LLMs Don’t Code—They Compile Your Prompts

The author argues that AI coding is best understood as a compiler workflow: prompts are code and outputs are compilations, not autonomous programming. English is a flawed medium for specifying software, leading to brittle, non-deterministic results, so perceived gains often don’t translate into real productivity. Instead of hype-driven ‘vibe coding,’ we should treat AI as a tool and focus on building better programming languages, compilers, and libraries.

Key Points

AI programming should be modeled as a compiler: prompts are code, outputs are compiled results, and iteration is like recompilation.
English is a poor programming language—imprecise, non-deterministic, and highly non-local—making AI coding workflows brittle.
Perceived productivity gains often mask real slowdowns; hype around ‘vibe coding’ mirrors past tech bubbles.
LLMs add value primarily through search, optimization, and pattern reuse; the human remains the programmer.
AI may eventually replace some programming tasks like compilers and spreadsheets did, but it should be treated as a workflow tool, not an autonomous coder.

Sentiment

Mixed but leaning pragmatic: many find AI coding tools genuinely useful for boilerplate, exploration, and acceleration within disciplined workflows, while a substantial minority agrees with the article’s skepticism about hype, non-determinism, reduced understanding, and long-term talent pipeline risks.

In Agreement

AI is a tool akin to a compiler or junior dev: it needs precise, structured guidance and works best on boilerplate and well-known patterns.
English is a poor, non-deterministic specification language; small prompt changes can unpredictably affect outcomes.
LLMs succeed via pattern reuse, search, and optimization—not true creative understanding or design.
Vibe coding can replace critical thinking with a gambler’s mentality; useful mainly for POCs, ad hoc scripts, and low-stakes tasks.
Perceived productivity can outpace actual throughput or quality; a cited study finds users feel ~20% more productive but are ~19% slower.
AI use can erode understanding of one’s codebase and shift effort to heavy code review and reading; rigorous workflows and oversight are required.
Hype parallels self-driving: geofenced successes aren’t the promised universal autonomy; huge spend has often outpaced realistic capability.
The real opportunity is better languages, compilers, libraries, and tooling design—not burning billions on hype.

Opposed

Many practitioners report large real productivity gains (sometimes 4x) by offloading repetitive coding, which accelerates delivery and reduces procrastination.
LLM autocomplete and chat integrate seamlessly and are the biggest day-to-day wins; even seniors benefit when scoping tasks well.
AI can improve discipline by forcing upfront thinking, decomposition, and clearer specs; it’s an effective personal assistant and research aide.
AI broadens access: non-experts can build useful apps; startups can move faster; the tool reveals weak seniors rather than creating them.
Self-driving is working in multiple cities; calling it ‘obviously nonsense’ ignores operational deployments and potential viable niches.
Non-determinism is acceptable at higher abstraction levels, similar to delegating to humans; the point is working outcomes, not exact internals.
Research is mixed and often small-n; cherry-picking studies is unhelpful—experience and process changes matter.
Architecting smaller, independent components and adding guardrails lets LLMs work better on larger systems.
The ‘anti-AI hype’ narrative dominates online discourse; many real users quietly find AI indispensable despite warts.