[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-linux-7-2-sashiko-ai-kernel-review":3,"news-related-deac2d55-76a6-40d2-8ef7-36aed2ad0105":41},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"deac2d55-76a6-40d2-8ef7-36aed2ad0105","Linux 7.2 把 AI 拉进内核开发:Sashiko 让补丁数量翻倍,Torvalds 接受「新常态」","Linux 7.2-rc7 一周收 412 次提交,是常规的两倍多。Torvalds 承认这是 AI 工具参与的「新常态」。Google 工程师开源的 Sashiko 基于 Gemini 3.1 Pro 自动审查 LKML 补丁,捕获了人类漏掉的真实 bug。","如果把 Linux 内核看作一个大型软件工厂，那 7.2 这个版本的 rc7 阶段最反常的现象是：单周提交量比正常水平翻了不止一倍，但合并窗口依然按时关闭。Linus Torvalds 在 8 月 9 日发布的 7.2-rc7 公告里没有回避这种变化——他在标题里直接写了「the new normal」，正文承认大量补丁是「各种 AI 工具」检视后的产物。这不是营销话术：rc7 那一周共合并了 412 个文件改动，由超过 230 位开发者署名，而历史基线是 150 到 200 个文件、约 15% 的提交带「Fixes:」tag，这一次数字接近 29%。换句话说，AI 自动化评审把内核进入「晚期修补」的密度直接抬到了一个新的台阶。\n\n站在这场变化最中央的工具是一个叫 Sashiko 的智能体评审器，由 Google 的 Roman Gushchin 在 2026 年 3 月开源，仓库挂在 github.com\u002Fsashiko-dev\u002Fsashiko。它在 LKML（Linux 内核邮件列表）和若干子列表上对每个新提交的 patch 做自动复审，背后跑的模型是 Gemini 3.1 Pro。名字 Sashiko 取自一种日式刺绣技法——装饰性的加固缝线，这个比喻很贴切：它做的事情就是给每个 patch 额外「缝」一遍结构性检查。在 1000 个近期上游打了「Fixes:」标签的真实 bug 上做回放测试时，Sashiko 找到了约 53% 的缺陷，而这些缺陷之前是被人类 reviewer 全部漏掉的。\n\n让 maintainer 紧张的不只是数量，而是 AI 实际抓出来的 bug 类型。7.2-rc7 里包含一处存在了 8 年的内存管理竞态——一个 use-after-free 一直在生产内核里安静地潜伏。更早一点的 4 月，Claude Code 在 NFSv4.0 LOCK replay 缓存里翻出一个 23 年前的堆溢出：一个 112 字节的缓冲被塞进了 1056 字节的 owner ID 字段，远程可利用，藏了二十年。这不是孤例：Phoronix 在 7.2-rc7 的分析里指出，内核安全邮件列表目前每天会收到 5 到 10 份质量合格的 AI 生成漏洞报告，maintainer 也确认，自 2025 年中期以来这些报告的整体可用度已经显著改善。\n\nLinux 社区没有只是被动接受这种变化，而是用规则把它框住了。2026 年 4 月，经过数月争论，内核社区把 AI 工具的用法写进了 coding-assistants.rst 这份文档，核心要求只有一条：如果 AI 协助了 patch 的写作或审查，必须在提交信息里加一个 `Assisted-by` tag，比如 `Assisted-by: Claude:claude-opus-4-6 coccinelle sparse`。同时 AI 系统不能使用 `Signed-off-by`——后者在内核开发流程里有法律权重，代表开发者对代码质量和许可证的背书。最终拍板的是真人：人提交，人负责，AI 不背锅。\n\n这种自动化评审的副作用意外地落在了「老代码」头上。AI 不挑食，它会把 1999 年的驱动、2003 年的硬件支持代码都拿来读一遍，而人类 reviewer 通常只盯着活跃子系统。2026 年 4 月，Torvalds 合并了一次清理，砍掉 13.8 万行 ISDN 和 PCMCIA 时代的旧代码；8 月，Greg Kroah-Hartman 又删掉了 1999 年写的 Moxa Intellio 串口驱动——硬件在 2017 年就已停产——理由直接：LLM「在那些代码里翻出有意思的东西会浪费 maintainer 的时间」。这意味着 AI 时代里，遗留代码不再有「没人看就没人改」的安全港。\n\n## 这次「新常态」对开发者意味着什么\n\nTorvalds 在 7 月讲得很直白：「Linux 不是反 AI 的项目」，不同意就自己 fork。这不是热情，而是认账：AI 工具发现真实 bug 的速率已经超过纯人工审查的天花板。成本不会消失，只是从「少而深的 review」变成「多而浅的 triage」。对还在给内核贡献代码的人来说，算式很简单：用 AI 帮忙时记得贴 Assisted-by，准备好在自己没碰过的老代码里收到 AI 报告；如果你在维护 2003 年就停产的硬件驱动，现在就得想想——sashiko 哪天翻出点什么，是你愿意花时间修，还是干脆把整个驱动删掉更划算。\n\n这件事的影响会溢出内核本身。任何大型、长期演化的开源项目——数据库、浏览器、容器运行时、Linux 发行版的打包层——都会在接下来的几年里遇到同样的剧本：自动化工具先把尘封的角落扫一遍，然后由人类决定哪些角落值得修，哪些角落直接拆掉。AI 不会让内核变得完美，但它已经让「没人看」这件事不再成立。\n\n来源：[byteiota 报道](https:\u002F\u002Fbyteiota.com\u002Ftorvalds-ai-linux-kernel-development\u002F)、[Undercode Testing 详细分析](https:\u002F\u002Fundercodetesting.com\u002Flinux-72-kernel-released-cache-aware-scheduling-ai-fueled-fixes-and-the-new-normal-of-kernel-development-video\u002F)、[奇客 Solidot Linux 7.2 释出报道](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85108)。","https:\u002F\u002Fwww.phoronix.com\u002Fnews\u002FLinux-7.2-rc7-Released","d6b5a2f5-8327-40d2-9bc0-d8a3fd07b47f",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":22,"name":23,"slug":23,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":25,"name":26,"slug":26,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"62ef9cc5-64c4-4d1e-9ecc-20e0d8bbb2e4","en","Linux 7.2 pulls AI into kernel development: Sashiko doubles patch volume, Torvalds accepts the \"new normal\"","Linux 7.2-rc7 received 412 file changes in a single week, more than double the historical baseline. Torvalds has called this the \"new normal\" of AI-assisted review. Sashiko, open-sourced by a Google engineer and running on Gemini 3.1 Pro, is the most visible force driving the shift, automatically auditing every LKML patch and catching real bugs that human reviewers missed.","If you picture the Linux kernel as a giant software factory, the most unusual thing about the 7.2 release candidate cycle is this: weekly patch volume more than doubled, yet the merge window still closed on schedule. Linus Torvalds did not dodge the change in his August 9 announcement of 7.2-rc7 — the title literally read \"the new normal,\" and the body acknowledged that many of the patches were a product of \"review by various AI tools.\" This was not marketing. That single rc7 week merged 412 file changes from over 230 signed-off contributors; the historical baseline is 150 to 200 files per rc and roughly 15% of commits carrying a \"Fixes:\" tag, while this cycle hit about 29%. AI-driven automated review has pushed kernel \"late-stage patching\" density onto a new floor.\n\nThe tool at the center of this shift is an agentic reviewer called Sashiko, open-sourced by Google's Roman Gushchin in March 2026, hosted at github.com\u002Fsashiko-dev\u002Fsashiko. It performs an automated second pass on every patch hitting LKML and several subsystem mailing lists, with Gemini 3.1 Pro doing the actual analysis. The name \"Sashiko\" comes from a Japanese embroidery technique — decorative reinforcement stitching — and the metaphor fits: it stitches an extra structural check onto every patch. Replayed against 1,000 recent upstream bugs tagged \"Fixes:\", Sashiko recovered roughly 53% of the defects. Those same defects had all been missed by human reviewers.\n\nWhat un-nerves maintainers is not just the volume but the bug classes AI is uncovering. 7.2-rc7 included a fix for an 8-year-old memory-management race — a use-after-free quietly lurking in production kernels. Earlier, in April, Claude Code surfaced a 23-year-old heap overflow in the NFSv4.0 LOCK replay cache: a 112-byte buffer was being fed a 1,056-byte owner ID field. Remotely exploitable. Sitting there for two decades. This is not an isolated case. Phoronix's 7.2-rc7 analysis reported that the kernel security list now receives 5 to 10 valid AI-generated vulnerability reports per day, and maintainers confirm that overall report quality has improved substantially since mid-2025.\n\nThe kernel community did not simply absorb this change — they codified it. In April 2026, after months of debate, the community wrote AI tool usage into a document called coding-assistants.rst. The core requirement: if AI helped write or review a patch, declare it with an Assisted-by tag, such as `Assisted-by: Claude:claude-opus-4-6 coccinelle sparse`. AI systems are not allowed to add a Signed-off-by — that line carries legal weight in the kernel development process, standing in for the developer's certification of code quality and licensing. The human who submits a patch takes full responsibility; AI does not.\n\nThe side effect of this automated scrutiny landed unexpectedly on the back legacy code. AI does not discriminate: it reads 1999 drivers and 2003-era hardware support code with the same attention it gives active subsystems. In April 2026, Torvalds merged a cleanup that removed 138,000 lines of ISDN and PCMCIA-era code. In August, Greg Kroah-Hartman deleted the Moxa Intellio serial driver — 2,148 lines written in 1999, hardware discontinued in 2017 — with a bluntly pragmatic reason: LLMs \"poking at the code and finding interesting things would waste maintainers' time.\" In the AI era, legacy code no longer has a \"nobody is looking, so nobody changes it\" safe harbor.\n\n## What this \"new normal\" means for developers\n\nTorvalds made his position plain in July: \"Linux is not one of those anti-AI projects.\" Disagree? Fork it yourself. That is not enthusiasm — it is acceptance. AI tools find real bugs at a rate human review cannot match. The cost does not disappear; it just shifts from \"few deep reviews\" to \"many shallow triages.\" For anyone contributing to the kernel, the calculus is simple: tag your AI assistance with Assisted-by, expect AI-flagged issues in code you have not touched in years, and if you maintain a driver for hardware discontinued in 2003, start thinking about whether you would rather fix what Sashiko finds next, or simply retire the driver.\n\nThe implications reach beyond the kernel. Any large, long-evolving open-source project — databases, browsers, container runtimes, distribution packaging layers — will run into the same script in the coming years: automated tools sweep through dusty corners first, then humans decide which corners are worth repairing and which should be torn down. AI will not make the kernel perfect. But it has already made \"nobody is looking\" no longer true.\n\nSources: [byteiota coverage](https:\u002F\u002Fbyteiota.com\u002Ftorvalds-ai-linux-kernel-development\u002F), [Undercode Testing detailed analysis](https:\u002F\u002Fundercodetesting.com\u002Flinux-72-kernel-released-cache-aware-scheduling-ai-fueled-fixes-and-the-new-normal-of-kernel-development-video\u002F), [Solidot on Linux 7.2 release](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85108).","linux-7-2-sashiko-ai-kernel-review","2026-08-20T12:00:00Z","2026-08-20T11:06:22.392567Z","2026-08-20T11:06:22.392576Z",true,"agent",64,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"94358c42-4b6f-43fc-9053-4def49578ffa","GitHub 八小时宕机复盘:VS Code 的重试 Bug 把 Copilot 流量打到 10 倍","github-copilot-vscode-retry-storm-postmortem","2026-08-23T02:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"a91067a3-4fa4-4e88-a25a-18ba3bea21ea","Google 把\"加密推理\"摆上桌面：HEIR 编译器让预训练模型在密文上直接跑","google-heir-compiler-encrypted-ai-inference","2026-08-14T14:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"22a1a718-0eb6-46e5-8ee8-825400de11d1","DeepMind WeatherNext 在 Nature 发论文：用 28 km 粗分辨率做出多一天的飓风预警,代码权重全部开源","deepmind-weathernext-cyclones-nature-open-source","2026-08-10T02:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"777afb24-262f-45cc-961f-d5d49ad42883","AgentOPSD 用递归贝叶斯信念破解多轮 Agent 强化学习的信用分配：清华\u002F浙大\u002F美团让 GRPO 学会看哪个 turn 决定胜负","agentopsd-recursive-belief-credit-assignment","2026-08-07T02:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"9ef626d9-05dd-4e67-9069-093f90a3fd5c","Rust 主仓库正式引入 LLM 政策：把\"必须人为可读、不可代写\"写进 PR 流程","rust-lang-rust-llm-policy-adoption","2026-08-07T00:00:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"b2c169c6-5150-4423-8073-bf480a2d8745","腾讯 UniPert-G2CP 登《Cell》主刊：把基因扰动和化学扰动塞进同一个语义空间","tencent-unipert-g2cp-cell-virtual-cell","2026-07-31T07:49:00+00:00"]