The human-in-the-loop is tired
使用大型语言模型进行编程带来既有实际效用又伴随高度不稳定性的双重体验。虽然这些工具能加速开发，但也改变了工作的本质：过去通过手动编码获得的那种小而满足的多巴胺回报，往往被持续监管带来的沉重认知负担取代。开发者大量时间用于澄清和重新定义任务，却常常只能发现那些缺乏人类逻辑的一些莫名其妙的错误。结果是一种疲惫感：在机器产出大量大体正确却常有瑕疵的输出时，必须不断维持高层意图。
当前的状况把开发者困在工作强度不断上升的循环里：一方面可以同时启动多个项目，另一方面人类注意力仍是不可并行、有限的资源，二者难以调和。这引发了激励机制的问题。传统编码通过逻辑与掌控带来即时满足，而以 LLM 辅助的工作流则更强调审查与监督。这个转变常让人感到孤立，因为软件开发中那些自然的协作时刻，往往被无休止的提示操作取代，许多程序员在适应新范式的过程中感到孤独。
这种颠覆类似于二〇〇〇年代末向响应式网页设计的转变：当时标准的改变对那些精通固定宽度布局的开发者来说，一度像生存威胁。正如那一时期要求设计师从像素级控制转向对系统的理解，AI 革命也要求工程关注点发生转移。工艺与专业性并未过时，但展示它们所需的具体技能在变化。如今最有价值的品质包括架构成熟度、细腻的判断力，以及分辨基本原则与陈旧习惯的能力。
归根结底，软件开发的瓶颈从来不是写代码本身，而是人类注意力与工程视野的运用。随着 AI 接管编程中机械性的部分，人类能力显现为真正的稀缺资源。行业正在经历根本性的重塑，工程师仍是日益复杂系统的质量把关者。整个行业也在为自己的激励机制进行调试，尽管被压垮的感受普遍存在，但这是一种共同经历，标志着这门技艺一次艰难却必要的进化。
Programming with large language models offers a dual experience of genuine utility and significant destabilization. While these tools can accelerate development, they also shift the nature of the work, often replacing the small, satisfying dopamine hits of manual coding with the exhausting cognitive load of constant supervision. Developers now spend significant time clarifying and re-specifying tasks, only to catch inexplicable errors that lack human coherence. This creates a state of fatigue that stems from needing to maintain high-level intent while the machine generates high volumes of mostly correct, but often flawed, output.
The current landscape traps developers in a cycle of increased intensity, where the ability to start multiple projects simultaneously is balanced against the reality that human attention remains a non-parallelizable, finite resource. This shift creates a reward function problem. Whereas traditional coding provided immediate gratification through logic and control, modern LLM-assisted workflows prioritize review and oversight. This transition often feels solitary, as the natural collaborative moments of software development are frequently replaced by endless prompting, leaving many programmers feeling isolated in their struggle to adapt to the new paradigm.
This moment of disruption mirrors the transition to responsive web design in the late 2000s, where a shift in standards initially felt like an existential threat to developers who had mastered fixed-width layouts. Just as that era required designers to evolve their understanding of systems rather than obsess over pixel-level control, the AI revolution demands a shift in engineering focus. Craft and expertise are not becoming obsolete, but the specific skills required to demonstrate them are changing. Today, the most valuable traits include architectural maturity, nuanced judgment, and the ability to distinguish between essential principles and outdated habits.
Ultimately, the bottleneck in software development was never the act of writing code, but the application of human attention and engineering vision. As AI takes over the mechanical parts of programming, human capacities are revealed as the true scarce resource. While the profession is undergoing a fundamental reshaping, the role of the engineer remains vital as the quality gate for increasingly complex systems. The industry is currently debugging its own reward functions, and while the feeling of being overwhelmed is widespread, it is a shared experience that marks a difficult but necessary evolution in the craft.
162 comments • Comments Link
• 将大型语言模型（LLM）引入编程后，开发者的体验从以解决问题和追求技艺为核心的旅程，变成了持续的审查与监督，因而丧失了手工编码带来的内在多巴胺回报。
• 编程思维分化为两类：一类注重技艺和过程，对 AI 辅助的工作流感到疏离；另一类只看重结果，重视高效交付，尽管个人主体感因此减弱。
• 开发者常感疲惫，源自验证 LLM 输出所需的高认知负荷：审查生成的散文或晦涩代码，往往比自己编写或调试更耗费精神。
• 一些开发者通过将 LLM 视为受限的代码生成器并采用细粒度、迭代式的规划（而非随性编码）成功适应，这有助于保持对代码库的控制感和归属感。
• 当前环境产生了"human on the hook"的动态：虽然开发者往往失去了通过手写代码获得的深刻实现理解，但仍需对错误承担全部责任。
• 受市场 FOMO 驱动的盲目赶进，使许多人陷入不可持续的"crunch"实践，这更像是游戏行业常见的有毒劳动文化，而非将效率提升用于改善工作生活平衡。
• 无论是代码、文档还是文章，AI 生成内容越来越被视为"粗糙产物"，导致创作者产生意义危机——他们发现很难为那些容易被机器复制的成果感到自豪。
• 精通工具与抽象的高级开发者通常对所谓的"生产力提升"持怀疑态度，认为对于具备深厚领域知识的人来说，编写代码从来不是主要瓶颈。
• AI 化写作风格的普遍存在，即便出现在真实的个人散文中，也引发读者的愤世嫉俗，难以分辨人类洞见与与企业立场一致的生成内容，这使在线讨论变得更复杂。
• 对一些资深开发者而言，AI 是强大的加速器，能让他们重拾初学者时那种快速创造的"神奇"感受，这表明满意度的差异可能更多与个人动机相关，而非纯粹技术能力。
这次讨论反映出生成式 AI 带来效率提升与开发者满意度下降之间的深层张力。部分人觉得 AI 让他们能更像架构师，专注更高层次的问题，但也有很多人感到"技能退化"和身份丧失——编程中那种沉思且令人满足的体验，被管理机器生成输出的繁重工作所取代。对 AI 创作内容的质疑进一步加剧了困境：社区在努力维系真实的人际联系，而在这个时代，主动创作与自动生产之间的界限正在迅速模糊。归根结底，这场辩论关乎软件开发究竟应被视作一种艺术与工艺，还是仅仅作为向终端用户交付功能性产品的工具性手段。 • The integration of LLMs into programming has shifted the developer experience from a journey of problem-solving and craftsmanship to a state of constant review and supervision, resulting in a loss of the intrinsic dopamine rewards associated with manual coding.
• Programming has diverged into two distinct mindsets: those focused on the craft and process, who feel alienated by AI-assisted workflows, and those focused purely on the end result, who value the ability to ship products efficiently despite the diminished sense of personal agency.
• A feeling of exhaustion often stems from the high cognitive load required to verify LLM output, as reviewing generated prose or opaque code requires more mental energy than writing or debugging one's own work.
• Some developers have successfully adapted by treating LLMs as highly constrained code generators, emphasizing granular, iterative planning rather than "vibe coding," which helps maintain both control and a connection to the codebase.
• The current AI-driven environment has introduced a "human on the hook" dynamic, where the developer remains solely responsible for errors despite losing the deep understanding of the implementation that traditionally came from manual authorship.
• The anxiety to move fast, driven by market FOMO, pushes many into unsustainable "crunch" practices, mirroring the toxic labor cultures seen in the gaming industry, rather than leveraging increased productivity for a better work-life balance.
• AI-generated content—whether it be code, documentation, or articles—is increasingly viewed as "slop," leading to a crisis of meaning for creators who find it harder to take pride in achievements that feel easily replicable by machines.
• Expert developers who have mastered their tools and abstractions often view the "productivity boost" of AI with skepticism, noting that for those with deep domain knowledge, writing code was never the primary bottleneck.
• The pervasiveness of AI-influenced writing styles, even in genuine personal essays, has triggered a cynical reaction among readers who struggle to distinguish human insight from corporate-aligned, generated content, further complicating online discourse.
• For some veteran developers, AI serves as a powerful accelerator that returns them to the "magical" feeling of rapid creation they experienced as beginners, suggesting that the divide in satisfaction may be less about technical ability and more about individual motivations.
The discussion reflects a deep-seated tension between the efficiency gains afforded by generative AI and the erosion of developer satisfaction. While some find that AI allows them to act as architects and focus on higher-level problem solving, many others report a profound sense of "skill rot" and a loss of identity as the meditative, satisfying aspects of coding are replaced by the grueling task of managing machine-generated outputs. The skepticism toward AI-authored content further complicates the landscape, as the community struggles to maintain a sense of authentic human connection in an era where the boundary between effortful creation and automated production is rapidly dissolving. Ultimately, the debate hinges on whether software development is viewed as an artistic craft or a purely instrumental means to deliver a functional product to the end user.