In recent weeks, Andrej Karpathy has shifted his coding process significantly, moving from a manual coding approach to relying heavily on language models (LLMs). He notes that his workflow has transformed from about 80% manual coding to 80% agent-assisted coding in just a month. While this change feels a bit uncomfortable, the efficiency gained from programming in English, directing LLMs to generate code, is undeniable. He believes many engineers are experiencing similar shifts, though the general public remains largely unaware of this transition. Karpathy highlights the limitations of current LLMs, emphasizing that they still make mistakes, often subtle conceptual errors rather than simple syntax issues. They can make incorrect assumptions without verifying them and struggle with clarifying questions or managing inconsistencies. While he acknowledges the improvements made by LLMs, he cautions that they can overcomplicate code and don't clean up after themselves. Despite these challenges, he finds it hard to imagine returning to traditional coding methods. He enjoys the stamina and persistence of LLMs, noting that they don't tire or get discouraged, allowing them to tackle problems longer than a human might. He also discusses the potential impact of LLMs on productivity, suggesting that they may not just speed up coding but expand what can be accomplished. By framing tasks with clear success criteria, LLMs excel in goal-driven programming. The experience of coding has become more enjoyable for him, with tedious tasks being handled by the model, freeing him to focus on creative aspects. However, he acknowledges a downside: as reliance on LLMs grows, he fears he may become less adept at manual coding, leading to a phenomenon he refers to as “comprehension debt.” He raises questions about the future of engineering roles and whether generalists will start to outperform specialists in this new landscape.