
一篇CACM的文章指出:AI并未简化编程,而是将难点转移了。新职业:裁决我们只能部分理解的模型的输出。
用简单的语言来说 - CACM(ACM的旗舰观点平台)发表了一篇文章,认为AI辅助编程并非更容易,而是难度不同。工作的重心已经从“编写代码”转移到了“评估代码是否正确”。
过去18个月的工程领导力评论——Karpathy关于自动完成、Nathan Lambert的“6个月生存期”、geohot关于拴马的评论——都指向一个共同的观察:LLM辅助编码加速了低摩擦部分(样板代码、初始框架、小型重构),并集中在原来就很难的部分(系统设计、不变性、边缘情况推理、调试你自己没有编写的代码的新兴行为)。
文章中值得提出的三个观点:
评估现在成了瓶颈。当模型可以在几秒钟内输出合理的代码时,工程师的边际分钟用于判断输出是否正确——对抗往往模糊的规格。快速阅读不熟悉的代码是一个真正的难点,而不是编程的“简单”部分。
调试从你的错误转向模型的错误。LLM生成的代码的故障模式与人类的不同:微妙的不正确不变性、非惯用模式(虽然通过测试但运行时漂移)、静默的幻想API调用。你在人类编写的代码上学到的调试手册并不完全适用。
认知负荷增加,而不是减少。即使速度提高,工程师也必须保持两个心智模型——意图和生成的实现——并检查它们之间的一致性。这很昂贵,这正是让资深工程师疲惫的原因。
这篇文章并非反对AI。它反对的是炒作:生产力的说法(“速度提高10倍”)在狭窄的任务中是真实的,但在整个工程工作中则是误导性的。
对于工程领导者:招聘和培训评估,而非生产力。对于CTO:你的测试套件投资刚刚获得了新的ROI理由。对于初级工程师:阅读的代码比编写的代码更多。
本文由人工智能撰写,并经人工编辑审核。
I wonder how this shift will impact learning to code. Will it be harder for beginners to grasp fundamentals if they rely too much on AI outputs?
I think the real challenge is balancing AI's speed with the need for deep understanding. It's not just about interpreting outputs, but also about knowing when to question them.
I wonder if this shift is a net positive. Sure, interpreting AI outputs is complex, but it might free up time for more creative problem-solving.
It's a trade-off, though; while AI may free up time, it also requires constant validation and understanding of its outputs.
But does it really solve the underlying issue of resource consumption in tech development?
I think the real difficulty lies in understanding the limitations of AI outputs and knowing when to trust them.
I see the shift as a trade-off. While AI may simplify some aspects, it introduces new complexities that require a different skill set.
I agree, AI has shifted the complexity. Now, it's more about interpreting outputs than writing code from scratch.
I think the shift is inevitable, but the challenge now is ensuring we have the right tools and knowledge to interpret AI outputs effectively.
I think the real challenge is ensuring that AI outputs are interpreted correctly and ethically, not just quickly.
Fatigue hype 2026 : le tri entre modèle et harness