{
 "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": "2026-10-22T13:34:49+08:00",
 "published_at": "2026-10-22T13:35:05+08:00",
 "fetch_window": {
  "start": "2026-10-19T07:25:00+08:00",
  "end": "2026-10-20T07:13:00+08:00"
 },
 "date": "2026-10-20",
 "title": "AI 要闻 · 2026-10-20",
 "url": "https://www.wangwu.ai/daily/2026-10-20/",
 "fetched": "2026-10-22 12:37–12:55（北京时间；10-22 补发，内容取自 10-19 07:25 起已入库的抓取）",
 "sections": [
  {
   "id": "top3",
   "title": "今日最该知道的 3 条",
   "items": [
    {
     "id": "1aa8f83b01a3a053",
     "canonical_url": "https://www.theverge.com/ai-artificial-intelligence/1013643/ai-agents-hack-john-deere-shark-ninja-cummins",
     "url": "https://www.theverge.com/ai-artificial-intelligence/1013643/ai-agents-hack-john-deere-shark-ninja-cummins",
     "title": "一名黑客用一群 AI agent 攻入 John Deere 和 SharkNinja",
     "summary": "攻击门槛正在从写代码降到写一段提示：安全公司 Norforge 发现，一名攻击者在开源 agent 平台 Cairn 上只用不到 100 个汉字交代入侵思路，agent 便基本自主地攻破了 100 多个网站。",
     "why": "被攻破的未必是核心系统：源里举的两例，Cummins 是第三方供应商运营的网站，SharkNinja 是独立的产品支持站。agent 能连续二十多个小时不停试探，企业该先清点暴露在外、没人维护的边缘站点，也要把 agent 平台自身公开暴露的实例当成新的攻击面来管。",
     "tags": [
      "安全",
      "AI Agent 攻击"
     ],
     "section": "top3",
     "published_at": "2026-10-19T21:58:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://www.theverge.com/ai-artificial-intelligence/1013643/ai-agents-hack-john-deere-shark-ninja-cummins",
       "title": "The Verge · 10-19 21:58"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/tdcasw4dwhyzne3hxo3dmizmz"
      }
     ],
     "source": "The Verge · 10-19 21:58"
    },
    {
     "id": "5a55e95dc0391c33",
     "canonical_url": "https://engineering.fb.com/2026/10/19/ml-applications/thin-orchestrator-ai-agent-architecture/",
     "url": "https://engineering.fb.com/2026/10/19/ml-applications/thin-orchestrator-ai-agent-architecture/",
     "title": "瘦编排器：让路由器少干活，我们的 AI agent 反而更可靠",
     "summary": "知识塞得越多越容易串台：不靠逐个任务手写提示时，单个全能 agent 的成功率只有 10%–20%；Meta 的解法是让中枢只负责派活和可靠性检查、自身不做任何领域判断，把领域推理交给 14 个专职 agent。",
     "why": "多 agent 的价值不在数量，而在隔离上下文：编排层越“笨”、越确定，就越好测试和排错。在做企业内部 agent 的团队，可以先把路由和校验从领域推理里拆出来。",
     "tags": [
      "多 Agent",
      "编排"
     ],
     "section": "top3",
     "published_at": "2026-10-19T23:30:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://engineering.fb.com/2026/10/19/ml-applications/thin-orchestrator-ai-agent-architecture/",
       "title": "Engineering at Meta · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/qh6pvehounuylklti7i0zl124"
      }
     ],
     "source": "Engineering at Meta · 10-19"
    },
    {
     "id": "fdff59609d837633",
     "canonical_url": "https://alignment.anthropic.com/2026/graft/",
     "url": "https://alignment.anthropic.com/2026/graft/",
     "title": "通过定位而非惩罚坏行为来减少奖励作弊",
     "summary": "",
     "why": "靠监控器扣分来压制作弊，会把模型训练得更会躲监控；GRAFT 改为把被标记轨迹的更新关进一组训练后整组删除的参数里。目前只在小模型和简单监控器上验证，但给“监控不可靠时怎么训练 agent”提供了新思路。",
     "tags": [
      "对齐",
      "强化学习"
     ],
     "section": "top3",
     "published_at": "2026-10-19T08:00:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://alignment.anthropic.com/2026/graft/",
       "title": "Anthropic Alignment · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/vyw8fvwn6v0st03h0e5llhoji"
      }
     ],
     "source": "Anthropic Alignment · 10-19"
    }
   ]
  },
  {
   "id": "edge",
   "title": "专栏：端侧模型",
   "intro": "本期看点：Mac 上可自训投机解码草稿模型，蚂蚁一次开源 5 个语音模型。",
   "items": [
    {
     "id": "07ec78d8b2d02a92",
     "canonical_url": "https://x.com/Prince_Canuma/status/2112210315303940538",
     "url": "https://x.com/Prince_Canuma/status/2112210315303940538",
     "title": "@Prince_Canuma: Mac 本地训练 drafter",
     "summary": "",
     "why": "投机解码用的草稿模型（drafter）现在能在 Mac 上用 MLX-VLM 自己训练。给本地模型配一个贴合自己负载的 drafter，是在 Apple Silicon 上提速解码最直接的办法。",
     "tags": [
      "MLX",
      "投机解码"
     ],
     "section": "edge",
     "published_at": "2026-10-19T23:52:36+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/Prince_Canuma/status/2112210315303940538",
       "title": "X · 10-19 23:52"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "X · 10-19 23:52"
    },
    {
     "id": "5840e2a190b4eb5d",
     "canonical_url": "https://x.com/AntLingAGI/status/2112210811997540475",
     "url": "https://x.com/AntLingAGI/status/2112210811997540475",
     "title": "@AntLingAGI: Ming-3.0 语音模型开源",
     "summary": "",
     "why": "一次放出语音合成、语音理解和语音识别三条线共 5 个模型，最小的只有 0.5B，最大的 126B 总参数、6.5B 激活；理解模型还支持工具调用。要做端侧语音 agent 的，可以按设备算力挑对应尺寸。",
     "tags": [
      "语音模型",
      "开源"
     ],
     "section": "edge",
     "published_at": "2026-10-19T23:54:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/AntLingAGI/status/2112210811997540475",
       "title": "X · 10-19 23:54"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/uqinrni7f70ef46rykfx1n4ia"
      }
     ],
     "source": "X · 10-19 23:54"
    },
    {
     "id": "cdd1752140a7a4df",
     "canonical_url": "https://blog.n8n.io/n8n-nvidia-dgx-spark/",
     "url": "https://blog.n8n.io/n8n-nvidia-dgx-spark/",
     "title": "在 NVIDIA DGX Spark 上运行 n8n AI 工作流，无需云 API",
     "summary": "",
     "why": "一台 128GB 统一内存的 DGX Spark 上同时跑七个本地模型，通过一个 OpenAI 兼容端点驱动整套工作流、全程不调云 API。数据不能出内网的小团队，可以照着它的内存分配和并发数据规划私有化部署。",
     "tags": [
      "本地部署",
      "工作流"
     ],
     "section": "edge",
     "published_at": "2026-10-19T21:22:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://blog.n8n.io/n8n-nvidia-dgx-spark/",
       "title": "n8n Blog · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/ezjbo6ic6xfvpyxyghjekaude"
      }
     ],
     "source": "n8n Blog · 10-19"
    }
   ],
   "note": "其余为小版本：llama.cpp b11722–b11741 构建。"
  },
  {
   "id": "col",
   "title": "专栏：Agent 记忆与协作",
   "intro": "个人与团队 Agent 的记忆、上下文共享与协作方式。",
   "items": [
    {
     "id": "d32b1681ebb7754d",
     "canonical_url": "https://www.oreilly.com/radar/ai-changed-how-we-code-it-hasnt-changed-how-we-work-together/",
     "url": "https://www.oreilly.com/radar/ai-changed-how-we-code-it-hasnt-changed-how-we-work-together/",
     "title": "AI 改变了我们写代码的方式，却没改变我们协作的方式",
     "summary": "",
     "why": "AI 会议纪要大多停在一封邮件里，只对参会的人有用。团队真正缺的，是“为什么这么定”能跟着任务跨团队、跨时区流动，这也是团队 agent 共享上下文要解决的难题。",
     "tags": [
      "团队上下文",
      "会议记忆"
     ],
     "section": "col",
     "published_at": "2026-10-19T18:54:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://www.oreilly.com/radar/ai-changed-how-we-code-it-hasnt-changed-how-we-work-together/",
       "title": "O'Reilly Radar · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/pqqqv0uwnqqzeeq18yb745yeh"
      }
     ],
     "source": "O'Reilly Radar · 10-19"
    },
    {
     "id": "afc5e1707712ba78",
     "canonical_url": "https://huggingface.co/tencent/Hy-Memory-30B-A3B-GGUF",
     "url": "https://huggingface.co/tencent/Hy-Memory-30B-A3B-GGUF",
     "title": "tencent/Hy-Memory-30B-A3B-GGUF：腾讯混元长期记忆写入模型的 GGUF 量化版",
     "summary": "",
     "why": "它只负责记忆框架里的“写入”这一步，可直接替换 Mem0、MemOS 里原本调用的通用大模型；激活约 3B，Q4 量化版约 18GB，本地就能跑。把记忆写入交给专用小模型，长期记忆的成本和隐私都更可控。",
     "tags": [
      "Agent 记忆",
      "本地运行"
     ],
     "section": "col",
     "published_at": "2026-10-19T15:52:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://huggingface.co/tencent/Hy-Memory-30B-A3B-GGUF",
       "title": "Hugging Face · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Hugging Face · 10-19"
    }
   ],
   "note": "长程 agent 的上下文组织见论文栏 CarryOn；Meta 的多 agent 编排经验见今日三条。"
  },
  {
   "id": "cl",
   "title": "主流产品 Changelog 精读",
   "intro": "窗口内只列值得注意的条目，安静的产品略过。",
   "items": [
    {
     "id": "015d2ff33d5d3e85",
     "canonical_url": "https://x.com/ChatGPT/status/2112265233192657153",
     "url": "https://x.com/ChatGPT/status/2112265233192657153",
     "title": "@ChatGPT: Slides 实时协作与编辑",
     "summary": "",
     "why": "幻灯片现在能分享给别人实时共同编辑、留评论，还能让 ChatGPT 把评论直接改进稿子，文字、图表和图片也能点了就改。它在往“协作文档”方向走，而不只是一次性生成。",
     "tags": [
      "ChatGPT"
     ],
     "section": "cl",
     "published_at": null,
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/ChatGPT/status/2112265233192657153",
       "title": null
      },
      {
       "title": "ChatGPT changelog",
       "url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/i5ongeb1uk4uox5z3rsf72if0"
      }
     ],
     "product": "ChatGPT"
    },
    {
     "id": "ad557c270c7c5a46",
     "canonical_url": "https://x.com/reach_vb/status/2112167632493269243",
     "url": "https://x.com/reach_vb/status/2112167632493269243",
     "title": "@reach_vb: ChatGPT Chrome 扩展",
     "summary": "",
     "why": "装上 Chrome 扩展后，ChatGPT 可以基于已打开的标签页和已登录的网站做跨标签研究；Work 模式还能填表、下载文件、在后台运行，所有付费档都能用。浏览器 agent 的门槛从装桌面应用降到装一个扩展。",
     "tags": [
      "ChatGPT"
     ],
     "section": "cl",
     "published_at": null,
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/reach_vb/status/2112167632493269243",
       "title": null
      },
      {
       "title": "ChatGPT changelog",
       "url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/e736dog3ts07bvc84g5gov8lx"
      }
     ],
     "product": "ChatGPT"
    }
   ],
   "note": "其余为小版本或预览构建：Gemini CLI v0.66.0 nightly、Codex rust-v0.164.0 alpha、Qwen Code v0.25.3 及 SDK、langgraph-cli 0.4.37。本期为补发，ChatGPT、Claude、Grok 等网页版更新日志没有当天的抓取。"
  },
  {
   "id": "gh",
   "title": "GitHub 热门",
   "intro": "AI / Agent 相关的 Trending 仓库。",
   "items": [],
   "note": "本期为补发：GitHub Trending 只有 10-22 当天的抓取，无法还原当天榜单，本栏从缺。"
  },
  {
   "id": "eng",
   "title": "Agent 工程",
   "items": [
    {
     "id": "759f54f88765fca3",
     "canonical_url": "https://blogs.microsoft.com/blog/2026/10/19/securing-the-agentic-insider/",
     "url": "https://blogs.microsoft.com/blog/2026/10/19/securing-the-agentic-insider/",
     "title": "应对“agent 型内部人员”的安全问题",
     "summary": "",
     "why": "微软的判断是：企业里大量部署、权限过大的 agent 就是新的“内部人员”，前沿模型无论出自哪家都不能默认信任，需要“守护 agent”在每个动作执行前实时审查。给 agent 授权时，应按内部员工的标准做最小权限和持续审计。",
     "tags": [
      "Agent 安全",
      "权限"
     ],
     "section": "eng",
     "published_at": "2026-10-19T21:00:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://blogs.microsoft.com/blog/2026/10/19/securing-the-agentic-insider/",
       "title": "Microsoft Blog · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/rbr1w8zq9iemkglglp72iuthp"
      }
     ],
     "source": "Microsoft Blog · 10-19"
    }
   ]
  },
  {
   "id": "infra",
   "title": "检索与基础设施",
   "items": [
    {
     "id": "f8f5bd3a08940cd1",
     "canonical_url": "https://vllm.ai/blog/2026-10-20-vllm-torchtpu",
     "url": "https://vllm.ai/blog/2026-10-20-vllm-torchtpu",
     "title": "推出 vLLM TorchTPU",
     "summary": "",
     "why": "对已经在 GPU 上用 vLLM 的团队，换到 TPU 的迁移成本大幅下降：多数模型不改代码就能跑，命令行和 OpenAI 兼容接口照旧，长上下文的 agent 流量是重点优化对象。算力选型多了一个现实选项。",
     "tags": [
      "推理",
      "TPU"
     ],
     "section": "infra",
     "published_at": "2026-10-20T00:00:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://vllm.ai/blog/2026-10-20-vllm-torchtpu",
       "title": "vLLM Blog · 10-20"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "vLLM Blog · 10-20"
    },
    {
     "id": "6f69c417ed46e29d",
     "canonical_url": "https://pytorch.org/blog/introducing-torchneuron-native-pytorch-backend-for-aws-trainium/",
     "url": "https://pytorch.org/blog/introducing-torchneuron-native-pytorch-backend-for-aws-trainium/",
     "title": "推出 TorchNeuron：AWS Trainium 的原生 PyTorch 后端",
     "summary": "",
     "why": "Trainium 成为 PyTorch 的标准设备后，训练和推理库不必再为它单独适配；AWS 称会在 10 月 20 日的 PyTorch 大会上开源。用 AWS 算力的团队，换芯片的代码成本会明显降低。",
     "tags": [
      "训练",
      "Trainium"
     ],
     "section": "infra",
     "published_at": "2026-10-19T20:45:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://pytorch.org/blog/introducing-torchneuron-native-pytorch-backend-for-aws-trainium/",
       "title": "PyTorch Blog · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/szf3v3re4888fzekwoohhu8b4"
      }
     ],
     "source": "PyTorch Blog · 10-19"
    },
    {
     "id": "818671a99ca03138",
     "canonical_url": "https://www.marktechpost.com/2026/10/18/nvidia-runai-open-source-gpu-orchestration-kubernetes/",
     "url": "https://www.marktechpost.com/2026/10/18/nvidia-runai-open-source-gpu-orchestration-kubernetes/",
     "title": "NVIDIA 以 Apache 2.0 开源 Run:ai，用于 Kubernetes GPU 编排",
     "summary": "",
     "why": "GPU 调度、配额和工作负载管理这一层现在可以免订阅自托管、按需修改。自建 Kubernetes GPU 集群的团队，多了一个经过生产检验的调度方案。",
     "tags": [
      "GPU 调度",
      "开源"
     ],
     "section": "infra",
     "published_at": "2026-10-19T11:19:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://www.marktechpost.com/2026/10/18/nvidia-runai-open-source-gpu-orchestration-kubernetes/",
       "title": "MarkTechPost · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/t4k9stbchn7bkvp9w31x3v4a5"
      }
     ],
     "source": "MarkTechPost · 10-19"
    },
    {
     "id": "bebd537ed004d584",
     "canonical_url": "https://modal.com/blog/serverless-routing-made-sticky",
     "url": "https://modal.com/blog/serverless-routing-made-sticky",
     "title": "让无服务器路由也能“粘住”会话",
     "summary": "",
     "why": "语音 agent、流式转写这类会话型负载，需要每次请求都回到同一个容器；Modal 把会话粘性做进无服务器路由，同时按活跃会话数自动扩缩。跑有状态 agent 的团队，不必再用单容器硬扛。",
     "tags": [
      "Serverless",
      "有状态 Agent"
     ],
     "section": "infra",
     "published_at": "2026-10-20T00:00:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://modal.com/blog/serverless-routing-made-sticky",
       "title": "Modal Blog · 10-20"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/s1spqur4ws02avzig36blr8of"
      }
     ],
     "source": "Modal Blog · 10-20"
    },
    {
     "id": "d73a613bf03e3491",
     "canonical_url": "https://engineering.fb.com/2026/10/19/data-center-engineering/inside-prometheus-meta-gigawatt-scale-supercluster/",
     "url": "https://engineering.fb.com/2026/10/19/data-center-engineering/inside-prometheus-meta-gigawatt-scale-supercluster/",
     "title": "走进 Prometheus：Meta 迈向吉瓦级超级集群之路",
     "summary": "",
     "why": "Meta 首次详细披露这座吉瓦级集群：第一期已建成并用于训练 Muse 系列模型，目标规模约 100 万块 GPU。前沿模型的竞争越来越落在电力、散热和集群可靠性这些工程能力上。",
     "tags": [
      "算力",
      "数据中心"
     ],
     "section": "infra",
     "published_at": "2026-10-19T21:00:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://engineering.fb.com/2026/10/19/data-center-engineering/inside-prometheus-meta-gigawatt-scale-supercluster/",
       "title": "Engineering at Meta · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/ebl13m9gich6kfjixzvfiw5bq"
      }
     ],
     "source": "Engineering at Meta · 10-19"
    }
   ]
  },
  {
   "id": "papers",
   "title": "论文",
   "intro": "取自 HF 每日论文 10-19 那期（窗口内新上榜，arXiv 提交日期更早，按上榜时间计）中与 agent 相关的几篇。",
   "items": [
    {
     "id": "768211190d781c4e",
     "canonical_url": "https://arxiv.org/abs/2610.21073",
     "url": "https://arxiv.org/abs/2610.21073",
     "title": "跳过压缩器：编码 agent 自己挑整文件时能修好更多 issue",
     "summary": "",
     "why": "压缩会改掉代码原文，检索又常漏掉要改的文件，两种失败都会让补丁打不上。给自己的 agent 设计上下文时，先保证要改的文件原样、完整，再谈省 token。",
     "tags": [
      "编码 Agent",
      "上下文"
     ],
     "section": "papers",
     "published_at": "2026-10-19T08:00:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.21073",
       "title": "HF 每日论文 · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "HF 每日论文 · 10-19"
    },
    {
     "id": "a897f73e39fbac23",
     "canonical_url": "https://arxiv.org/abs/2610.18784",
     "url": "https://arxiv.org/abs/2610.18784",
     "title": "轻装上阵，CarryOn：为长程 agent 构建面向对象的上下文",
     "summary": "",
     "why": "与其反复压缩“怎么走到这一步”的历史，不如围绕已经完成的进展来组织上下文：作者称在 OSWorld 2.0 上效果提升 5.8%、总 token 用量少一半多。长程 agent 的上下文管理，可以从“摘要历史”转向“维护任务状态”。",
     "tags": [
      "长程 Agent",
      "上下文管理"
     ],
     "section": "papers",
     "published_at": "2026-10-19T08:00:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.18784",
       "title": "HF 每日论文 · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "HF 每日论文 · 10-19"
    },
    {
     "id": "c982c9b0b7ebaef3",
     "canonical_url": "https://arxiv.org/abs/2610.20424",
     "url": "https://arxiv.org/abs/2610.20424",
     "title": "BizAgentBench：揭示长程 agent 在真实业务落地中的隐藏挑战",
     "summary": "",
     "why": "只看最后是否完成，会掩盖过程里的违规：在这套源自一家头部支付公司真实流程的基准上，Claude Opus 5 的严格成功率也只有 64%。把 agent 放进支付这类高风险流程前，评估要覆盖中间步骤是否合规。",
     "tags": [
      "Agent 评测",
      "业务流程"
     ],
     "section": "papers",
     "published_at": "2026-10-19T08:00:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.20424",
       "title": "HF 每日论文 · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "HF 每日论文 · 10-19"
    },
    {
     "id": "84bd64fc85e5c762",
     "canonical_url": "https://arxiv.org/abs/2610.23787",
     "url": "https://arxiv.org/abs/2610.23787",
     "title": "DLite：面向 agent 的无历史投机解码",
     "summary": "",
     "why": "agent 场景上下文越拉越长、轮次频繁，现有草稿模型的接受长度最多会掉 80%；DLite 干脆不维护历史，直接复用目标模型的隐藏状态做局部预测。给 agent 做推理加速，投机解码要针对长对话重新设计。",
     "tags": [
      "投机解码",
      "推理加速"
     ],
     "section": "papers",
     "published_at": "2026-10-19T08:00:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.23787",
       "title": "HF 每日论文 · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "HF 每日论文 · 10-19"
    },
    {
     "id": "f2ac90f65e06d654",
     "canonical_url": "https://arxiv.org/abs/2610.07898",
     "url": "https://arxiv.org/abs/2610.07898",
     "title": "FC-SWE：面向长程软件工程 agent 的失败条件强化学习",
     "summary": "",
     "why": "失败的补丁和测试反馈本身就是训练信号：把它们当成下一次尝试的上下文纳入强化学习，4B 小模型在 SWE-bench Verified 上两次内的解决率从 48.5% 提到 52.8%。训练编码 agent 时，别把失败轨迹直接丢掉。",
     "tags": [
      "编码 Agent",
      "强化学习"
     ],
     "section": "papers",
     "published_at": "2026-10-19T08:00:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.07898",
       "title": "HF 每日论文 · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "HF 每日论文 · 10-19"
    },
    {
     "id": "91cd3d2811f7ef10",
     "canonical_url": "https://arxiv.org/abs/2610.23194",
     "url": "https://arxiv.org/abs/2610.23194",
     "title": "OfficeTown：用生成式组织推动 agent 智能演进",
     "summary": "",
     "why": "用模拟的“公司”生成 agent 训练数据：任务、权限和历史决策都来自持续运转的组织，比孤立任务更接近真实工作。作者称用这些数据微调的 9B 模型在 GDPval 等外部基准上明显提升。",
     "tags": [
      "Agent 训练数据",
      "模拟环境"
     ],
     "section": "papers",
     "published_at": "2026-10-19T08:00:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.23194",
       "title": "HF 每日论文 · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "HF 每日论文 · 10-19"
    }
   ]
  },
  {
   "id": "biz",
   "title": "大厂与模型",
   "items": [
    {
     "id": "7d7ee730e6105e8b",
     "canonical_url": "https://the-decoder.com/stripe-says-ai-agent-purchases-are-adding-up-but-it-wont-share-the-numbers/",
     "url": "https://the-decoder.com/stripe-says-ai-agent-purchases-are-adding-up-but-it-wont-share-the-numbers/",
     "title": "Stripe 称 AI agent 代购交易在增长，但不愿公布具体数字",
     "summary": "Stripe 让 agent 代购落地的办法是给 agent 一张单独的虚拟卡号、由用户设消费上限；接入 Link 的 agent 已从约 70 个增至约 200 个，但它只给增长倍数、不给绝对规模，起点很低，真实体量仍看不清。",
     "why": "agent 支付正在成为新的基础设施之争，PayPal 和 Shopify 的 Shop Pay 也在跟进。做购物、订票类 agent 的，优先接入能限额、不暴露真实卡号的支付方式。",
     "tags": [
      "Agent 支付",
      "电商"
     ],
     "section": "biz",
     "published_at": "2026-10-19T20:51:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://the-decoder.com/stripe-says-ai-agent-purchases-are-adding-up-but-it-wont-share-the-numbers/",
       "title": "The Decoder · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/f6cnyr9kgnmhd8kmz5o4r4xtm"
      }
     ],
     "source": "The Decoder · 10-19"
    },
    {
     "id": "cb7726337a7c64de",
     "canonical_url": "https://factory.com/news/databricks-partnership",
     "url": "https://factory.com/news/databricks-partnership",
     "title": "Factory 与 Databricks：打造由数据驱动的软件",
     "summary": "",
     "why": "大型数据平台迁移最难的往往不是表，而是上面依赖这些数据的上百个应用；这次合作把数据迁移和应用改写放进同一个项目，由编码 agent 承担应用侧的重建。企业级编码 agent 的落地点正从写新代码转向搬旧系统。",
     "tags": [
      "编码 Agent",
      "企业迁移"
     ],
     "section": "biz",
     "published_at": "2026-10-19T08:00:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://factory.com/news/databricks-partnership",
       "title": "Factory · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/ixosv1z032fmdmpq1o31yhyic"
      }
     ],
     "source": "Factory · 10-19"
    },
    {
     "id": "ef736b2c39c4d865",
     "canonical_url": "https://www.marktechpost.com/2026/10/18/mistral-timestral-time-series-foundation-model/",
     "url": "https://www.marktechpost.com/2026/10/18/mistral-timestral-time-series-foundation-model/",
     "title": "Mistral AI 发布 Timestral：2.09 亿参数、表现超过 Google TimesFM-3 的预测模型家族",
     "summary": "",
     "why": "这是 Mistral 第一次做时序基础模型，四个尺寸从 5M 到 209M，主打零样本预测。需求预测、监控告警这类任务多了一个可比的基线；榜单成绩为 Mistral 自报，落地前要用自己的数据复测。",
     "tags": [
      "时序模型",
      "开源模型"
     ],
     "section": "biz",
     "published_at": "2026-10-19T09:18:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://www.marktechpost.com/2026/10/18/mistral-timestral-time-series-foundation-model/",
       "title": "MarkTechPost · 10-19"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/vd68ub9lfujnxxim0cwkga917"
      }
     ],
     "source": "MarkTechPost · 10-19"
    },
    {
     "id": "8c169bdd3cfe449f",
     "canonical_url": "https://x.com/thexpin/status/2112121758811562346",
     "url": "https://x.com/thexpin/status/2112121758811562346",
     "title": "@thexpin: DeepSeek 成立产品部",
     "summary": "",
     "why": "据称 DeepSeek 把 Harness、Web、App 和平台团队合并成产品部，由原 Harness 团队负责人带队，还在招产品设计师。一向以模型见长的 DeepSeek 开始补消费端体验，agent 产品形态会是它下一步的重点。",
     "tags": [
      "DeepSeek",
      "组织调整"
     ],
     "section": "biz",
     "published_at": "2026-10-19T18:00:00+08:00",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/thexpin/status/2112121758811562346",
       "title": "X · 10-19 18:00"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/hdymmkxhqhudn0cb7fi0kmsqp"
      }
     ],
     "source": "X · 10-19 18:00"
    }
   ],
   "note": "加密：窗口内无值得收录的 AI × 加密动态。"
  },
  {
   "id": "x",
   "title": "X 精选",
   "items": [
    {
     "id": "98e97b6511d9ae53",
     "canonical_url": "https://x.com/emollick/status/2112030115752456456",
     "url": "https://x.com/emollick/status/2112030115752456456",
     "title": "@emollick: 自动化早在做管理决策",
     "summary": "",
     "why": "他反驳“计算机不能做管理决策”这张常被引用的幻灯片：交易、物流、客服、保险里的自动化系统一直在做决策，假装没有只会模糊责任。部署 agent 时，与其争论能不能让它决策，不如先写清出了错谁负责。",
     "tags": [
      "Agent 与工作方式"
     ],
     "section": "x",
     "published_at": "2026-10-19",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/emollick/status/2112030115752456456",
       "title": "@emollick · 10-19 11:56 · ♥ 1157"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 1157
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/d3aeg55owdcjj3bf1ur71oe21"
      }
     ],
     "source": "@emollick · 10-19 11:56 · ♥ 1157"
    },
    {
     "id": "1e425438d3f2649b",
     "canonical_url": "https://x.com/emollick/status/2112165213202157907",
     "url": "https://x.com/emollick/status/2112165213202157907",
     "title": "@emollick: STATA 技能正被淘汰",
     "summary": "",
     "why": "他说写好 STATA 代码曾是自己学术生涯的关键，如今这项技能基本过时，并提醒还把 AI 当聊天机器人的学者：一年里变化很大。研究者的竞争力，正从会写分析代码转向会指挥 agent 做分析。",
     "tags": [
      "Agent 与工作方式"
     ],
     "section": "x",
     "published_at": "2026-10-19",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/emollick/status/2112165213202157907",
       "title": "@emollick · 10-19 20:53 · ♥ 439"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 439
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/kj5yboipm9173hc965tc3xttx"
      }
     ],
     "source": "@emollick · 10-19 20:53 · ♥ 439"
    },
    {
     "id": "123137e77db911e5",
     "canonical_url": "https://x.com/mattpocockuk/status/2112194888511222205",
     "url": "https://x.com/mattpocockuk/status/2112194888511222205",
     "title": "@mattpocockuk: agent 翻出漏读回复",
     "summary": "",
     "why": "他漏看了 Michael Feathers 的回复，是自己的 agent 整理消息时翻了出来。个人 agent 最先显出价值的场景之一，就是替人捞回被信息洪流淹没的重要消息。",
     "tags": [
      "Agent 与工作方式"
     ],
     "section": "x",
     "published_at": "2026-10-19",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/mattpocockuk/status/2112194888511222205",
       "title": "@mattpocockuk · 10-19 22:51 · ♥ 264"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 264
     },
     "via": [],
     "source": "@mattpocockuk · 10-19 22:51 · ♥ 264"
    },
    {
     "id": "da396806f7daa30d",
     "canonical_url": "https://x.com/tison1096/status/2112203431486369818",
     "url": "https://x.com/tison1096/status/2112203431486369818",
     "title": "@tison1096: 开源就做基础构件",
     "summary": "",
     "why": "他认为基础设施代码生命周期长、场景可以预先设计，适合做形式化证明；做开源最好做 building blocks，并事先划定范围，打磨才有意义。代码越来越便宜时，边界清晰的底层库反而更值得长期投入。",
     "tags": [
      "开源与工程"
     ],
     "section": "x",
     "published_at": "2026-10-19",
     "first_seen_at": "2026-10-20",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/tison1096/status/2112203431486369818",
       "title": "@tison1096 · 10-19 23:25 · ♥ 79"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 79
     },
     "via": [],
     "source": "@tison1096 · 10-19 23:25 · ♥ 79"
    }
   ]
  }
 ]
}
