The Optimal Amount of Slop Is Non-Zero
5 hours ago
- #software development
- #risk management
- #AI-generated code
- The article argues for a non-zero optimal amount of 'slop' (unreviewed AI-generated code), meaning that skipping human code review can be appropriate depending on the context and risk level.
- Rigor in code review should be proportional to risk: high-risk projects (like medical devices or banking systems) require high rigor, while low-risk casual software (personal tools, prototypes) may not need extensive review.
- Software can be categorized using the 'ABC' taxonomy: Acute (mission-critical, high scrutiny), Business (professional, moderate risk), and Casual (limited distribution, low rigor). The appropriate level of review depends on factors like longevity, potential harm, reach, and severity.
- Vibe coding (accepting LLM-generated code without reading it) can be acceptable for casual software with limited impact, but it's risky for business or acute software where defects could cause significant harm or financial loss.
- LLMs should not fully replace human code review because their 'adversarial reviews' lack independence and epistemological soundness—relying solely on LLMs reduces knowledge to mere testimony without reason or empirical verification.
- Examples of unreviewed 'slop' shipped by the author include personal macOS apps, CLI tools, and private apps for family use, while more widely distributed or higher-risk projects received some level of review or scrutiny.
- When deciding whether to review code, consider your optimization goals: speed, business value, learning, or ethics. Unreviewed code can enable faster shipping but may increase technical debt and maintenance costs long-term.
- The author advises assessing risk-rigor alignment for each project: evaluate potential harm, distribution, and longevity before skipping human review to avoid future problems.