从零开始开发一个操作系统内核，通常需要一支经验丰富的工程师团队耗费数年时间。然而，CentOS联合创始人、Rocky Linux项目发起者Gregory Kurtzer近日展示了一项颇为特别的成果：他利用人工智能代理，仅用11天便打造出一个全新的操作系统内核Ginux。这个内核没有使用任何Linux内核代码，却已经能够启动Rocky Linux 10。尽管目前仍属于研究原型，但这一项目展示了AI参与复杂底层软件开发的全新可能。
Kurtzer上周通过社交平台X首次公布了Ginux项目，相关代码也已经上传至GitHub，并采用Apache 2.0开源许可证。Ginux这个名称融合了Linux、GPT和Greg三个元素，项目介绍甚至开玩笑称，为了避免作者过于自我膨胀，其中的字母G应当解释为GPT。
Ginux并不是Linux的一个分支，也没有从现有Linux代码中复制内核组件。它的内存管理、进程调度、文件系统和网络功能均从头编写，主要使用Rust语言开发。它与Linux真正共享的是应用程序二进制接口（ABI），也就是说，Ginux通过模拟Linux向软件提供的系统接口，使原本为Linux编写的程序能够在不修改代码的情况下运行。
项目开发者将这种设计描述为：Linux并不只是一个内核实现，也可以被视为内核向应用软件呈现的一种“个性”或兼容环境。目前Ginux只实现了Linux兼容模式，未来则不排除增加BSD兼容模式的可能。
不过，Ginux距离完整兼容Linux还有相当大的差距。根据项目文档，以Linux 6.12的x86-64接口为参照，其375个系统调用中，目前只有82个已经得到完整实现，另有245个只能部分工作，剩余48个则被有意设置为不可用。这意味着它虽然已经能够运行部分现有Linux软件，但尚不能视为成熟的通用Linux替代品。
Ginux还引入了一套名为“域”（Domains）的隔离机制。软件运行在彼此隔离的环境中，每个域拥有独立的进程、内存空间和文件系统。项目希望这种机制能够提供比传统容器更强的隔离边界。不过，Ginux仍然可以使用标准OCI容器镜像来构建各个域的文件系统。正是通过这种方式，它已经能够加载Rocky Linux 10，并将系统启动推进至systemd的多用户运行目标。
这项开发工作最引人关注的地方，是人工智能在其中承担的角色。Kurtzer在技术报告中将自己的方法称为“人类指导的智能体工程”（Human-Directed Agentic Engineering）。他负责确定整体架构、制定开发阶段和验收标准，而人工智能代理则承担了后续绝大部分执行工作，包括编写代码、启动内核、运行测试，以及发现问题后修改代码。
整个过程由一个持续运行的OpenAI Codex任务完成，使用的模型是GPT-5.6 Sol。在11天内，AI累计生成约13.9万行代码和测试内容，完成了1000多次代码提交。开发过程中，真实发行版Rocky Linux也发挥了重要作用：它在实际启动和运行时暴露出一些人工构造的测试用例未能发现的问题。
不过，Kurtzer并没有将Ginux包装成已经可以投入生产环境的操作系统。目前项目仍然完全运行在QEMU虚拟机中，尚不支持直接在物理硬件上启动，也没有GPU和NVMe存储设备支持。更重要的是，Ginux尚未接受独立的安全审计，其安全性还无法得到充分验证。
Kurtzer认为，人工智能如今已经具备编写复杂系统软件的能力，但真正困难的部分并没有因此消失。软件兼容性、安全性以及长期维护能力，仍然是决定一个操作系统能否真正成功的关键壁垒。换句话说，AI能够显著缩短代码从无到有的过程，却不能自动保证这些代码足够可靠，更不能替代长期的工程验证。
与此同时，Linux内核开发社区也在逐步尝试引入AI工具。上个月，Linux创始人Linus Torvalds曾使用Claude Code编写一组实验性的页面缓存补丁，但审查人员随后发现其中存在竞态条件，因此相关代码仍在审核之中。大约在同一时期，Linux内核维护者也同意不再要求贡献者必须披露代码是否经过AI辅助编写。
Ginux的意义或许并不在于它能够立即取代Linux，而在于它展示了一种新的软件工程模式：人类负责目标、架构和关键决策，AI代理负责大量具体实现和测试工作。过去需要数年才能初步成形的复杂系统，如今可能在很短时间内获得一个可以运行的原型。不过，从原型走向真正可靠、能够适配广泛硬件并长期维护的操作系统，仍然是一段漫长的工程道路。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The development of an operating system from scratch usually requires years of experience in a team of engineers. However, Gregory Kurtzer, co-founder of CentOS and co-founder of the Rocky Linux project, has recently demonstrated a remarkable success: using artificial intelligence agents, he built a completely new system of nuclear Ginux in just 11 days. The kernel does not use any Linux kernel code, but has been able to activate Rocky Linux 10. Although it is still a research prototype, the project demonstrates the new possibilities for AI to participate in the development of complex bottom software.
Kurtzer announced the Ginux project for the first time last week through social platform X, and the code has been uploaded to GitHub and used the Apache 2.0 open-source license. The name Ginux combines Linux, GPT and Greg, and the project presentation even makes a joke that, in order to avoid the author being too self-inflated, the letter G should be interpreted as GPT.
Ginux is not a branch of Linux and does not copy the inner nuclear component from the existing Linux code. Its memory management, process movement, document system and network functions are developed from scratch, mainly in Rust language. What it really shares with Linux is the application binary interface (ABI), that is, the system interface that Ginux provides to software by simulating Linux, enabling the application originally written for Linux to run without changing the code.
The project developer described this design as if Linux was not just a kernel realization, but could also be considered an “personal” or compatible environment presented by kernel applications. At present, Ginux has only achieved the Linux compatibility model, and the possibility of increasing the BSD compatibility model is not excluded in the future.
Ginux, however, is far from being fully compatible with Linux. According to the project document, based on the Linux 6.12 x86-64 interface, only 82 of its 375 systems are currently fully operational and 245 are only partially operational, while the remaining 48 are deliberately set as not available. This means that while it is already able to run some of the existing Linux software, it cannot yet be considered a mature generic Linux alternative.
Ginux also introduced a separation mechanism called Domains. The software operates in a segregated environment and each domain has a separate process, memory space and file system. The project hopes that this mechanism will provide a stronger separation border than traditional containers. However, Ginux can still construct document systems for each domain using a standard OCI container mirror. It is in this way that it has been able to load Rocky Linux 10 and move system start-up to a multi-user running goal to systemd.
The most interesting aspect of this development is the role played by artificial intelligence. In its technical report, Kurtzer referred to its methodology as “Human-Directed Human Engineering”. He was responsible for defining the overall architecture, developing the development phase and acceptance criteria, while the artificial intelligence agent was responsible for most of the follow-up work, including the preparation of codes, the activation of kernels, operational testing and the modification of codes when problems were identified.
The whole process was completed by an ongoing Open AI Codex mission using the GPT-5.6 Sol model. Within 11 days, AI accumulated approximately 139,000 line codes and test content and completed more than 1,000 code submissions. The real issue of Rocky Linux also played an important role in the development process: It exposed a number of artificially constructed test examples that were not detected during actual start-up and operation.
However, Kurtzer did not package Ginux into an operating system that could already be put into a production environment. The project is still fully operational on the QEMU virtual machine and does not yet support direct start-up on physical hardware, nor does it have GPU and NVME storage equipment. More importantly, Ginux has not yet been subject to an independent security audit and its security has not yet been fully validated.
According to Kurtzer, artificial intelligence is now capable of developing complex systems software, but the truly difficult part has not disappeared. Software compatibility, security and long-term maintenance capacity remain key barriers to the true success of an operating system. In other words, AI can significantly shorten the process from non-existent codes, but does not automatically guarantee that they are reliable enough, let alone substitute for long-term engineering certification.
Meanwhile, the Linux Nuclear Development Community is gradually trying to introduce the AI tool. Last month, Linus Torvalds, the founder of Linux, used Claude Code to prepare a set of experimental cache patches, but the reviewers subsequently discovered that there were competing conditions and the code was still under review. Around the same period, the Linux nuclear maintainer also agreed that the contributor would no longer be required to disclose whether the code had been prepared with AI support.
The significance of Ginux may not be that it can immediately replace Linux, but rather that it demonstrates a new software engineering model: human responsibility for objectives, architecture and key decision-making, and AI agent for a great deal of practical realization and testing. A complex system that has taken several years to form in its initial form may now be available in a very short period of time. However, moving from the prototype to a truly reliable operating system capable of adapting to a wide range of hardware and long-term maintenance remains a long engineering path.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.