{
 "schema_version": "1.3",
 "$schema": "https://www.wangwu.ai/daily/schema/digest.v1.3.json",
 "generator": {
  "name": "personal-ai-daily",
  "version": "0.1.0"
 },
 "curator": {
  "name": "wangwu.ai",
  "url": "https://www.wangwu.ai/daily/"
 },
 "topics": [
  "agents",
  "agent-memory",
  "skills",
  "multi-agent",
  "on-device-models",
  "retrieval",
  "ai-infra",
  "papers",
  "ai-industry"
 ],
 "language": "zh-CN",
 "generated_at": "2026-10-29T10:20:00+08:00",
 "published_at": "2026-10-29T10:24:59+08:00",
 "fetch_window": {
  "start": "2026-10-28T07:24:37+08:00",
  "end": "2026-10-29T10:20:00+08:00"
 },
 "date": "2026-10-29",
 "title": "AI 要闻 · 2026-10-29",
 "url": "https://www.wangwu.ai/daily/2026-10-29/",
 "fetched": "2026-10-29 10:15–10:20（北京时间）",
 "sections": [
  {
   "id": "top3",
   "title": "今日最该知道的 3 条",
   "items": [
    {
     "id": "74eba3ef62ad631b",
     "canonical_url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/",
     "url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/",
     "title": "Gemini 4 Carbon：我们的下一代前沿模型",
     "summary": "官方称 Carbon 是目前最强 Gemini，重点放在网络安全与长程能力；距 Argon 约两周。第三方 Artificial Analysis 综合指数 v5.0 转述为 59 分（Argon 50、Opus 5.5 58、Astra 57），九项里四项第一，编程 Agent 指数仍低于 Opus 5.5（59 vs 62）。",
     "why": "前沿模型窗口从「综合分」切到「可部署的网安与长程 agent」。做防御/长任务编排的，先核官方可用范围与价格阶梯，再决定要不要把路由从 Argon 迁走。",
     "tags": [
      "Gemini",
      "网络安全",
      "长程"
     ],
     "section": "top3",
     "published_at": "2026-10-29T01:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.662,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/",
       "title": "Google · 10-29 01:00"
      },
      {
       "title": "sundarpichai",
       "url": "https://x.com/sundarpichai/status/2115489219804352921"
      },
      {
       "title": "0xLogicrw",
       "url": "https://x.com/0xLogicrw/status/2115612510602809723"
      }
     ],
     "signals": {
      "source_count": 3
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "gemini-4-carbon-2026-10",
     "source": "Google · 10-29 01:00"
    },
    {
     "id": "16933ee91d8224f9",
     "canonical_url": "https://blogs.microsoft.com/blog/2026/10/28/introducing-the-microsoft-copilot-harness-unlocking-the-potential-of-truly-multi-model-agentic-work/",
     "url": "https://blogs.microsoft.com/blog/2026/10/28/introducing-the-microsoft-copilot-harness-unlocking-the-potential-of-truly-multi-model-agentic-work/",
     "title": "推出 Microsoft Copilot harness：释放真正多模型智能体工作的潜力",
     "summary": "GitHub 工程侧那套多模型 Copilot harness 经 Microsoft Foundry 对外开放：管指令、上下文、执行环境与治理，宣称支持 MCP、agent skills 与 plugins，并可在 20+ 前沿模型与自带 key 间路由。运行时已迁 Rust，强调降会话开销与首 token。",
     "why": "大厂开始把「harness 本身」当产品卖，而不只卖模型。自建 agent 运行时的，对一下：你缺的是路由/沙箱/技能协议，还是又一层套壳；Foundry 托管是否锁死观测与记忆落点。",
     "tags": [
      "Harness",
      "Foundry",
      "多模型"
     ],
     "section": "top3",
     "published_at": "2026-10-28T12:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.704,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://blogs.microsoft.com/blog/2026/10/28/introducing-the-microsoft-copilot-harness-unlocking-the-potential-of-truly-multi-model-agentic-work/",
       "title": "Microsoft · 10-28"
      },
      {
       "title": "rohanpaul_ai",
       "url": "https://x.com/rohanpaul_ai/status/2115612239193686050"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "ms-copilot-harness-2026-10",
     "source": "Microsoft · 10-28"
    },
    {
     "id": "8228f1dd462d8181",
     "canonical_url": "https://arxiv.org/abs/2610.30537",
     "url": "https://arxiv.org/abs/2610.30537",
     "title": "PLE：物理工程最后一考",
     "summary": "85 道电/土木/结构/机械长程任务，要求在多格式项目材料上交付可计费工程包（含图纸与约束）。七种公开 harness 评测下，最强配置均分仅 50.6%，均耗约 2700 万 token、63 分钟/题；设计题约束满足率可到约 91%，但「每条硬约束都过」仅约 41%。",
     "why": "工程 agent 的瓶颈不在玩具题，而在硬安全/规范门禁。做领域 harness 时，先把硬约束 checker 写成一票否决，再谈脚手架调参。",
     "tags": [
      "基准",
      "工程 Agent",
      "长程"
     ],
     "section": "top3",
     "published_at": "2026-10-28T08:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.30537",
       "title": "arXiv · 2610.30537"
      },
      {
       "title": "rohanpaul_ai",
       "url": "https://x.com/rohanpaul_ai/status/2115567033773674980"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "ple-physical-engineering-2610-30537",
     "source": "arXiv · 2610.30537"
    }
   ]
  },
  {
   "id": "edge",
   "title": "专栏：端侧模型",
   "intro": "浏览器里跑编码 agent、MediaPipe 端侧任务更新，以及开源语音转写上桌面。",
   "items": [
    {
     "id": "8544fd5820f98f73",
     "canonical_url": "https://x.com/xenovacom/status/2115277783022862613",
     "url": "https://x.com/xenovacom/status/2115277783022862613",
     "title": "@xenovacom: 浏览器里跑 Pi + LFM3",
     "summary": "",
     "why": "称在 M4 Max 上用自研 WebGPU kernel + 4-bit GGUF，把 Pi coding agent 配 LFM3-1.7B 跑进浏览器，峰值逾 300 tok/s。端侧叙事从「能聊天」推进到「浏览器里可改代码」；复现时先核 kernel/权重许可与热路径是否仍依赖本机桥。",
     "tags": [
      "WebGPU",
      "LFM3",
      "编码 agent"
     ],
     "section": "edge",
     "published_at": "2026-10-28T11:01:38+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.412,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://x.com/xenovacom/status/2115277783022862613",
       "title": "X · xenova · 10-28 11:01"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "xenova-pi-lfm3-browser-2026-10",
     "source": "X · xenova · 10-28 11:01"
    },
    {
     "id": "7713a03df72ac931",
     "canonical_url": "https://github.com/google-ai-edge/mediapipe/releases/tag/v1.2.0",
     "url": "https://github.com/google-ai-edge/mediapipe/releases/tag/v1.2.0",
     "title": "MediaPipe v1.2.0",
     "summary": "Tasks 侧补 Android Kotlin TextToSpeech SDK、InteractiveSegmenter 接 LiteRT；iOS 开源文本校对与摘要任务；修 AHWB/SSBO 与池化纹理的 GPU 泄漏，并为 Parakeet ASR 流式降延迟。另升 OpenCV 4.10、Web Worker 平台探测等。",
     "why": "端侧多媒体流水线仍在补「语音出 + 分割入」拼图。已嵌 MediaPipe 的，优先回归 GPU 泄漏与 ASR 延迟项，再决定是否跟 Tasks 新接口。",
     "tags": [
      "MediaPipe",
      "端侧",
      "ASR"
     ],
     "section": "edge",
     "published_at": "2026-10-28T16:05:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.455,
     "importance": 0.3,
     "sources": [
      {
       "url": "https://github.com/google-ai-edge/mediapipe/releases/tag/v1.2.0",
       "title": "google-ai-edge · 10-28"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "mediapipe-v1-2-0",
     "source": "google-ai-edge · 10-28"
    },
    {
     "id": "cea7948628549b02",
     "canonical_url": "https://huggingface.co/blog/nvidia/nemotron-transcribe-granary",
     "url": "https://huggingface.co/blog/nvidia/nemotron-transcribe-granary",
     "title": "NVIDIA 发布开放权重语音模型 Nemotron 3.5 Transcribe 及 Granary v2 数据管线",
     "summary": "单检查点覆盖转写（约 40 语种）、翻译（约 31 语种英向）、说话人标注（约 13 语种、最多 8 人）与音频问答，宣称可处理约 1 小时录音；英文 Open ASR 榜与多项 diarization 对比靠前。量化版称将登陆 RTX/DGX Spark 与 Jetson。",
     "why": "开源语音栈开始把转写/翻译/说话人捆成「可提示」一体模型。本地会议 agent 选型时，先看你要的是多语 diarization 还是纯 ASR 延迟，再比参数量。",
     "tags": [
      "语音",
      "开源权重",
      "Apple Silicon"
     ],
     "section": "edge",
     "published_at": "2026-10-28T22:58:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.571,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://huggingface.co/blog/nvidia/nemotron-transcribe-granary",
       "title": "Hugging Face Blog · 10-28"
      },
      {
       "title": "Prince_Canuma",
       "url": "https://x.com/Prince_Canuma/status/2115567892175720458"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "nemotron-3-5-transcribe-2026-10",
     "source": "Hugging Face Blog · 10-28"
    }
   ],
   "note": "窗内 llama.cpp b11938–b11941 多为日常构建，未单列。LFM3 权重家族见昨日。"
  },
  {
   "id": "col",
   "title": "专栏：Agent 记忆与协作",
   "intro": "记忆「净收益」账本、设计向 skill，以及开源工作流合集信号。",
   "items": [
    {
     "id": "aaa0a39f9469ada8",
     "canonical_url": "https://arxiv.org/abs/2610.34073",
     "url": "https://arxiv.org/abs/2610.34073",
     "title": "记忆税：剖析 Agent 记忆的净价值",
     "summary": "多基准、多骨干上发现不少专用记忆系统相对朴素 RAG 净值为负。作者把相对 RAG 的增益/损失拆成 memory benefit 与 memory tax：重组事实或蒸馏偏好本意为增益，却也可能抹掉证据。代码与结果称将开源。",
     "why": "别再只报记忆模块绝对分。做个人或团队记忆库对比时，应用「相对直接检索的净收益」做账，并单列税项（证据被压掉、过时偏好）。",
     "tags": [
      "记忆",
      "RAG",
      "评测"
     ],
     "section": "col",
     "published_at": "2026-10-28T08:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.676,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.34073",
       "title": "arXiv · 2610.34073"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "memory-tax-2610-34073",
     "source": "arXiv · 2610.34073"
    },
    {
     "id": "8b8e06a8f61d4042",
     "canonical_url": "https://github.com/Leonxlnx/taste-skill",
     "url": "https://github.com/Leonxlnx/taste-skill",
     "title": "Leonxlnx/taste-skill",
     "summary": "",
     "why": "宣称用艺术指导、参考图与浏览器检查，让编码 agent 少出「通用 AI 页」。热度极高；接入前先看 skill 是否会改仓库外文件，以及检查步骤是否可在 CI 无头跑。",
     "tags": [
      "Skills",
      "设计",
      "Agent"
     ],
     "section": "col",
     "published_at": "2026-10-28T12:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.707,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://github.com/Leonxlnx/taste-skill",
       "title": "GitHub · ~105k★"
      }
     ],
     "signals": {
      "source_count": 1,
      "stars_total": 104566
     },
     "via": [],
     "key": "leonxlnx-taste-skill-2026-10",
     "source": "GitHub · ~105k★"
    },
    {
     "id": "f5baeeee6be1c93a",
     "canonical_url": "https://x.com/shao__meng/status/2115428696672866729",
     "url": "https://x.com/shao__meng/status/2115428696672866729",
     "title": "MiniMax Code 团队开源日常 Agent 工作流与 Skills 合集 MCode Stack",
     "summary": "",
     "why": "当作 Skills 目录信号；接入前核对仓库与许可，别只信转发标题。",
     "tags": [
      "Skills",
      "工作流"
     ],
     "section": "col",
     "published_at": "2026-10-28T20:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.64,
     "importance": 0.43,
     "sources": [
      {
       "url": "https://x.com/shao__meng/status/2115428696672866729",
       "title": "X · AIHOT"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "minimax-mcode-stack-2026-10",
     "source": "X · AIHOT"
    }
   ],
   "note": "HABIT-Bench 见论文栏。豆包「录制回放」仅短帖未展开。Codex alpha 无实质说明。"
  },
  {
   "id": "cl",
   "title": "主流产品 Changelog 精读",
   "items": [],
   "note": "今日无重要更新。Codex rust-v0.166.0-alpha.4–6 预发布页无变更说明；ChatGPT/Grok/DeepSeek/Gemini API changelog 抓取失败（jina/playwright）。"
  },
  {
   "id": "gh",
   "title": "GitHub 热门",
   "intro": "日榜里和 Agent / 感知相关的仓库。",
   "items": [
    {
     "id": "ff74511b47309146",
     "canonical_url": "https://github.com/expo/skills",
     "url": "https://github.com/expo/skills",
     "title": "expo/skills",
     "summary": "",
     "why": "Expo 官方 skill 集合。做移动/跨端 agent 工作流时可当参考目录，装前看是否绑定 Expo 专有 CLI。",
     "tags": [
      "Skills",
      "Expo"
     ],
     "section": "gh",
     "published_at": "2026-10-29T00:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.663,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://github.com/expo/skills",
       "title": "GitHub · ~3.6k★"
      }
     ],
     "signals": {
      "source_count": 1,
      "stars_total": 3612
     },
     "via": [],
     "key": "expo-skills-2026-10",
     "source": "GitHub · ~3.6k★"
    },
    {
     "id": "6947080b4e91ce77",
     "canonical_url": "https://github.com/ruvnet/RuView",
     "url": "https://github.com/ruvnet/RuView",
     "title": "ruvnet/RuView",
     "summary": "",
     "why": "用普通 WiFi 做空间感知与生命体征，不靠视频像素。和视觉 agent 路线对照有趣，但落地受射频环境与隐私法规约束更大。",
     "tags": [
      "感知",
      "WiFi",
      "无摄像头"
     ],
     "section": "gh",
     "published_at": "2026-10-29T00:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.434,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://github.com/ruvnet/RuView",
       "title": "GitHub Trending 日榜"
      }
     ],
     "signals": {
      "source_count": 1,
      "stars_total": 101107
     },
     "via": [],
     "key": "ruvnet-ruview-2026-10",
     "source": "GitHub Trending 日榜"
    },
    {
     "id": "1be80d0ac922f006",
     "canonical_url": "https://github.com/jakubkrehel/skills",
     "url": "https://github.com/jakubkrehel/skills",
     "title": "jakubkrehel/skills",
     "summary": "",
     "why": "个人 skills 合集进候选热度。装前扫一遍会不会覆盖全局 skill 目录，并和 taste-skill / Expo 官方包对照定位。",
     "tags": [
      "Skills"
     ],
     "section": "gh",
     "published_at": "2026-10-29T00:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.688,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://github.com/jakubkrehel/skills",
       "title": "GitHub · 候选"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "jakubkrehel-skills-2026-10",
     "source": "GitHub · 候选"
    }
   ],
   "note": "HKUDS/DeepTutor 昨日已收，无新版本不重复。jailbreaks 仓库不收。"
  },
  {
   "id": "eng",
   "title": "Agent 工程",
   "items": [
    {
     "id": "4d32a2a81239bfd0",
     "canonical_url": "https://arxiv.org/abs/2610.30435",
     "url": "https://arxiv.org/abs/2610.30435",
     "title": "自评估强化学习：检索增强大模型中的搜索状态意识",
     "summary": "提出 SARL：训练期让模型识别「已拿到什么、还缺什么」，推理期不再外挂评估器。双奖励既抓纠正效果又罚过度依赖评估。相对最强基线相对提升约 24.3%，多跳约 +35.9%、深研约 +27.9%；长搜索视野下训练更稳。",
     "why": "搜索 agent 的失败常是「不知道自己还缺证据」。做检索环时，把搜索状态显式进奖励，比只加工具调用轮数更对症。",
     "tags": [
      "RAG",
      "RL",
      "搜索 agent"
     ],
     "section": "eng",
     "published_at": "2026-10-28T08:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.477,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.30435",
       "title": "arXiv · 2610.30435"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "sarl-search-state-2610-30435",
     "source": "arXiv · 2610.30435"
    },
    {
     "id": "6e34c191fd559aa8",
     "canonical_url": "https://replit.com/blog/vibench-2-long-horizon-software-benchmark",
     "url": "https://replit.com/blog/vibench-2-long-horizon-software-benchmark",
     "title": "ViBench 2.0：长程软件构建基准",
     "summary": "",
     "why": "长程写软件基准又一轮。和昨日 Trial by Trace 对照：既要看任务成功，也要看 harness 是否在中途违规。原文细节以官博为准，短卡仅作入口。",
     "tags": [
      "基准",
      "长程",
      "软件工程"
     ],
     "section": "eng",
     "published_at": "2026-10-29T02:50:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.52,
     "importance": 0.8,
     "sources": [
      {
       "url": "https://replit.com/blog/vibench-2-long-horizon-software-benchmark",
       "title": "Replit · 快讯已见"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "vibench-2-2026-10",
     "source": "Replit · 快讯已见"
    }
   ]
  },
  {
   "id": "infra",
   "title": "检索与基础设施",
   "items": [
    {
     "id": "4334615c71f64c69",
     "canonical_url": "https://engineering.fb.com/2026/10/28/ml-applications/meta-andromeda-next-gen-ads-retrieval-gpu-co-design/",
     "url": "https://engineering.fb.com/2026/10/28/ml-applications/meta-andromeda-next-gen-ads-retrieval-gpu-co-design/",
     "title": "Meta Andromeda：通过模型-系统协同设计驱动下一代广告检索",
     "summary": "检索阶段以近排序级模型质量、在延迟预算内从海量候选里捞广告，吞吐称逾 10 亿广告/秒，较 2024 Andromeda 约 100×；称自 2025 起为 Facebook 转化带来约 2.4% 提升。路径是模型、系统与 GPU（延续 GH200 一代）一起共设计。",
     "why": "检索基础设施的增量来自「模型-系统-硬件」同改，而不是只换更大 embedding。做大规模召回的，先问哪一层是真正瓶颈。",
     "tags": [
      "检索",
      "GPU",
      "协同设计"
     ],
     "section": "infra",
     "published_at": "2026-10-29T00:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.531,
     "importance": 0.475,
     "sources": [
      {
       "url": "https://engineering.fb.com/2026/10/28/ml-applications/meta-andromeda-next-gen-ads-retrieval-gpu-co-design/",
       "title": "Meta Engineering · 10-28"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "meta-andromeda-2026-10",
     "source": "Meta Engineering · 10-28"
    }
   ]
  },
  {
   "id": "papers",
   "title": "论文",
   "intro": "Agent 记忆与检索相关。",
   "items": [
    {
     "id": "4116e478e4c11aa9",
     "canonical_url": "https://arxiv.org/abs/2610.24520",
     "url": "https://arxiv.org/abs/2610.24520",
     "title": "HABIT-Bench：从纵向弱证据评估 Agent 记忆中的习惯诱导",
     "summary": "用跨数百会话的合成用户史测习惯归纳与后请求应用；八种记忆法最好宏平均约 33.6%，直接喂标注证据仍受限。含临时例外、习惯变更与未采纳建议。",
     "why": "长期记忆不等于会用习惯。做偏好沉淀时要单测例外与习惯变更。",
     "tags": [
      "记忆",
      "习惯"
     ],
     "section": "papers",
     "published_at": "2026-10-28T08:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.666,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.24520",
       "title": "arXiv · 2610.24520"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "habit-bench-2610-24520",
     "source": "arXiv · 2610.24520"
    },
    {
     "id": "65cfc6d79b4b10ce",
     "canonical_url": "https://arxiv.org/abs/2610.37666",
     "url": "https://arxiv.org/abs/2610.37666",
     "title": "ThinkWorld：世界想象强化动力学思维链推理",
     "summary": "",
     "why": "世界模型与思维链结合的论文入口，可与昨日 Genie 4 对照。短摘要源，细节以 PDF 为准。",
     "tags": [
      "世界模型",
      "推理"
     ],
     "section": "papers",
     "published_at": "2026-10-28T08:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.449,
     "importance": 0.225,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.37666",
       "title": "arXiv · 2610.37666"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "thinkworld-2610-37666",
     "source": "arXiv · 2610.37666"
    }
   ],
   "note": "Memory Tax 见专栏；PLE 见三条；SARL 见工程。"
  },
  {
   "id": "biz",
   "title": "大厂与模型",
   "items": [
    {
     "id": "3e290668e478db3a",
     "canonical_url": "https://the-decoder.com/google-opens-android-development-to-third-party-ai-agents-and-unveils-new-gemini-apis/",
     "url": "https://the-decoder.com/google-opens-android-development-to-third-party-ai-agents-and-unveils-new-gemini-apis/",
     "title": "Google 向第三方 AI 智能体开放 Android 开发并推出新 Gemini API",
     "summary": "Android Studio「Bring Your Own Agent」称可接入 40+ 外部编码 agent（含 Claude、Codex）；并扩终端工具以减耗时与 token。另提 Gemini Nano 4 Agentic API、MCP 接入，以及与 XREAL 的 Android XR 眼镜 AURA（约 1099 美元）。",
     "why": "移动开发生态把第三方 coding agent 当成一等公民。做端侧 agent 分发的，盯 IDE 插件位与本机 Agentic API 权限模型，比再发一个聊天套壳更关键。",
     "tags": [
      "Android",
      "Agent",
      "Gemini"
     ],
     "section": "biz",
     "published_at": "2026-10-28T12:00:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.645,
     "importance": 0.8,
     "sources": [
      {
       "url": "https://the-decoder.com/google-opens-android-development-to-third-party-ai-agents-and-unveils-new-gemini-apis/",
       "title": "THE DECODER · 10-28"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "android-byo-agent-2026-10",
     "source": "THE DECODER · 10-28"
    },
    {
     "id": "18df32737decb1a7",
     "canonical_url": "https://techcrunch.com/2026/10/28/anthropic-just-launched-tools-to-show-companies-what-their-ai-spending-is-buying/",
     "url": "https://techcrunch.com/2026/10/28/anthropic-just-launched-tools-to-show-companies-what-their-ai-spending-is-buying/",
     "title": "Anthropic 推出企业报告工具，量化 Claude 的 token 支出回报",
     "summary": "面向企业的用量/产出/故障点报告，结合内部分析与员工调研给改进建议；称默认不用该数据训模型，并推团队共享预算池。文内转述 Anthropic 对投资者的季度数据（含约 900% NRR）作背景，ROI 叙事针对大客户留存。",
     "why": "企业买的是可解释的 token 账单，不是又一个聊天窗。做团队 agent 计费/审计时，把「产出物与失败步骤」和用量绑在同一张表。",
     "tags": [
      "Anthropic",
      "企业",
      "ROI"
     ],
     "section": "biz",
     "published_at": "2026-10-28T12:19:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.645,
     "importance": 0.8,
     "sources": [
      {
       "url": "https://techcrunch.com/2026/10/28/anthropic-just-launched-tools-to-show-companies-what-their-ai-spending-is-buying/",
       "title": "TechCrunch · 10-28"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "anthropic-enterprise-roi-reports-2026-10",
     "source": "TechCrunch · 10-28"
    },
    {
     "id": "c84572f8132ffeb4",
     "canonical_url": "https://the-decoder.com/new-artificial-analysis-index-widens-the-gap-between-open-and-closed-models/",
     "url": "https://the-decoder.com/new-artificial-analysis-index-widens-the-gap-between-open-and-closed-models/",
     "title": "Artificial Analysis 新版指数拉大开源与闭源差距",
     "summary": "",
     "why": "AA 改闭卷基准后，开源相对闭源差距被测得更大。读榜单时先看题集是否仍公开可刷，再拿来做选型。",
     "tags": [
      "评测",
      "开源",
      "闭源"
     ],
     "section": "biz",
     "published_at": "2026-10-28T23:04:00+08:00",
     "first_seen_at": "2026-10-29",
     "score": 0.432,
     "importance": 0.55,
     "sources": [
      {
       "url": "https://the-decoder.com/new-artificial-analysis-index-widens-the-gap-between-open-and-closed-models/",
       "title": "THE DECODER · 10-28"
      },
      {
       "title": "emollick",
       "url": "https://x.com/emollick/status/2115277286123704654"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "key": "aa-index-closed-bench-2026-10",
     "source": "THE DECODER · 10-28"
    }
   ],
   "note": "Carbon / Copilot harness / PLE 见今日三条。Genie 4、Skill Adaptation、Codex 0.165、LFM3 见昨日。"
  },
  {
   "id": "x",
   "title": "X 精选",
   "items": [
    {
     "id": "d574cb3ce50b4e85",
     "canonical_url": "https://x.com/mattpocockuk/status/2115538750147023323",
     "url": "https://x.com/mattpocockuk/status/2115538750147023323",
     "title": "@mattpocockuk: 公司级软件工厂课",
     "summary": "",
     "why": "宣称拍完职业生涯最野心课程，11/9 起两周。关注 Skill/软件工厂叙事的，可当公开课大纲信号，不必等完整讲义。",
     "tags": [
      "工程与习惯",
      "课程",
      "软件工厂"
     ],
     "section": "x",
     "published_at": "2026-10-29",
     "first_seen_at": "2026-10-29",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/mattpocockuk/status/2115538750147023323",
       "title": "@mattpocockuk · 10-29 04:18 · ♥ 716"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 716
     },
     "via": [],
     "source": "@mattpocockuk · 10-29 04:18 · ♥ 716"
    },
    {
     "id": "8a104b9a87e21f92",
     "canonical_url": "https://x.com/addyosmani/status/2115480412151762979",
     "url": "https://x.com/addyosmani/status/2115480412151762979",
     "title": "@addyosmani: Hunger beats tenure",
     "summary": "",
     "why": "转发 Anthropic 弃传统编程面试的讨论：新同学执行力追上资深。招 agent 工程时，可对照「能交付闭环」权重是否高于题海。",
     "tags": [
      "工程与习惯",
      "招聘",
      "Anthropic",
      "执行"
     ],
     "section": "x",
     "published_at": "2026-10-28",
     "first_seen_at": "2026-10-29",
     "score": 0.406,
     "importance": 0.475,
     "sources": [
      {
       "url": "https://x.com/addyosmani/status/2115480412151762979",
       "title": "@addyosmani · 10-28 16:26 · ♥ 479"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 479
     },
     "via": [],
     "source": "@addyosmani · 10-28 16:26 · ♥ 479"
    },
    {
     "id": "bdd031b45d61a000",
     "canonical_url": "https://x.com/0xLogicrw/status/2115396120717451612",
     "url": "https://x.com/0xLogicrw/status/2115396120717451612",
     "title": "@0xLogicrw: RWE-Bench 真工程题",
     "summary": "",
     "why": "Xplore 20 道机电航任务：Astra 单次约 37%，五次后 Opus 5.5 约 70%。和今日 PLE 对照：真工程交付仍远未饱和。",
     "tags": [
      "工程与习惯",
      "RWE-Bench",
      "工程"
     ],
     "section": "x",
     "published_at": "2026-10-28",
     "first_seen_at": "2026-10-29",
     "score": 0.56,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://x.com/0xLogicrw/status/2115396120717451612",
       "title": "@0xLogicrw · 10-28 18:51 · ♥ 6"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 6
     },
     "via": [],
     "source": "@0xLogicrw · 10-28 18:51 · ♥ 6"
    },
    {
     "id": "dbc882e49d74ff0e",
     "canonical_url": "https://x.com/tison1096/status/2115591732972998728",
     "url": "https://x.com/tison1096/status/2115591732972998728",
     "title": "@tison1096: 催 Carbon 开放",
     "summary": "",
     "why": "Ultra 用户称 Argon 写稿体验不错、催 Carbon。产品节奏信号，不作评测依据。",
     "tags": [
      "工程与习惯",
      "Carbon",
      "体验"
     ],
     "section": "x",
     "published_at": "2026-10-29",
     "first_seen_at": "2026-10-29",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/tison1096/status/2115591732972998728",
       "title": "@tison1096 · 10-29 07:49 · ♥ 11"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 11
     },
     "via": [],
     "source": "@tison1096 · 10-29 07:49 · ♥ 11"
    },
    {
     "id": "88b26e1304b82aee",
     "canonical_url": "https://x.com/0xLogicrw/status/2115354482473361599",
     "url": "https://x.com/0xLogicrw/status/2115354482473361599",
     "title": "@0xLogicrw: Stainless 删库牵连客户",
     "summary": "",
     "why": "称 Anthropic 收购的 SDK 工具公司删私有仓导致客户代码没了。用托管生成 SDK 时，备份与退出条款要默认按最坏情况写。",
     "tags": [
      "工程与习惯",
      "Anthropic",
      "Stainless",
      "供应链"
     ],
     "section": "x",
     "published_at": "2026-10-28",
     "first_seen_at": "2026-10-29",
     "score": 0.562,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://x.com/0xLogicrw/status/2115354482473361599",
       "title": "@0xLogicrw · 10-28 16:06 · ♥ 7"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 7
     },
     "via": [],
     "source": "@0xLogicrw · 10-28 16:06 · ♥ 7"
    }
   ]
  }
 ]
}
