这个问题虽然可以归咎于项目文档不全，但是却切实给我造成了麻烦：文档似乎永远不够全，规范似乎永远补不完。每次 AI 写一个功能我都需要打回修改至少 3~4 次，经常比我自己动手还要费事、烦人。而且随着规则越来越多，AI 的服从性也越来越差，后来甚至开始忘记之前的规范重新犯错误。所以到最后，我基本上放弃了让 AI 写后端代码，只用它调查问题源头或是写无所谓的内容。
Parameter cliff at ~800: Sharp accuracy transition observed by multiple researchers
，推荐阅读搜狗输入法2026获取更多信息
In recent years, LLMs have shown significant improvements in their overall performance. When they first became mainstream a couple of years before, they were already impressive with their seemingly human-like conversation abilities, but their reasoning always lacked. They were able to describe any sorting algorithm in the style of your favorite author; on the other hand, they weren't able to consistently perform addition. However, they improved significantly, and it's more and more difficult to find examples where they fail to reason. This created the belief that with enough scaling, LLMs will be able to learn general reasoning.
Heavy deluges and saturated soils have the potential to cause significant damage to homes, transport infrastructure and food supply.