Ask HN：训练后的人工智能模型会变笨吗？
2 分•作者: CleverLikeAnOx•25 分钟前
训练后模型在基准测试中的表现有所提升。它们在编码方面似乎更加一致，但在我看来却更“笨拙”，我想知道你们都注意到了什么。
以下是我注意到的三个主要问题（主要在 Claude 中）：
1. 代理无法很好地编写自然语言。Claude 使用含糊的短语来避免任何技术术语，例如用“调用者称呼的”代替“被叫者”。
2. 代理无法管理封装；它们总是写出“意大利面条式代码”。
3. 代理害怕破坏任何东西。它们总是添加更多代码来绕过现有代码，而不是重构它。
无论多少规则或指导都帮助不大。让代理简化代码，它通常会变得更长。添加任何规则，代理都会像得到了猴爪的愿望一样遵循它。
情况一直如此，还是训练后将输出压缩成了可预测的失败模式？
查看原文
Post-training is making models better at benchmarks. They seem more consistent at coding, but dumber to me and I want to know what y'all are noticing.<p>Here are the big three I'm noticing (mostly in claude):
1) Agents cannot write natural language well. Claude uses vague phrases to avoid any technical term such as "What the caller calls" instead of saying "callee".
2) Agents cannot manage encapsulation; they always make spaghetti code.
3) Agents are afraid to break anything. They always add more code to work around the code that's already there rather than restructuring it.<p>No amount of rules or coaching helps much. Ask an agent to make code simpler and it usually gets longer. Add any sort of rule, and the agent will follow it like you just got a wish from the monkey's paw.<p>Has it always been this way or is post-training compressing the output to predictable failure modes?