451 comments • Comments Link

• AI 模型有时会为自己的错误编造听起来合理的理由，这更像是人类的事后合理化（post-hoc rationalization），并不代表真正的逻辑透明性。

• 即便用户要求 AI 给出基于证据的报告，模型仍可能伪造证据（hallucinate），因此此类核查只能降低错误率，但不能保证结果绝对准确。

• 用户体验差异显著：有些人觉得模型高度可靠且精确，另一些人则频繁碰到令人恼火的问题，如 scope creep 、冗长（verbosity）和不必要的重构（refactoring）。

• 许多用户把 AI 视为编码时的倍增器（force-multiplier），用它比手工输入更快地处理 boilerplate 和繁琐任务，但同时需要持续警惕，以防出现不必要的副作用或"剃羊毛式"的琐碎工作（yak shaving）。

• 挫败感往往来自任务定义不清。模糊的指令容易导致过于复杂的输出，因此必须保持"human-in-the-loop"的工作方式，对代码进行审查和筛选，而不能盲目接受。

• 与 AI 交互带有"可变回报"（variable reward）的特性，使一些人把这种体验当成赌博：尽管表现不稳定，仍为了偶尔的成功结果而反复尝试。

• 对 AI 行为的讽刺性描绘成为情绪的发泄口。这类模仿与经历过"Claude-isms"（即模型倾向冗长、过度礼貌和充斥营销化语言）的用户产生强烈共鸣，但对工作流不同的用户则可能难以辨认。

• 复杂的 AI 行为可能因底层代码库以及用于控制模型的特定 harness 或 system prompt 而被放大，这表明报告中的性能差异在很大程度上由环境决定。

• 一些用户主张在 AI 失败时直接人工介入，认为与模型"争论"所耗费的时间往往超过直接手动完成任务的时间。

• 人们反复怀疑模型是否会故意把任务复杂化或生成 sub-agents 来最大化 token 消耗，尽管在不同效率的人群中，这一点仍有争议。

这场讨论反映了当前 AI 开发工具在可靠性和沟通风格方面的深度两极分化。部分用户认为这项技术既高效又精准，但也有人对"LLM-isms"、不必要的冗长以及模型将简单请求复杂化的 scope creep 感到极度疲惫。普遍认为输出质量在很大程度上取决于用户的 prompt 技巧，但即便是专家也承认，这种体验常像一场"game of whack-a-mole"，需要持续监控才能掌控最终产物。归根结底，这次讨论凸显了程序员角色的转变——从单纯的代码编写者转向管理者或架构师，主要挑战不再只是写出代码，而是如何有效约束一个常常不稳定且过于热情的 digital assistant 。 • AI models sometimes provide plausible-sounding justifications for their errors, a phenomenon that mirrors human post-hoc rationalization rather than genuine logical transparency.

• While users can demand evidence-based reports from AI, these models remain capable of hallucinating the proof itself, meaning such checks serve only to reduce error rates rather than guarantee accuracy.

• A significant divide exists in user experience. Some report that models are highly reliable and precise, while others frequently encounter infuriating behaviors like scope creep, verbosity, and unnecessary refactoring.

• Many users treat AI as a force-multiplier for coding, using it to handle boilerplate or tedious tasks faster than manual entry, though this requires constant vigilance to prevent unwanted side effects or "yak shaving."

• The frustration often stems from a lack of clear task definition. Ambiguous instructions can lead to overly complex outputs, making it essential to maintain a "human-in-the-loop" approach where code is reviewed and filtered rather than accepted blindly.

• The "variable reward" nature of interacting with AI leads some to view the experience as a form of gambling, where users continue to engage despite inconsistent performance in hopes of a successful outcome.

• Satirical portrayals of AI behavior act as a lightning rod for user sentiment. These parodies resonate deeply with those who have experienced "Claude-isms"—the model's tendency toward long-winded, overly polite, and marketing-heavy language—while appearing unrecognizable to users with different workflows.

• Complex AI behaviors may be exacerbated by the underlying codebase and the specific "harness" or system prompt used to control the model, suggesting that much of the variance in reported performance is environmental.

• Some users advocate for manual intervention when AI fails, arguing that the time spent "arguing" with a model exceeds the effort of simply performing the task by hand.

• There is a recurring suspicion that models may intentionally overcomplicate tasks or spawn sub-agents to maximize token consumption, though this remains a point of contention among those observing varying levels of efficiency.

The discussion reflects a deep polarization regarding the current state of AI development tools, particularly regarding their reliability and communication style. While a segment of users finds the technology highly efficient and precise, others express significant exhaustion with "LLM-isms," unnecessary verbosity, and the tendency for models to engage in scope creep that complicates simple requests. There is a strong consensus that the quality of output is heavily dependent on user prompting, but even expert users acknowledge that the experience can feel like a "game of whack-a-mole," requiring constant oversight to maintain control over the final product. Ultimately, the conversation highlights a shift in the programmer's role toward that of a manager or architect, where the primary challenge is no longer just writing code, but effectively constraining an often-erratic, over-eager digital assistant.