The author reflects on their experience with code reviews in the context of using large language models (LLMs) like Claude to assist in coding, particularly within the PyTorch project. Earlier this year, LLMs struggled with production-quality code, but recent improvements have led to more reliable outputs. The author argues that code reviews should evolve into a mechanism for human alignment rather than just fixing mechanical issues. For instance, LLMs tend to generate overly defensive code, which can lead to unnecessary complexity. The author emphasizes that the real challenge lies in aligning the human author's understanding of what constitutes overly defensive coding with their own perspective. When submitting LLM-generated pull requests, the focus shifts to conveying higher-level concepts about how the new code functions and what context is necessary for understanding it. The author is willing to iterate on LLM outputs until they reach a clear and correct solution, reflecting a shift in expectations regarding the cost and effort of generating code. The article also addresses concerns about engineers who don't adapt to AI tools. The author believes that skills like code comprehension, communication, and critical thinking are becoming more valuable than raw coding ability. They suggest that junior engineers with strong soft skills will be particularly important in this changing landscape. Today's LLMs lack memory and must rediscover system knowledge from scratch each time. Therefore, the author's vision for code review is to help maintain a shared understanding among team members about the system's intended function. By reorienting the review process toward this collective knowledge, teams can ensure that both human and machine contributions to coding efforts are aligned and productive.