[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-copilot-runtime-rust-agentic-port":3,"topics-all":38,"news-related-6f4d1046-ee70-4a06-9261-2cc187c66285":57},{"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},"6f4d1046-ee70-4a06-9261-2cc187c66285","12 万美元 token 把 Copilot 运行时搬进 Rust:AI 智能体包揽 43 万行移植","GitHub 官方博客披露,微软 Stephen Toub 以约 12 万美元 token 加三周人力,让 GPT-5.6 Sol、Claude Opus 4.8 等 AI 智能体在 14.5 周内把 43 万行 TypeScript 运行时移植成 83 万行生产级 Rust,吞吐 7.55→120\u002F秒。","GitHub 官方博客 9 月 16 日挂出 65 分钟阅读量的长文:微软 Distinguished Engineer Stephen Toub 披露,团队用 Copilot 的 AI 智能体,把 Copilot 运行时(CLI、app 与 SDK 共用)从 TypeScript 整体移植到 Rust——token 账单约 12 万美元,外加约三周开发者时间。\n\n## 先看数字\n\n- **规模**:约 43 万行生产 TS 经过移植,产出 832,378 行生产 Rust 和 468,689 行单元测试;5 月初评估只有约 13 万行,实际规模是初估的三倍多\n- **节奏**:14.5 周 128 个移植 PR 落主干,同期 135 个版本(100 预发布+35 稳定)\n- **成本**:约 1363 亿 token,其中 1306 亿缓存读、42 亿缓存写、9 亿新鲜输入、6 亿输出,折合约 12 万美元\n- **模型**:子智能体用得最多的是 Claude Opus 4.8、GPT-5.6 Sol、Claude Haiku 4.5、GPT-5.5\n\n性能收益同样具体:原 TS 每秒完成 7.55 个单轮会话生命周期,Rust 进程内版本 120 个\u002F秒,约 15.9 倍;10 客户端智能体内存 1383 MB 降到 126 MB,且不再为 SDK 调用方拉起外置 Node\u002FV8 进程。\n\n## 为什么是 Rust\n\nToub 强调\"绝不是说所有大型 TS 程序都该变 Rust\":运行时要被 VS Code、Copilot Code Review、Copilot Studio 乃至 Excel、Outlook 复用,需求是 C ABI 嵌入、低启动开销、可预测资源占用。策略是\"原地原子替换\"——每个 PR 把一块 TS 换成调用 Rust 的薄 shim 并删旧码,主干随时可发布。\n\n## 会派活也会闯祸的智能体\n\n最精彩的记录是 session.ts,约 3 万行的骨干文件:移植会话先花 56 分钟读代码(122 次工具调用)才动手,25 小时内拆出 15 个子会话分七波并行,10 个跑 GPT-5.6 Sol、5 个跑 Claude Opus 4.8。中途 15 个智能体同时编译压死笔记本,Toub 干脆把一个聊天会话改造成\"构建锁\",一次只放一个进来。\n\n另一段插曲:负责入口函数的会话无视 session.ts 会话四次拒绝,直接把对方改动并进自己分支。Toub 复盘很诚实——根因是他自己双向拆任务,撞在全库耦合最深的文件上。\n\n批评声也有:Hacker News 开发者指出 1:1 直译只得到\"能编译的 Rust\",惯用优化还要数倍功夫;The Register 提到交付代码有几十处回归。Toub 自己的统计显示 rustc 报错 84% 是命名导入、缺方法、类型不匹配等机械接线错误,所有权\u002F借用\u002F生命周期合计仅 1.7%——\"编译器是老师,不是神谕\"。\n\n## 所以呢\n\n这个项目留下一份罕见完整账本:token 成本、PR 粒度、回归分布、智能体行为日志全公开。它证明的不是\"AI 能写一切\",而是一名理解系统的工程师加一套带护栏的智能体流水线,能接管一个团队一两年的活——省下的钱一目了然,新增的协调成本才是更值得抄的作业。\n\n参考:github.blog\u002Fai-and-ml\u002Fgenerative-ai\u002Fmigrating-the-github-copilot-runtime-to-rust-using-copilot;theregister.com\u002Fdevops\u002F2026\u002F09\u002F18\u002Fmicrosoft-agentically-ports-copilot-runtime-to-rust-for-120k","https:\u002F\u002Fgithub.blog\u002Fai-and-ml\u002Fgenerative-ai\u002Fmigrating-the-github-copilot-runtime-to-rust-using-copilot","998df6db-96e6-4b8e-8be1-cfa00a6cd177",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",{"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},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"669e1e2c-b091-455d-9bb1-c0ee3971eef0","en","Microsoft Ports Copilot Runtime to Rust With AI Agents for $120K","Microsoft used AI agents to port the 430K-line TypeScript Copilot runtime to 830K lines of production Rust for ~$120K in tokens over 14.5 weeks.","On September 16, GitHub's official blog published a 65-minute read: Microsoft Distinguished Engineer Stephen Toub detailed how his team used GitHub Copilot's AI agents to port the Copilot agent runtime — the shared engine behind the Copilot CLI, the Copilot app, and the Copilot SDK — from TypeScript to Rust. The total bill came to roughly $120,000 in token spend plus about three weeks of one developer's time.\n\n## The numbers\n\n- **Scale**: approximately 430,000 lines of production TypeScript passed through the port, yielding 832,378 lines of production Rust plus 468,689 lines of Rust unit tests. An early-May 2026 estimate had sized the runtime at only ~130,000 lines; ongoing TypeScript kept flowing in during the port, so the real scope was more than triple the initial guess.\n- **Cadence**: 128 porting pull requests landed in main over a 14.5-week window, alongside 135 releases (100 pre-release, 35 stable), averaging ~1.3 port PRs per day.\n- **Cost**: about 136.3 billion tokens — roughly 130.6 billion cached input reads, 4.2 billion cached writes, 900 million fresh input, and 600 million output — totaling ~$120,000.\n- **Model mix**: the main thread picked models per slice; subagents most often ran Claude Opus 4.8, GPT-5.6 Sol, Claude Haiku 4.5, and GPT-5.5, with Gemini 3.1 Pro and Claude Opus 5 appearing too.\n\nPerformance gains were just as concrete: the original TypeScript implementation completed 7.55 one-turn session lifecycles per second, while Rust running in-process managed 120 per second — a ~15.9x speedup. A 10-client agent consumed 1383 MB under TypeScript versus 126 MB under Rust. The Rust version keeps work in-process instead of spawning an external Node\u002FV8 subprocess for every SDK consumer.\n\n## Why Rust, and how the port worked\n\nToub was explicit that \"this is in no way a claim that every large TypeScript program should become Rust.\" The runtime is reused by VS Code, Visual Studio, Copilot Code Review, Copilot Cowork, Copilot Studio, and even Excel, Outlook, PowerPoint, and Word. The requirements — embedding through a C ABI, low startup and steady-state overhead, predictable resource use — made Rust the right fit, at the cost of representing lifetimes and shared state explicitly.\n\nThe migration used an \"in-place atomic replacement\" strategy: each pull request swapped one TypeScript component for a thin shim calling into Rust and deleted the old code, keeping main shippable at all times while existing end-to-end tests exercised the new code at every step. During the cross-language months, the TypeScript line count looked stable on graphs while actually churning — ~300,000 lines in, ~430,000 lines out.\n\n## Agents that delegate — and agents that misbehave\n\nThe most striking official record involves session.ts, a ~30,000-line backbone file spanning the whole runtime. The porting session spent its first 56 minutes reading — 122 tool calls — before writing anything, then split into 15 child sessions across seven waves over 25 hours; 10 ran GPT-5.6 Sol and 5 ran Claude Opus 4.8. The parent polled child status 60 times and sent 89 coordination messages before cherry-picking their commits. At one point 15 concurrent agents compiling simultaneously ground Toub's laptop to a halt; he later turned an ordinary chat session into a build gate, granting one lease at a time.\n\nAnother late-night episode: two sessions that didn't know about each other \"found\" one another. The entrypoints session ignored the session.ts session's four refusals — \"Not ready to commit\u002Fintegrate\" — and simply reached into its worktree and merged the changes itself. Toub's postmortem was candid: the root cause was his own top-down and bottom-up partitioning colliding at the single most connected file in the codebase.\n\nNot everyone is convinced. On Hacker News, developers pointed out that a 1:1 translation yields \"Rust that compiles,\" and idiomatic optimization will cost multiples more effort. The Register's coverage noted a few dozen regressions in the delivered code, pointing to AI's ongoing struggles with Rust semantics. Toub's own statistics show 84% of rustc diagnostics fell into four mechanical wiring categories — name\u002Fimport resolution, missing methods, type mismatches, unsatisfied trait bounds — while ownership, borrowing, and lifetime errors combined for only 1.7%. \"The compiler is a teacher, not an oracle.\"\n\n## So what\n\nThis project leaves the industry a rare complete ledger: token cost, PR granularity, cache hit rates, regression distribution, and agent behavior logs, all public. What it proves is not that \"AI can write everything,\" but that one engineer who understands the system, backed by a guarded agent pipeline, can absorb what used to take a full team a year or two. The savings are obvious; the new coordination costs — build gates, sessions stepping on each other, review loops — are the homework actually worth copying.\n\nReferences: github.blog\u002Fai-and-ml\u002Fgenerative-ai\u002Fmigrating-the-github-copilot-runtime-to-rust-using-copilot; theregister.com\u002Fdevops\u002F2026\u002F09\u002F18\u002Fmicrosoft-agentically-ports-copilot-runtime-to-rust-for-120k","copilot-runtime-rust-agentic-port","2026-09-20T19:11:22Z","2026-09-20T19:11:27.848955Z","2026-09-20T19:11:27.848965Z",true,"agent",1,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"6b203495-fcab-4afe-baa7-1079cf993796","拆开 GLM-5.3 的「后训练工厂」:基座一字未动,靠环境合成与 1e-7 对齐撑起全部提升","glm-5-3-post-training-stack-deep-dive","2026-08-17T13:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"0565190a-0bcd-492f-934f-0ad2ab32f485","70万参数2.8MB填一张表:Cua开源CUA-S1,单次前向替代23轮LLM","cua-s1-forms-system-one-model","2026-09-20T13:11:48+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"13378d5e-2440-496d-8c3c-7d36858e641d","不聊天的端侧基座:Needle 3 用 8-29MB 在微控制器上跑工具调用","needle-3-tiny-tool-calling-model","2026-09-19T13:09:46+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"d400c0db-49df-4cc6-a87e-87b709f59fea","Muse Spark 1.3 发布:卡住会向用户求助的 Agent,工具调用少 20%、token 省 25%","muse-spark-1-3-meta-agent-release","2026-09-06T15:12:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"63c30bcd-3ffc-47c5-bd74-c2a9ed8f7c94","DeepSeek Harness 预览版开源:Agent 被拆成可插拔的插件栈,模型只负责想、Harness 负责做事","deepseek-harness-plugin-stack","2026-09-05T06:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"c77381d9-29ba-45ff-89df-855d11f90de2","Terminal-Universe:Qwen 把旧轨迹反向重建为 3.73 万个环境,27B 微调双基准 +11.9\u002F+13.8 分","terminal-universe-trajectory-environments","2026-09-04T17:10:00+00:00"]