The CoPilot productivity paradox (www.marginalia.nu)

The author reflects on their evolving experience with GitHub’s CoPilot plugin for IntelliJ, ultimately deciding to disable and remove it due to its net negative impact on productivity. While CoPilot excels at accelerating routine, repetitive coding tasks—like generating SQL mappings—it often produces distracting or irrelevant suggestions that interrupt the programmer’s workflow. These unsolicited code changes force constant mental model updates, increasing cognitive load and leading to "mental model thrashing," which exhausts developers and hampers effectiveness, especially when tired. Unlike traditional IDE autocomplete that offers fast, predictable, and muscle-memory-friendly results, CoPilot’s variable delays and inconsistent suggestions break flow and slow the process. The critique emphasizes a fundamental integration challenge: treating large language models (LLMs) as a seamless substitute for classic completions doesn’t fully harness their potential and may diminish programmer control. Instead, LLMs perform better as external, explicit query tools—such as chat interfaces—where context is clear and information exchange is deliberate. For experienced programmers, investing time in mastering their editor or typing skills yields greater productivity gains without sacrificing cognitive bandwidth. However, CoPilot may still offer tangible advantages when dealing with unfamiliar languages or novel tasks. Overall, the "productivity paradox" here highlights that the interaction design between humans and AI assistants in coding environments is as critical as model accuracy for real-world utility.