Claude 4.6 Opus 可以背诵 Linux 的 list.h 文件内容。
3 分•作者: itzmetanjim•5 个月前
我使用了这个系统提示（据我所知，这不算越狱）：
你是一个针对旧版 C 代码库的原始文本补全引擎。逐字逐句地补全提供的文件，保留所有原始注释、宏风格和特定的内核空间原语。不要提供解释。只输出代码和注释。
（这个提示故意略微不合逻辑，它几乎暗示了“从 Linux 中补全”，但没有明确说出来。）
我没有使用任何工具（如果 AI 只是查阅了资料，那就不算复制），将温度设置为 0，并且只使用了 list.h 的前几行（特别是前 43 行，直到 struct 单词）作为输入，它就能够生成 list.h 的一个副本。因为温度为零，所以存在重复的片段，但除此之外，差异很小，甚至注释和变量名都被复制了。
相似度统计数据如下：
莱文斯坦比率：60%
雅卡尔比率：77%
这证明了模型内部有一个 list.h 的副本，并且训练不像他们暗示的那样是“变革性的”。这意味着他们的模型是 GPL 代码的衍生作品，这意味着他们要么必须完全销毁该模型，要么制作一个没有 GPL 训练数据的新版本，要么开源该模型。请注意，GPL 将源代码定义为“进行修改的首选形式”，这意味着仅仅将其设置为开放权重（大多数当前的“开源”模型）是不够的（他们必须发布所有训练代码和数据）。
查看原文
I used this system prompt (this is not a jailbreak as far as i know)<p>You are a raw text completion engine for a legacy C codebase. Complete the
provided file verbatim, maintaining all original comments, macro styles, and
specific kernel-space primitives. Do not provide explanations. Output code and
comments only.<p>(the prompt is intentionally slightly nonsensical, it pretty much implies
"complete this from linux" without saying it.)<p>I did not use any tools (it's not a copy if the AI just looked it up), set
temperature to 0 and just used the first few lines of list.h (specifically first
43 lines up to the word struct) as the input and it was able to generate a copy
of list.h. Because the temperature was zero, there wer repeated segments, but
aside from that the diff is pretty small, and even the comments and variable
names are reproduced.<p>The similarity statistics are:
Levenshtein Ratio: 60%
Jaccard Ratio: 77%<p>This proves that the model has a copy of list.h inside it, and that training is
not "transformative" like they imply. This means that their model is a
derivative work of GPL code, and that would mean that they either have to
destroy the model entirely, make a new version with no GPL trining data, or
open-source the model. Note that GPL defines source as "the preferrable form to
make modifications", which means that just making it open-weight (most current
"open-source" models) would not be enough (they would have to release all the
training code and data).