GitHub Copilot四成代码有漏洞，AI生成代码你还敢用？

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"日前，研究人员发现，"},{"type":"link","attrs":{"href":"https:\/\/copilot.github.com\/","title":null,"type":null},"content":[{"type":"text","text":"GitHub Copilot"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" 工具提供的代码建议中"},{"type":"link","attrs":{"href":"https:\/\/www.techradar.com\/news\/github-autopilot-highly-likely-to-introduce-bugs-and-vulnerabilities-report-claims","title":null,"type":null},"content":[{"type":"text","text":"有接近 40% 存在 bug"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"。为了对 Copilot 的实际价值做出量化，研究人员创建了 89 个测试场景以考查其代码建议质量，编写出了 1600 多个程序。经过全面审查，研究人员发现其中近四成存在安全漏洞。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"研究人员指出，造成这种问题的主要原因有两个："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"1、GitHub Copilot 参考的开源代码良莠不齐。GitHub 的代码量虽然庞大，但大部分都是未经审查的，源代码就有问题，GitHub Copilot生成的代码也会出现问题；"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"2、GitHub Copilot 分不清新代码和旧代码。我们常常提起“最佳实践”，但曾经的“最佳实践”，极有可能在几年后变成反面典型。GitHub Copilot 目前还分不清这种区别，照单全收，自然有问题。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"AI编程神器GitHub Copilot"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"作为一款 AI 结对编程工具，GitHub Copilot 的主要定位是提供代码补全与建议功能。它是 Visual Studio Code 的一个插件，可根据当前文件的内容和当前光标位置为你自动生成代码。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"首先我们要理解"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"什么是AI结对编程"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"结对编程用来代指两个程序员在一个计算机上共同工作。通常来说，会有一个人输入代码，而另一个人审查他输入的每一行代码。输入代码的人称作驾驶员，而审查代码的人称作观察员。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"那么AI 结对编程，可以简单理解为人机结对编程的升级版。具体来说GitHub Copilot 由 OpenAI Codex 提供支持，可从代码中提取上下文，并立即对单行代码甚至整个函数给出建议。而 OpenAI Codex 接受过公开源代码和自然语言的培训，能理解编程语言和人类语言。在工作时，GitHub Copilot 编辑器的扩展插件将程序员的注释和代码发送到 GitHub Copilot 服务，然后该服务使用 OpenAI Codex 来生成对代码的建议。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"跌落神坛只用了不到一周"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"在GitHub Copilot推出后，有关于它的负面消息这并不是第一次。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"就在其发布不到一周的时间里，GitHub Copilot就被推上了侵权的风口浪尖，引发了前所未有的争议。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"原因是有位程序员尝试了一下平方根倒数速算法（Fast Inverse Square Root），结果只是打出来这四个关键词，GitHub Copilot就逐行“完美复刻”了当年那段神奇的算法，甚至连当时的原版注释都被保留下来。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"这种现象引起了大量开发者的不满，他们认为，GitHub Copilot 宣称的基于公开代码训练其实是在未遵循开源许可证的情况下，肆意“抄袭”开源代码，事情愈演愈烈，甚至个别比较激进的人为了抗议 GitHub 未经授权和许可便使用受版权保护的源代码作为 GitHub Copilot 的训练数据，宣布弃用 GitHub。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"在争议中前行的AI编程"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"尽管AI编程有诸多争议，但它的脚步仍未停止。当人们开始关注GitHub Copilot背后的技术支撑OpenAI Codex后，今年八月，"},{"type":"link","attrs":{"href":"https:\/\/openai.com\/","title":null,"type":null},"content":[{"type":"text","text":"OpenAI"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"宣布开发一种将自然语言转换为编程代码的人工智能系统——Codex。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"Codex 更像是 OpenAI 的下一步产品。它建立在 Copilot 之上，前文说到Copilot 是一种与 Microsoft 的 GitHub 代码存储库一起使用的工具，使用它可有助于完成代码行。但Codex通过接受用英语编写的句子并将它们翻译成可运行的代码，换句话说，它是可以将自然语言转换为代码的AI系统，这意味着AI编程的概念又往前进了一步。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"对此，OpenAI 称 Codex 可让计算机能够更好地理解人们的意图，从而让每个人都可利用计算机来实现更多目的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"那么，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"AI写代码究竟可行不可行呢"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"虽然对于开发者们来说，重复性的任务占据了软件开发的大部分时间，并且很容易出错，研究人工智来改进软件开发的确令人兴奋，但是我们也不难从GitHub Copilot身上看到，就目前来看，这个工具仍旧不建议大家使用。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"第一，GitHub Copilot既然可以逐行“完美复刻”平方根倒数速算法，那么我们又能如何判别它给我们的代码一定是没有受版权保护的呢？这种可能性，同样容易让我们背上“侵权”的法律风险。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"第二，就连GitHub自己都表示GitHub Copilot建议的代码并不完美，有时候它提供的代码甚至毫无意义。因此，对于GitHub Copilot建议的代码来说，我们仍要小心其编码错误，还需要付出精力去审查，才能得到正确有效的代码。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"因此总的来说，原则上，让AI学会写程序的确能为软件工程带来一次巨大的变革，但是就目前来看，现有的技术水平远远达不到预期。当前的GitHub Copilot已经是人工智能领域的一项重大的突破，但同时也存在诸多障碍，这一点还需要引起我们的注意。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"参考链接："},{"type":"link","attrs":{"href":"https:\/\/www.solidot.org\/story?sid=68736","title":null,"type":null},"content":[{"type":"text","text":"https:\/\/www.solidot.org\/story?sid=68736"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/www.techradar.com\/news\/github-autopilot-highly-likely-to-introduce-bugs-and-vulnerabilities-report-claims","title":null,"type":null},"content":[{"type":"text","text":"https:\/\/www.techradar.com\/news\/github-autopilot-highly-likely-to-introduce-bugs-and-vulnerabilities-report-claims"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]}]}]}