The compiler was supposed to end programming too
Claims that AI coding agents will “end programming” are being weighed against earlier promises that compilers, COBOL, SQL and visual tools would let non‑experts replace software engineers — promises that mostly reshaped the work rather than eliminating it. Commenters contrast deterministic compilers with probabilistic LLMs, arguing that natural-language prompts lack the precise contracts needed to fully automate software design, architecture and debugging, even if they vastly accelerate routine coding. Many expect programming and other white‑collar jobs to shrink and shift toward writing specifications, managing agents and higher‑level product thinking, while others doubt LLMs can ever reliably capture messy real‑world requirements well enough to remove humans from the loop.
Historical context and prior “end of programming” waves
- Commenters recall COBOL, SQL, BASIC, 4GLs, visual tools, and RAD/low‑code as earlier attempts to let “ordinary people” program, none of which eliminated programmers.
- Clarifications that early “automatic coding” meant replacing hand-written machine code, not ending programming itself; compilers were once called “automatic coders.”
- Some argue predictions were partly right: compilers eliminated whole layers of manual work and created more, higher-level programming.
Are LLMs comparable to compilers?
- One side: LLM-assisted coding is just another abstraction layer, akin to moving from assembly to Fortran; the “source of truth” may shift to specs/tests.
- Opposing view: bad analogy; compilers implement precise, deterministic mappings with well-defined contracts, while LLMs are stochastic, fuzzy, and always return some output.
- Debate over determinism: modern compilers are complex and heuristic, but still must preserve program semantics; LLMs do not offer that guarantee.
Specs, natural language, and ambiguity
- Several argue natural language is inherently ambiguous; trying to remove ambiguity ends up recreating code or formal specs.
- Others counter that many business workflows already use natural-language-like rules in tools and agentic workflows.
Quality and role of LLM-generated code
- Some report they hardly “program” anymore; LLMs generate most of their code, with humans focusing on direction and review.
- Others say LLMs over-engineer, make “boneheaded” mistakes, and produce messy, hard-to-maintain systems without careful constraints.
- There’s disagreement on whether top models already outperform most developers; some see that claim as exaggerated or context-dependent.
Jobs, economics, and future workflows
- Strong concern that AI will significantly reduce demand for programmers and other white-collar workers; comparisons to horses being replaced by cars.
- Counterarguments: software is never “finished,” humans still need to define products, handle shifting reality, and resolve underspecified requirements.
- Some foresee 2030s workflows where agents do most coding, tests, and operations, with humans specifying behavior, adjudicating edge cases, and managing “office politics.”
- Others view current enthusiasm as part of a recurring cycle; tools change, but debates over “how software should be built” persist.