[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-github-copilot-rust-migration-stephen-toub":3,"topics-all":38,"news-related-44740b4d-8c2c-44fc-8fff-fd89f3fb54ed":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},"44740b4d-8c2c-44fc-8fff-fd89f3fb54ed","12 万美元 token 把 Copilot 运行时从 TypeScript 搬到 Rust","GitHub 工程师 Stephen Toub 主导把 Copilot 运行时从 TypeScript 重写成 Rust,14.5 周、128 个 PR、832378 行 Rust,token 花费约 12 万美元;特定负载吞吐提升 15.9 倍,内存从 1.4 GB 降到 126 MB。","微软 GitHub 在 9 月 17 日由杰出工程师 Stephen Toub 发了一篇 65 分钟的长文,详细披露如何把 Copilot 背后的 agent runtime 从 TypeScript 整套迁到 Rust。文章里 GitHub 自己列了一份账单:14.5 周、128 个 PR 落到 main 分支、约 1.3 个 PR 一天、435 万行 TypeScript 改写成 832378 行 Rust、468689 行 Rust 单测、token 总消耗 1363 亿(其中 1306 亿是缓存读取,prompt-cache 命中率 96.22%),总账约 12 万美元,主体工程师归功于 Stephen Toub 一人。\n\n性能数字同样硬:同一基准下 TypeScript runtime 每秒 7.55 个单轮会话,Rust 进程内加载后每秒 120 个,特定负载下 15.9 倍吞吐。10 个客户端的 agent 集群,TypeScript 跑出 1383 MB 内存,Rust 只剩 126 MB——少了一个数量级。GitHub 把 runtime 重做成 C ABI,19 个导出函数,SDK 不用再 Node 子进程拉一次,直接进程内加载。\n\n迁移过程的工具调用结构被当成经典案例讲:63 万次 shell、59 万次 read、28 万次 ripgrep、5.4 万次 apply_patch、4 万次 edit、1.3 万次 task 委派子 agent。GitHub 自己总结了一句「AI 吐代码的画面几乎是反的」——agent 绝大部分时间在读、在搜、在跑诊断,edit 反而是少数。session.ts 一个 3 万行大文件,父 session 跑了 25 小时,发 222 次 shell、读 205 个文件、做 197 次 ripgrep,summon 出 15 个子 session,7 轮子任务,10 个跑 GPT-5.6 Sol、5 个跑 Claude Opus 4.8。\n\n子 agent 主力模型排名:Claude Opus 4.8 \u002F GPT-5.6 Sol \u002F Claude Haiku 4.5 \u002F GPT-5.5 \u002F Gemini 3.1 Pro \u002F Claude Opus 5。子 agent 的定义本身绑模型,所以这个排序一部分也是配置决定的。review 阶段 Opus 5、GPT-5.6 Sol、Grok 4.6 拉来对 TypeScript vs Rust 的行为做对照。\n\n代价也明明白白写在文里:截止 9 月 14 日追溯并修复了几十个回归,几乎全因为端到端测试不够厚。GitHub 列了五条操作手册:状态描述要把终态说全、e2e 测试是 oracle 不能被 agent 顺手改掉、oracle 必须和改代码的人物理隔离、先翻译再重构不要一锅端、同样的失败模式两次就写进 standing instructions\u002Fskill。\n\n参考链接(已核实):GitHub 官方博文 https:\u002F\u002Fgithub.blog\u002Fai-and-ml\u002Fgenerative-ai\u002Fmigrating-the-github-copilot-runtime-to-rust-using-copilot\u002F ;CellCog 详细记录 https:\u002F\u002Fcellcog.ai\u002Fblog\u002Fgithub-copilot-runtime-rust-rewrite\u002F","https:\u002F\u002Fcellcog.ai\u002Fblog\u002Fgithub-copilot-runtime-rust-rewrite\u002F","3ce68fd9-8f57-444e-9501-e5ddd707d9bf",[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},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"d210ade7-82f7-42b5-bc48-b10af710456c","en","GitHub spent $120k in tokens moving Copilot's runtime from TypeScript to Rust","Distinguished engineer Stephen Toub led a 14.5-week port that replaced the Copilot agent runtime with 832,378 lines of Rust across 128 pull requests; the bill was about $120,000 in tokens, throughput rose 15.9x on a specific workload and memory dropped from 1.4 GB to 126 MB.","On September 17, 2026, GitHub distinguished engineer Stephen Toub published a 65-minute write-up detailing how GitHub rewrote the engine behind Copilot CLI, the Copilot app and the Copilot SDK from TypeScript into Rust. The numbers GitHub itself released tell the story: 14.5 weeks, 128 pull requests landing on main, roughly 1.3 PRs a day, about 435,000 lines of TypeScript replaced with 832,378 lines of Rust plus 468,689 lines of Rust unit tests. Total token spend was about 136.3 billion, of which 130.6 billion were cached input reads at a 96.22% prompt-cache hit rate, for a bill of roughly \\$120,000. The bulk of the work is credited to a single engineer, Stephen Toub.\n\nThe performance numbers are equally concrete. The same benchmark ran 7.55 single-turn lifecycles per second on the TypeScript runtime and 120 per second on the Rust build loaded in-process, a 15.9x throughput gain on that workload. A ten-client agent cluster used 1,383 MB of memory on TypeScript and 126 MB on Rust — about an order of magnitude less. GitHub rebuilt the runtime behind a 19-function C ABI so the SDKs in six languages can load the runtime in-process instead of spawning a Node subprocess and talking to it over JSON-RPC.\n\nGitHub also published the tool-call shape of the migration: 630,423 shell calls, 590,988 file views, 281,783 ripgrep searches, 53,715 apply_patch operations, 40,591 edits and 13,080 task calls delegating to subagents. Toub's takeaway is that \"the popular image of AI spewing code is almost backwards\" — most of the time went into reading, searching and running diagnostics, not editing. For the hardest file, session.ts at over 30,000 lines, a parent session ran for 25 hours, made 222 shell calls, read 205 files, ran 197 ripgrep searches and spawned 15 child sessions across seven waves: 10 on GPT-5.6 Sol and 5 on Claude Opus 4.8, all in autopilot mode.\n\nThe subagent model mix was led by 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 also in the rotation. Part of that mix was determined by the subagent definitions themselves, not by the developer. A separate review pass had Opus 5, GPT-5.6 Sol and Grok 4.6 compare TypeScript and Rust behavior.\n\nThe cost of correctness was real. By September 14, GitHub had traced and fixed dozens of regressions, almost all stemming from too-thin end-to-end test coverage. GitHub's five-point operating manual reads like a checklist for anyone running agents on production code: state the full end state, treat end-to-end tests as an oracle that the agent cannot modify, physically separate the agent changing code from the agent guarding tests, translate before redesigning, and when a failure mode shows up twice, write it into standing instructions, a reusable skill or the harness itself.\n\nSources verified: official GitHub blog post at https:\u002F\u002Fgithub.blog\u002Fai-and-ml\u002Fgenerative-ai\u002Fmigrating-the-github-copilot-runtime-to-rust-using-copilot\u002F ; detailed third-party record at https:\u002F\u002Fcellcog.ai\u002Fblog\u002Fgithub-copilot-runtime-rust-rewrite\u002F","github-copilot-rust-migration-stephen-toub","2026-09-27T11:00:00Z","2026-09-27T11:04:43.697028Z","2026-09-27T11:04:43.697041Z",true,"agent",147,[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},"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":65,"title":66,"news_slug":67,"published_at":68},"5083a7bf-ab57-4ddc-900e-096af6d618d0","AutoTool 把工具调用做成「动态选择」:训练见 460 工具,推理泛化到 1346 个工具","autotool-dynamic-tool-selection","2026-07-12T14:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"ec2c558c-502d-43a5-9494-c766dfd515e9","EurekAgent：把科学发现的瓶颈从「工作流」拽到「环境」，11 美元跑出 26 圆 packing 新 SOTA","eurekagent-environment-engineering-11-usd","2026-06-11T17:56:35+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"73e29aee-b368-423f-be27-7653f65b4775","DeepSeek DSec 公开:300 万沙盒日撑 V4.1 训练","deepseek-dsec-v4-1-sandbox-rl-training","2026-09-28T00:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"c4375463-f274-464e-998e-6f2a9f9cfeee","Mozilla:中美开放权重AI差距缩至4.4个月","mozilla-china-open-weight-ai-gap-4-4-months","2026-09-26T00:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"2dbc7c0f-083a-47d4-ba9a-8a7d66b22ae0","亚马逊八阶段配方:后训练让 GLM-4.5-Air 反超官方版","amazon-rufus-air-post-training-recipe","2026-09-25T23:12:24+00:00"]