Source note
When Code Is Cheap, Does Quality Still Matter?
Summary
LLM-generated code lowers the cost of writing code, but the article argues that code quality still matters because humans still own changes, incidents, and maintenance.
Problem
- Cheap generated code can increase volume faster than teams can understand it.
- Polished AI output can hide poor boundaries, duplicate logic, vague naming, and behavior that is hard to remove.
- The main cost in software work is understanding, change, review, debugging, and operations, not typing.
Approach
- Treat an LLM as a fast assistant inside a human-owned system.
- Keep AI changes narrow, with existing repo patterns, explicit boundaries, tests, tool permissions, and success criteria.
- Ask the model for constrained implementation work, mechanical rewrites, migrations, tests, and small functions instead of open-ended feature design.
- Judge generated code by whether engineers can explain, review, refactor, delete, and operate it.
Results
- The article reports 0 experiments, 0 datasets, 0 baselines, and 0 benchmark metrics.
- It claims LLMs reduce code production cost, while understanding, review, debugging, and operations stay expensive.
- It claims weak engineers can now produce more code than they understand, which raises review and maintenance risk.
- It claims strong AI coding stories use constraints such as narrow diffs, typed contracts, reproducible commands, strong tests, equivalence checks, and review checkpoints.
- It defines ownership as the quality bar: if a team cannot safely refactor or delete generated code, the code is already expensive.