What we lost the last time code got cheap
Cheap, abundant AI-generated code is reshaping how software is built, shifting effort from typing implementations to understanding intent, architecture and long‑term maintenance. Commenters describe LLMs as powerful aids for reading and navigating large codebases, generating tests and design docs, yet warn that over-reliance can erode developers’ understanding, motivation and quality control. There is broad agreement that human oversight, clear specifications and captured rationale (design docs, ADRs, commit messages) matter more than ever, even as opinions diverge on whether LLM-written code and prose should be treated with suspicion or as just another tool.
LLMs as Code Readers and Explainers
- Many commenters use LLMs primarily to read code: summarizing repos, explaining unfamiliar modules, or visualizing structure (HTML/SVG diagrams, even videos).
- They’re seen as especially useful for onboarding to large codebases and quickly locating relevant files/lines.
- Some constrain prompts (e.g., ignore domain naming, focus on data flow and math/state transitions) to reduce hallucinations and abstraction bloat.
- Several argue LLMs now make “understanding the codebase” less of a bottleneck than the article suggests.
Quality and Reliability of AI-Generated Code
- Experiences diverge sharply:
- Some report code “almost never” works on first try and needs several iterations, especially for non-trivial tasks.
- Others say recent models often produce working code on first run, especially with clear specs.
- Recommended mitigations: red/green TDD, having agents run tests and linters, and giving explicit goals so the model iterates until tests pass.
Documentation, Intent, and Comments
- Broad agreement that “why” and “why not” decisions are more valuable than “what this line does.”
- Mixed experiences: some see over-documentation with irrelevant change history in comments; others find insightful rationales, especially on bug fixes.
- One major concern: agents rarely capture architectural/product decisions unless explicitly instructed. Proposed solutions:
- decisions.md / ADRs, design docs in-repo, richer commit messages.
- PRs centered on plan/spec files, with code as a derived artifact.
- Some tools default to minimal comments, prompting users to override that behavior.
Impact on Skills, Engagement, and Responsibility
- Worry that relying on LLMs to read/write code can erode individual understanding; critics liken it to “letting your brain atrophy.”
- Counterpoint: humans already wrote poor code without full context; tools can be force multipliers if used thoughtfully.
- Real-world anecdotes: seniors skipping requirements and blaming agents; AI-generated proposals violating org policies. Some managers respond with clear expectations and, if needed, performance processes.
- Suggestions for keeping engineers engaged: involve them in design and intent-setting, framing them as “implementers of functionality” rather than just “writers of code.”
Attitudes Toward AI-Generated Text vs Code
- Recurrent thread: suspicion of “LLM-sounding” prose in the article and elsewhere, and frustration with derivative, unedited AI text.
- Others warn against a “witch hunt” where any punchy or formulaic style is labeled AI, including genuine human writing.
- Noted double standard: many accept AI-authored code but show strong contempt for AI-authored essays and blog posts, even though low-cost generation in both domains shifts the burden onto readers/reviewers.