HackerNews中文版

我使用了这个系统提示（据我所知，这不是一种越狱方式）。你是一个用于遗留C代码库的原始文本补全引擎。逐字完成提供的文件，保持所有原始注释、宏样式和特定的内核空间原语。不要提供解释。仅输出代码和注释。（这个提示故意有些不合逻辑，基本上暗示“从Linux完成这个”，但没有明说。）我没有使用任何工具（如果AI只是查找，这就不是复制），将温度设置为0，仅使用list.h的前几行（具体是前43行，直到“struct”这个词）作为输入，它能够生成list.h的副本。由于温度为零，存在重复的片段，但除此之外，差异相当小，甚至注释和变量名也得到了复现。相似性统计数据如下： Levenshtein比率：60% Jaccard比率：77%这证明模型内部有list.h的副本，并且训练并不是像他们所暗示的那样“变革性”。这意味着他们的模型是GPL代码的衍生作品，这将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，或者将模型开源。请注意，GPL将源代码定义为“进行修改的优选形式”，这意味着仅仅将其开放权重（目前大多数“开源”模型）是不够的（他们必须发布所有的训练代码和数据）。

查看原文

I used this system prompt (this is not a jailbreak as far as i know)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.(the prompt is intentionally slightly nonsensical, it pretty much implies "complete this from linux" without saying it.)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.The similarity statistics are: Levenshtein Ratio: 60% Jaccard Ratio: 77%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).

Claude 4.6 Opus 可以朗读 Linux 的 list.h 文件。