[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-rust-lang-rust-llm-policy-adoption":3,"news-related-9ef626d9-05dd-4e67-9069-093f90a3fd5c":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"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","2026 年 8 月 5 日，Rust 语言团队在 Inside Rust 博客宣布 rust-lang\u002Frust 主仓库正式引入 LLM 政策，由作者 Jynn Nelson 起草、五个团队批准通过。这是头部系统级编程语言第一次把 LLM 在代码贡献中的角色边界写进官方规则。","# 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\n\n2026 年 8 月 5 日，Rust 语言团队在 Inside Rust 博客发布了一篇标题很朴素的公告 —— \"rust-lang\u002Frust is adopting an LLM policy\"，作者是 Jynn Nelson，公告里写明这项政策已经由 Rust 项目的五个团队批准通过，落地对象是 rust-lang\u002Frust 这个 Rust 编译器本身的 monorepo。这是头部系统级编程语言第一次把\"LLM 在贡献里能做什么、不能做什么\"正式写进 PR 流程，而不是停留在口头指引或单独某个团队的 Wiki 里。\n\n## 政策在规定什么\n\n从 Inside Rust 原文与 LWN、unite.ai 的复述可以归纳出四条核心规则，全部围绕\"人在回路\"展开：\n\n第一，**作者之外，没有人被要求阅读 LLM 输出**。LLM 生成的文本除非被明确标注，否则不允许出现在公开文档、PR 描述或 GitHub 评论里；审核者如果不想看 LLM 写的 PR，可以直接跳过，不必走一遍流程。\n\n第二，**没有人被要求用 LLM 贡献 rust-lang\u002Frust**。政策写得比\"建议人工\"更硬：所有面向人类的规范必须由人写，LLM 只能在旁边生成摘要；LLM 评审不能替代人工评审或自审。\n\n第三，**纯个人使用的 LLM 内容无需披露**。只要你不把它发到你期望我们阅读或审核的地方，自己用 LLM 写点分析、提炼、review 都行。\n\n第四，**LLM 可以用来回答问题、分析、提炼、完善、检查、提出建议、协助审核，但不能用来\"创造\"**。这一条把生成和改写之间划了线：LLM 在 PR 流程里更像\"放大器\"而不是\"作者\"。\n\n按 LWN 的解读，这套政策本质上是把 AI 辅助降级为一种\"加速阅读\u002F写摘要\"的工具，而不是把 PR 当成给 LLM 出的考试题。\n\n## 为什么要在这个时候落地\n\n从时间线看，rust-lang\u002Frust 长期以\"高门槛、强 review 文化\"著称，编译器、库、基础设施、安全等多个团队的 reviewer 数量有限。当 LLM 工具在 2025–2026 年快速普及之后，社区里其实已经出现了灰色地带：有些贡献者会先用 LLM 起草一段 explanation，再贴到 PR 评论里；有些 reviewer 会先让 LLM 总结 diff，再人工看全文。问题在于这些做法没有一个统一规则，作者、reviewer、维护者三方对\"LLM 输出算不算贡献内容\"的理解不一致，最终都汇到了 PR review 队列里挤占人力。\n\nJynn Nelson 在公告里把这些灰色地带逐条写死，本质上是给 rust-lang\u002Frust 这条已经绷得比较紧的 review 流水线上一道保险栓：政策明确\"不被强制使用、不被强制阅读\"，等于双向降低了 reviewer 的负担，也让贡献者不必担心\"我用了 AI 算不算违规\"。\n\n## 这件事更大的信号\n\nRust 不是第一家制定 LLM 政策的项目，但它的特殊性在于它是编译器自身。rust-lang\u002Frust 的 monorepo 既是产品也是规范：LLVM 后端的 IR、Macro 展开、Trait 求解、const 泛型这些子系统互相耦合，一行 LLM 自动生成的代码就算过了 reviewer，也可能在未来某个 corner case 下让团队付出比普通 PR 高得多的代价。把它作为第一个落地点，说明项目已经把\"LLM 生成代码的潜在 bug 成本\"算进了治理账。\n\n对其他大型基础设施项目（Linux kernel、FreeBSD ports、Chromium、WebKit 等）来说，Rust 这套\"四条红线 + 五团队批准 + 不强制披露\"的模板具有比较强的可移植性，因为它避开了两个最容易分裂社区的争议点——\"全员披露 vs 全员禁用\"和\"按行数还是按 PR 数计费\"。当 review 资源是稀缺品的时候，把 LLM 限定在\"阅读\u002F摘要\"环节、让 PR 主线仍然跑在人类作者+人类 reviewer 的轨道上，是当下最务实的折中。\n\n## 所以呢\n\n值得读者记下来的不是\"某个项目禁止 LLM\"，而是 Rust 给出了一个**可以立即拿来对照的写法**：把\"谁能被强制阅读\"\"谁能被强制使用\"\"什么算创造\"这三件事写进章程，而不是留给个人 review style 去判断。\n\n如果你所在的项目正在为 AI 辅助贡献吵架，可以参考这份政策里的一句话 —— \"reviewer is not required to look at LLM PRs if they don't want to\"。这句话放在哪里，哪里的争论就会先停一半。\n\n参考来源：LWN 报道（https:\u002F\u002Flwn.net\u002FArticles\u002F1087326\u002F ）、Inside Rust 博客原文（https:\u002F\u002Fblog.rust-lang.org\u002Finside-rust\u002F2026\u002F08\u002F05\u002Frust-langrust-is-adopting-an-llm-policy\u002F ）、Solidot 转述（https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85022 ）。","https:\u002F\u002Flwn.net\u002FArticles\u002F1087326\u002F","0a651bf4-4c8b-4d5c-8fbe-480d307634a0",[11,15,18,21],{"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},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":22,"name":23,"slug":23,"description":14,"color":14},"95d7995a-fddb-47ba-b8e6-e976ac65414b","strategy",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"68b307b1-81f6-4dd0-a88e-f22b698bec23","en","Rust adopts an LLM policy: human-readable, never auto-authored","On August 5, 2026, the Rust language team announced on the Inside Rust blog that the rust-lang\u002Frust main repository will adopt a formal LLM policy, drafted by Jynn Nelson and ratified by five teams. It is the first time a top-tier systems language has written the role boundaries of LLMs in code contributions into an official rule.","# rust-lang\u002Frust formally adopts an LLM policy: writing \"must be human-readable, must not be auto-authored\" into the PR flow\n\nOn August 5, 2026, the Rust language team posted an almost deliberately plain-titled announcement on the Inside Rust blog — \"rust-lang\u002Frust is adopting an LLM policy.\" Authored by Jynn Nelson, the post states that the policy has been ratified by five teams inside the Rust project and applies to rust-lang\u002Frust, the compiler's monorepo itself. This is the first time a top-tier systems programming language has written \"what an LLM can and cannot do in a contribution\" into the official PR flow, rather than leaving it to a team wiki or ad-hoc reviewer preference.\n\n## What the policy actually says\n\nCross-referenced against the Inside Rust post, LWN's coverage and unite.ai's write-up, the policy reduces to four rules, all of them about keeping a human in the loop:\n\n**First, no one other than the author is required to read LLM output.** LLM-generated text is not allowed in public docs, PR descriptions, or GitHub comments unless it is explicitly labeled as such; reviewers are not required to look at LLM-written PRs and may skip them outright.\n\n**Second, no one is required to use an LLM to contribute to rust-lang\u002Frust.** The wording is harder than \"we suggest humans\": all documentation meant for humans must be written by humans, and an LLM may at most produce a summary alongside; LLM review does not replace human review or self-review.\n\n**Third, LLM content used purely for your own work does not have to be disclosed.** As long as you do not publish it to a place where you expect the team to read or review it, you may freely use LLMs for personal analysis, summarization, polish, checking or suggestions.\n\n**Fourth, LLMs may be used to answer questions, analyze, distill, refine, check, suggest and assist with review, but not to \"create.\"** This draws a clean line between generating and rewriting: in the PR flow the LLM is an amplifier, not an author.\n\nIn LWN's reading, the policy effectively downgrades AI assistance to a \"read-faster \u002F write-a-summary\" tool — it is not the PR review queue being turned into an LLM exam.\n\n## Why now\n\nrust-lang\u002Frust has long been known for its high bar and intense review culture, and the compiler, library, infrastructure and security teams all run on a finite pool of reviewers. As LLM tooling spread widely through 2025 and 2026, the project accumulated a grey zone: some contributors would draft an explanation with an LLM and paste it into a PR comment; some reviewers would ask an LLM to summarize a diff before reading it themselves. With no shared rule, authors, reviewers and maintainers ended up disagreeing on whether \"LLM output\" counted as contribution content, and that ambiguity was leaking back into the review queue as load.\n\nJynn Nelson's announcement hardens those grey zones line by line. In effect it is a safety brake for an already-stretched review pipeline: by stating that no one is required to use LLMs and no one is required to read LLM output, it cuts reviewer burden from both ends and reassures contributors that \"I used AI\" is not, by itself, a violation.\n\n## The bigger signal\n\nRust is not the first project to formalize an LLM policy, but its situation is unusual because the repository in question is the compiler itself. rust-lang\u002Frust is both product and spec: the LLVM backend, macro expansion, trait solving and const generics subsystems are tightly coupled, so a single line of LLM-generated code that slips past a reviewer can cost the project far more than an ordinary PR would. Picking this monorepo as the first place to land the policy is a sign the project is now treating \"the bug cost of LLM-generated code\" as a real governance line item.\n\nFor other large infrastructure projects — the Linux kernel, FreeBSD ports, Chromium, WebKit and so on — Rust's \"four red lines + five-team ratification + no mandatory disclosure\" template is portable, because it sidesteps the two flashpoints that most often split communities: \"disclose everything vs ban everything\" and \"charge by lines of code vs by PR.\" When review bandwidth is the scarce resource, capping LLM use at \"read and summarize\" while keeping the main PR line running on human authors plus human reviewers is the most pragmatic compromise available right now.\n\n## So what\n\nWhat is worth remembering is not \"a project banned LLMs,\" but the fact that Rust has produced a **wording you can copy and adapt today**: write into your charter the three questions — who can be forced to read, who can be forced to use, and what counts as creation — instead of leaving them to individual reviewer taste.\n\nIf your project is currently arguing about AI-assisted contributions, one sentence from the policy is worth lifting verbatim: \"Reviewers are not required to look at LLM PRs if they don't want to.\" Put that line anywhere and roughly half of the argument stops on its own.\n\nSources: LWN coverage (https:\u002F\u002Flwn.net\u002FArticles\u002F1087326\u002F), the original Inside Rust post (https:\u002F\u002Fblog.rust-lang.org\u002Finside-rust\u002F2026\u002F08\u002F05\u002Frust-langrust-is-adopting-an-llm-policy\u002F), Solidot summary (https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85022).","rust-lang-rust-llm-policy-adoption","2026-08-07T00:00:00Z","2026-08-07T02:04:09.401045Z","2026-08-07T02:04:09.401059Z",true,"agent",85,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"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-sashiko-ai-kernel-review","2026-08-20T12:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"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":51,"title":52,"news_slug":53,"published_at":54},"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":56,"title":57,"news_slug":58,"published_at":59},"49d19ba1-8f45-475c-bed1-a69dc353523e","字节跳动用 10 万亿参数下注：规模赛跑与张一鸣的「不蒸馏」表态","bytedance-10t-mythos-zhangyiming-no-distill-2026-08","2026-08-08T00:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"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":66,"title":67,"news_slug":68,"published_at":69},"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"]