Andrej Karpathy recently observed a significant shift in coding practices, moving from 80% manual coding to 80% AI-generated code in a matter of weeks. This transition reflects advancements in AI models and coding workflows, particularly in new projects. Boris Cherney, creator of Claude Code, echoed this by reporting that he hasn't manually edited code in over two months, relying entirely on AI for his contributions. A poll of 5,000 developers revealed that 44% write less than 10% of their code manually, signaling a major change in how coding is approached. However, the rise of AI in coding introduces new challenges. Karpathy notes that AI often makes conceptual errors rather than mere syntax mistakes. These include misunderstanding initial requirements, creating overly complex code, failing to clean up, and not questioning vague or contradictory instructions. Developers may find themselves accumulating “comprehension debt,” where they become less familiar with their own code as they rely on AI to generate it. A survey indicated that many developers don’t thoroughly check AI-generated code, leading to potential issues down the line. Interestingly, while teams using AI report a 98% increase in code output, they also face a 91% increase in pull request review times. This paradox highlights that even though code generation becomes faster, the need for review grows, creating new bottlenecks. The article suggests that the 80% threshold for AI-generated code works best in certain contexts, such as personal projects or startups without legacy systems. In these scenarios, the benefits of rapid iteration can outweigh the risks of AI's limitations.