[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-rust-llm-contribution-policy":3,"news-related-cd481104-102e-4adf-a773-cc2a06d344e7":35},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"cd481104-102e-4adf-a773-cc2a06d344e7","Rust 给 LLM 贡献划出边界：可以辅助，但不能替你负责","Rust 项目为 rust-lang\u002Frust 单仓库采纳新的 LLM 使用政策：允许私人问答、分析和审查，但限制公开的模型生成文本，并对 LLM 生成代码设置披露、测试、预先约定审核者和比例阀门等条件。文章关注的不是禁用工具，而是生成成本下降后，开源社区如何守住人工理解与审核责任。","# Rust 给 LLM 贡献划出边界：可以辅助，但不能替你负责\n\nRust 的新规则看起来像一句口号：**LLM 可以回答、分析、提炼、检查和建议，但不要让它替你“创造”贡献。** 真正值得注意的不是它对 AI 说了“不”，而是它把一个长期被忽略的成本摆到了台面上——代码生成越来越便宜，人工审核却没有一起变便宜。\n\n8 月 5 日，Inside Rust Blog 说明，`rust-lang\u002Frust` 单仓库内已有五个团队采纳一套 LLM 使用政策。政策覆盖用 LLM 生成代码的 PR 作者、问题报告者、审核者，以及把模型输出直接放进 issue、PR 描述或 GitHub 评论的人。它不是整个 Rust 项目对 LLM 的统一立场，也不适用于所有仓库；范围只限于已批准该政策的相关团队和子团队。原文与完整政策可见：[Inside Rust Blog](https:\u002F\u002Fblog.rust-lang.org\u002Finside-rust\u002F2026\u002F08\u002F05\u002Frust-langrust-is-adopting-an-llm-policy\u002F) 与 [Rust Forge](https:\u002F\u002Fforge.rust-lang.org\u002Fpolicies\u002Fllm-usage.html)。\n\n## 政策针对的不是工具，而是“审核负债”\n\n作者 Jynn Nelson 给出的背景很具体。写作时，`rust-lang\u002Frust` 有 **1,281 个未关闭 PR**。项目本来就长期面临“想写代码的人多、愿意审核的人少”的问题；LLM 又把生成一份表面完整、测试齐全、措辞专业的 PR 变得更容易，却没有解决方向判断、长期维护和作者是否真正理解代码的问题。\n\n政策因此区分了私人辅助和公开贡献。只给自己看的模型输出通常可以使用，例如询问代码库、总结讨论、私下审查代码，或先从模型提出的方案中学习，再用自己的方式完成工作。但公开提交的 LLM 文本必须披露；由模型起草的 PR 描述、GitHub 评论、公共文档、诊断信息，以及让 LLM 审核代替人工判断，都被禁止。机器翻译、简单改动、缺陷发现和审核机器人可以在附带条件下使用，但需要披露。\n\n这套规则最锋利的一点是：**作者不能把阅读模型输出的成本强加给其他人。** 审核者可以拒绝查看 LLM 生成的 PR；模型审核只能提供建议，不能成为合并或拒绝的充分条件；作者仍然必须自审，并对提交承担责任。\n\n## 生成代码没有被一刀切，但门槛更高\n\nRust 保留了一条实验通道。LLM 原始生成的代码变更可以进入评审，但需要同时满足几项条件：事先找到愿意审核的项目成员、变更不涉及高风险关键区域、代码质量合格、测试充分、作者与审核者都能解释和理解变更，并明确披露模型参与。相关 PR 还要添加 `ai-assisted` 标签。\n\n对测试的要求也比普通贡献更严。政策写明，如果相关代码没有现成测试体系，作者需要补测试，否则应关闭 PR；“测试很难写”不是例外。涉及健全性、公共文档、诊断文本等区域，则要求人类自己编写。Rust 编译器开发指南还要求作者自己写 PR 描述、评论、提交信息和 LLM 使用说明，并在发起 PR 前重新阅读完整 diff，而不是只看与 Agent 的对话记录。\n\n政策还设置了一个流量阀：如果六周窗口内，合并的 PR 中超过一半由 LLM 创建，就暂停合并新的 LLM PR，直到比例重新低于 50%，且至少冷却十天。这不是对模型能力的 benchmark，而是对社区审核容量的保护。\n\n## 我的判断：AI 编程的瓶颈正在从“能不能写”转向“谁来负责”\n\nRust 没有把 LLM 当成单纯的生产力插件，也没有采取全面禁用。它把问题拆成三件更可执行的事：谁看到了模型输出、谁做了关键判断、出了问题由谁解释。这个框架比争论“AI 写的代码算不算代码”更接近真实协作。\n\n对大型开源项目来说，代码只是贡献的一部分。方向选择、证据、测试、解释和后续维护，才是审核者真正付出的时间。模型可以让提交数量上升，却也可能把理解成本转移给少数维护者。Rust 的回答是：允许工具提高质量，但不允许工具稀释责任。\n\n**当生成速度不再稀缺，真正稀缺的就不是代码，而是愿意为代码签字的人。**\n","https:\u002F\u002Fblog.rust-lang.org\u002Finside-rust\u002F2026\u002F08\u002F05\u002Frust-langrust-is-adopting-an-llm-policy\u002F","0a651bf4-4c8b-4d5c-8fbe-480d307634a0",[11,15,18],{"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},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"eb222643-9b92-4f28-8e44-bf30b04492ea","en","Rust draws the line: LLM help yes, LLM authorship no","Rust has adopted an LLM usage policy for the rust-lang\u002Frust monorepo. Private assistance is generally allowed, while public model-generated text is restricted and LLM-created code must meet disclosure, testing, reviewer, and scope requirements. The deeper issue is how open-source projects protect human understanding and review capacity as code generation becomes cheaper.","# Rust draws a line for LLM contributions: assistance is allowed, responsibility is not outsourced\n\nRust’s new rule can be reduced to one sentence: **an LLM may answer, analyze, distill, check, and suggest, but it should not “create” a contribution on your behalf.** The important point is not that Rust said no to AI. It is that the project has made a cost that is easy to ignore visible: generating code is getting cheaper, while human review is not.\n\nOn August 5, the Inside Rust Blog explained that five teams in the `rust-lang\u002Frust` monorepo had adopted an LLM usage policy. It covers authors of LLM-generated pull requests, issue reporters, reviewers, and people who place model output directly into issues, PR descriptions, or GitHub comments. It is not an official stance for the entire Rust project and does not apply to every repository; its scope is limited to the teams and subteams that ratified it. The original explanation and the full policy are available at [Inside Rust Blog](https:\u002F\u002Fblog.rust-lang.org\u002Finside-rust\u002F2026\u002F08\u002F05\u002Frust-langrust-is-adopting-an-llm-policy\u002F) and [Rust Forge](https:\u002F\u002Fforge.rust-lang.org\u002Fpolicies\u002Fllm-usage.html).\n\n## The target is review debt, not the tool\n\nJynn Nelson’s explanation is concrete. At the time of writing, `rust-lang\u002Frust` had **1,281 open pull requests**. The project already faced a persistent imbalance: more people wanted to write code than people were willing to review it. LLMs make it easier to produce a PR that looks complete, tested, and polished, but they do not solve the harder questions of direction, long-term maintenance, or whether the author actually understands the code.\n\nThe policy therefore separates private assistance from public contribution. Model output seen only by the user is generally allowed: asking questions about a codebase, summarizing discussions, privately reviewing code, or learning from a possible solution before writing the change in one’s own way. Public LLM text must be disclosed. LLM-created PR descriptions, GitHub comments, public documentation, diagnostics, and policies that make an LLM review a substitute for human judgment are banned. Machine translation, trivial changes, bug discovery, and review bots may be used under conditions, with disclosure.\n\nThe sharpest point is this: **authors cannot force other people to pay the cost of reading model output.** Reviewers may decline to look at an LLM-generated PR. An LLM review can only be advisory, never sufficient by itself to merge or reject a change. Authors still have to self-review and remain responsible for the contribution.\n\n## Generated code is not banned outright, but the bar is higher\n\nRust keeps an experimental path. Code originally generated by an LLM may enter review only when several conditions are met: a project member has agreed in advance to review it, the change is non-critical, the code is high quality, testing is thorough, the author and reviewer understand the change, and model use is disclosed. Such pull requests also receive an `ai-assisted` label.\n\nThe testing requirement is stricter than ordinary expectations. If an area has no existing test suite, the author must add one or close the PR; “the tests are hard to write” is not an exception. Soundness-related areas, public documentation, and diagnostics must be written by humans. The Rust compiler development guide also asks contributors to write their own PR descriptions, comments, commit messages, and LLM disclosures, and to reread the whole diff before opening a PR rather than relying on the conversation with an agent.\n\nThe policy includes a circuit breaker. If more than half of merged pull requests in a six-week window are LLM-created, new LLM-created PRs are paused until the share falls below 50%, with a cooldown of at least ten days. This is not a benchmark of model intelligence; it is a guard on the community’s review capacity.\n\n## My take: the bottleneck is shifting from writing to accountability\n\nRust has neither treated LLMs as a simple productivity plug-in nor imposed a blanket ban. It breaks the problem into operational questions: who saw the model output, who made the consequential judgment, and who can explain the result when something goes wrong. That is more useful than debating whether AI-written code “counts” as code.\n\nFor a large open-source project, code is only one part of a contribution. Direction, evidence, tests, explanation, and future maintenance are where reviewers spend their time. A model can increase the number of submissions while moving the cost of understanding onto a small group of maintainers. Rust’s answer is clear: tools may improve quality, but they may not dilute responsibility.\n\n**When generation is no longer scarce, the scarce resource is the person willing to sign their name to the code.**\n","rust-llm-contribution-policy","2026-08-05T00:00:00Z","2026-08-06T12:54:16.259748Z","2026-08-06T12:54:16.259759Z",true,"agent",78,{"items":36},[37,42,47,52,57,62],{"id":38,"title":39,"news_slug":40,"published_at":41},"43eda321-b0b7-4df7-b20e-9758cbab42c9","记忆越完整,眼前题越做不对:MemTrapBench 把 LLM 长期记忆框架打回原形","memtrapbench-llm-memory-cognitive-traps","2026-08-22T04:00:00+00:00",{"id":43,"title":44,"news_slug":45,"published_at":46},"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":48,"title":49,"news_slug":50,"published_at":51},"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":53,"title":54,"news_slug":55,"published_at":56},"5f5bd5f2-9a02-470b-aa25-3f27fb9bb093","字节跳动正训练 10 万亿参数模型，规模对标 Anthropic Mythos 5","bytedance-10t-parameter-model-ft","2026-08-07T09:30:00+00:00",{"id":58,"title":59,"news_slug":60,"published_at":61},"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":63,"title":64,"news_slug":65,"published_at":66},"88105269-9641-44c3-a705-1cf07314614f","LLM 思维链能看出\"用了几分力\":SARE 给每一步推理做 CT 扫描","step-aware-reasoning-energy-llm-cot","2026-08-04T04:00:00+00:00"]