, scripts, bash/SQL/regex, translations, documentation, explaining APIs, and acting as a rubber duck.
- Often fails on: complex multi-module systems, niche domains (industrial embedded, AAA game systems, IAM, security), non‑standard APIs, difficult debugging, and larger refactors.
- Hallucinations and subtle bugs are common; it frequently invents APIs or omits edge cases. Many say review effort can match or exceed writing the code yourself.
- GPT‑4 is regarded as vastly better than 3.5, but still unreliable as an autonomous programmer.
Prospects for future AI progress
- Optimists expect rapid, possibly exponential gains (pointing to 2017–2023 progress), foresee mid‑level dev work largely automated within a decade, and speculate about self‑play and compiler‑in‑the‑loop training.
- Skeptics cite physical limits (compute, power), LLM architecture constraints, data quality, and historical over‑promises (self‑driving, flying cars); expect long‑running plateaus or only incremental improvements.
Impact on software jobs and career paths
- Broad agreement that junior and “glue” roles are most exposed; seniors doing architecture, hard debugging, and cross‑team coordination seem safer for now.
- Concern that if juniors are replaced, future seniors will have nowhere to learn; skills pipeline may hollow out.
- Some expect fewer but more productive, better‑paid experts; others anticipate downward pressure on wages and “cushy” tech jobs, especially in mid‑tier enterprise work.
Code quality, maintenance, and security
- Fear of a flood of mediocre, copy‑pasted AI code: more legacy to maintain, harder git archaeology, security issues reminiscent of 2000s-era bugs.
- Counterpoint: with tests, linters, and human review, LLMs can offload low‑value work (tests, mappings, wrappers) while humans own design and correctness.
The “craft” of programming and learning
- Many argue the craft is shifting, not dying: from memorizing arcana to problem decomposition, requirement clarification, and system design.
- Others feel genuine loss: AI removes the “puzzle” aspect and satisfaction of hard‑won solutions, especially for those who enjoy low‑level details.
- Worry that learners leaning on LLMs will skip the frustration that builds deep understanding and resilience.
Economic and social implications
- Strong thread on inequality: automation gains likely accrue to AI owners unless political choices change; parallels drawn to past automation and offshoring.
- Some see potential for cheaper, ubiquitous software and “post‑scarcity” for many tasks; others fear “technofeudalism” and shrinking paths to a middle‑class life.