Two Kinds of Code Review (haydenbleasel.com)

Anthropic has significantly transformed its code review process, skyrocketing from merging about 2,500 code changes per day in May to over 30,000 by September. This shift underscores a pivotal moment in the AI/ML community, as developers report that AI models like Claude now outperform them in code writing efficiency, leading engineers to focus more on strategic decision-making and intent evaluation rather than exhaustive line-by-line reviews. The merging of advanced AI in code generation means that while software engineers still exercise oversight, their roles are evolving towards a tech lead position that emphasizes design direction and high-level judgments. The article divides code review into two distinct types: objective and subjective. Objective reviews—where AI excels—focus on functionality, identifying bugs and vulnerabilities without fatigue, as demonstrated by Anthropic’s use of front-line AI models. In contrast, subjective reviews require human judgment to gauge whether the code aligns with project goals and best practices. This division highlights the expanding role of AI as a tool that aids but does not replace the critical, interpretative aspects of software engineering—namely, deciding not just if code works, but if it is the right code, particularly in contexts where the cost of errors can have severe consequences. As AI continues to advance, integrating these models into the development cycle is reshaping how code is reviewed, allowing for more efficient and effective software development.