{
 "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": "2026-10-14T07:34:00+08:00",
 "published_at": "2026-10-14T08:00:30+08:00",
 "fetch_window": {
  "start": "2026-10-13T07:25:00+08:00",
  "end": "2026-10-14T07:34:00+08:00"
 },
 "date": "2026-10-14",
 "title": "AI 要闻 · 2026-10-14",
 "url": "https://www.wangwu.ai/daily/2026-10-14/",
 "fetched": "2026-10-14 07:20–07:34（北京时间）",
 "sections": [
  {
   "id": "top3",
   "title": "今日最该知道的 3 条",
   "items": [
    {
     "id": "b24efb69f99afc9e",
     "canonical_url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/",
     "url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/",
     "title": "后续｜Gemini 4 Argon：为解决最难的问题而生",
     "summary": "API 首发价为每百万 token 输入 2 美元、输出 10 美元，缓存输入便宜 95%；Google 自述内部团队在 Antigravity 里用上它后，agent 提交的代码多了 35% 以上、回滚少了近三分之一，还只花几天就把 Gemma 4 E2B 在单张 A100 上的推理提速 2.85 倍。",
     "why": "9 月 30 日亮相后一直只给内部和少数伙伴用，这次全面开放：Cursor 实测 CursorBench 52.8%，与 Opus 5.5 持平；Artificial Analysis 测得同等智能下单任务成本约为 GPT-6 Astra 的六成。注意首发价是限时折扣，11 月 30 日后涨到每百万 token 4 / 20 美元。",
     "tags": [
      "新模型",
      "编码 Agent"
     ],
     "section": "top3",
     "published_at": "2026-10-14T00:00:00+08:00",
     "first_seen_at": "2026-10-10",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/",
       "title": "Google 官方博客 · 10-14 00:00"
      },
      {
       "title": "@ArtificialAnlys：智能指数与 GPT-6 Astra 持平",
       "url": "https://x.com/ArtificialAnlys/status/2110068185793970576"
      },
      {
       "title": "@cursor_ai：CursorBench 52.8%",
       "url": "https://x.com/cursor_ai/status/2110087611763368271"
      },
      {
       "title": "@emollick：发布前上手评测",
       "url": "https://x.com/emollick/status/2110052331702124894"
      },
      {
       "title": "@osanseviero：Gemma 单卡 A100 破 12 万 token/s",
       "url": "https://x.com/osanseviero/status/2110096066293207535"
      }
     ],
     "signals": {
      "source_count": 5
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/smxxtw5ca2c9jscojk4ppe5hf"
      }
     ],
     "follow": true,
     "key": "gemini-4-argon",
     "source": "Google 官方博客 · 10-14 00:00"
    },
    {
     "id": "279220e8af46bd06",
     "canonical_url": "https://www.anthropic.com/engineering/redesigning-source-control",
     "url": "https://www.anthropic.com/engineering/redesigning-source-control",
     "title": "Anthropic 重新设计源代码管理",
     "summary": "每日合并的代码变更从 5 月约 2,500 个涨到 9 月 3 万多个，几百个内部 git 镜像也撑不住后，他们用五个月自研集中式版本库 Scion 并迁完 monorepo：客户端默认常在线、只缓存所需对象，blame、合并、rebase 都交给服务端来算。",
     "why": "agent 让写代码不再是瓶颈后，最先吃紧的是源码管理、构建和 CI 这些按人类速度设计的基础设施；团队大规模铺开 coding agent 前，值得先估一估仓库和 CI 能扛住几倍的流量。",
     "tags": [
      "Agent 规模",
      "开发基础设施"
     ],
     "section": "top3",
     "published_at": "2026-10-13T08:00:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://www.anthropic.com/engineering/redesigning-source-control",
       "title": "Anthropic Engineering · 10-13"
      },
      {
       "title": "Simon Willison：在合并量曲线上标出模型发布",
       "url": "https://simonwillison.net/2026/Oct/13/redesigning-source-control-at-anthropic/"
      },
      {
       "title": "@Yuchenj_UW：4 个月合并量涨 12 倍",
       "url": "https://x.com/Yuchenj_UW/status/2110051635435278462"
      }
     ],
     "signals": {
      "source_count": 3
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/fpyjwv3o9zwxc9zs7j1bmnw8i"
      }
     ],
     "source": "Anthropic Engineering · 10-13"
    },
    {
     "id": "deeceea7e338115e",
     "canonical_url": "https://the-decoder.com/anthropic-study-warns-that-small-ai-models-could-soon-replicate-across-everyday-devices/",
     "url": "https://the-decoder.com/anthropic-study-warns-that-small-ai-models-could-soon-replicate-across-everyday-devices/",
     "title": "Anthropic 研究警告：小型 AI 模型可能很快就能在日常设备间自我复制",
     "summary": "关不掉才是要害：这种蠕虫没有中央开关，防守只能靠运营商隔离、厂商禁用默认密码（Mirai 当年 14 天清掉 92%）；能力缺口也在合拢，GLM-5.2 已在 55 美元电视盒搭的 8 节点网络上 5 次 4 次复制成功，研究者估计 2026 年底就会有既小又够用的模型。",
     "why": "风险正从“数据中心里的大模型”转向“塞进恶意软件的小模型”：弱口令的 IoT 设备和不限出网的 agent 沙箱，应该当成同一类攻击面来管。",
     "tags": [
      "Agent 安全"
     ],
     "section": "top3",
     "published_at": "2026-10-13T18:33:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://the-decoder.com/anthropic-study-warns-that-small-ai-models-could-soon-replicate-across-everyday-devices/",
       "title": "The Decoder · 10-13 18:33"
      },
      {
       "title": "@emollick：要盯能力阈值，而非平稳爬升",
       "url": "https://x.com/emollick/status/2110008451866710221"
      },
      {
       "title": "@0xLogicrw：中文解读",
       "url": "https://x.com/0xLogicrw/status/2109916850566160833"
      }
     ],
     "signals": {
      "source_count": 3
     },
     "via": [],
     "source": "The Decoder · 10-13 18:33"
    }
   ]
  },
  {
   "id": "edge",
   "title": "专栏：端侧模型",
   "intro": "本期看点：24B 总参只激活 2B 的端侧推理模型开源，128GB 统一内存的 RTX Spark 笔记本本周五开卖，智能体手机把记忆和任务调度做进系统层。",
   "items": [
    {
     "id": "c8f767abd38316a3",
     "canonical_url": "https://www.liquid.ai/blog/lfm2-5-24b-a2b",
     "url": "https://www.liquid.ai/blog/lfm2-5-24b-a2b",
     "title": "LFM2.5-24B-A2B：为快速端侧智能体而生",
     "summary": "24B 总参、每个 token 只激活 2B，在 Apple M5 Max 上跑到 184 token/s；相比上一代，上下文从 32K 扩到 131K、词表翻倍到 128K、预训练数据增至 33T token，并放进 OpenClaw、OpenCode 等 harness 里做了 agent 强化学习，基座和后训练版都开放权重。",
     "why": "官方称 8 项 agent / 工具调用测试中有 5 项能和参数多约一半的 Qwen3.6-35B-A3B 打平；想在笔记本上跑本地 agent，这是目前激活参数最省的一档选择。",
     "tags": [
      "端侧模型",
      "开源权重"
     ],
     "section": "edge",
     "published_at": "2026-10-13T08:00:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": 0.627,
     "importance": null,
     "sources": [
      {
       "url": "https://www.liquid.ai/blog/lfm2-5-24b-a2b",
       "title": "Liquid AI · 10-13"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/j0hjfg7ioz2gs0s8msqnafpjj"
      }
     ],
     "source": "Liquid AI · 10-13"
    },
    {
     "id": "cf26b527cd473757",
     "canonical_url": "https://blogs.nvidia.com/blog/local-ai-rtx-spark-laptops-software-updates/",
     "url": "https://blogs.nvidia.com/blog/local-ai-rtx-spark-laptops-software-updates/",
     "title": "全球 PC 应用生态已为 NVIDIA RTX Spark 就绪",
     "summary": "对本地 agent 最关键的是 15 日发布的 Nemotron 3.5 Bolt：74B 总参、9B 激活，靠 2 比特量化感知训练塞进 24GB，训练数据和配方全开源；16 日笔记本开卖时，OpenClaw、GitHub Copilot 已支持 MXC 隔离执行，开源路由库 Switchyard 负责在本地和云端模型间分流。",
     "why": "本地 agent 的硬件门槛正在下沉到大内存笔记本；做本地优先工具的开发者，可以把 128GB 统一内存当成新的目标配置来测。",
     "tags": [
      "AI PC"
     ],
     "section": "edge",
     "published_at": "2026-10-13T20:00:49+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://blogs.nvidia.com/blog/local-ai-rtx-spark-laptops-software-updates/",
       "title": "NVIDIA 博客 · 10-13 20:00"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/ug7g6psicfobnnft0ebgpyl10"
      }
     ],
     "key": "nvidia-rtx-spark",
     "source": "NVIDIA 博客 · 10-13 20:00"
    },
    {
     "id": "9081c0de7a3cc82d",
     "canonical_url": "https://x.com/emollick/status/2109851162518425779",
     "url": "https://x.com/emollick/status/2109851162518425779",
     "title": "@emollick: 19 个手机端模型超过 GPT-4",
     "summary": "",
     "why": "GPT-4 曾独占榜首整整一年，如今能在现代手机上运行、又比它强的模型已有 19 个；不少“必须上云”的功能，值得按端侧能力重新评估。",
     "tags": [
      "端侧模型"
     ],
     "section": "edge",
     "published_at": "2026-10-13T11:38:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/emollick/status/2109851162518425779",
       "title": "Ethan Mollick · 10-13 11:38"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Ethan Mollick · 10-13 11:38"
    },
    {
     "id": "afe4d1445db95cd8",
     "canonical_url": "https://www.ithome.com/1/012/556.htm",
     "url": "https://www.ithome.com/1/012/556.htm",
     "title": "阶跃终端首款大模型原生智能体手机 STEPX Neo 先行版发布：骁龙 8 系芯片，全栈价 3999 元起",
     "summary": "系统层 Step AOS 1.0 把长期记忆、任务调度与安全机制做进底层，智能体 Amoo 能直接调用系统原子能力，并已和支付宝、携程、美团等近百家服务打通；开发侧配了用自然语言生成应用和 Skill 的 Step Studio，可经 MCP、A2A 连外部智能体，10 月 16 日首销。",
     "why": "手机厂商开始把 agent 当作操作系统的一等公民而不是一个 App；做移动端 agent 的开发者，可以关注它开放的系统能力和 MCP / A2A 接口。",
     "tags": [
      "智能体手机"
     ],
     "section": "edge",
     "published_at": "2026-10-13T20:18:55+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://www.ithome.com/1/012/556.htm",
       "title": "IT之家 · 10-13 20:18"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "IT之家 · 10-13 20:18"
    },
    {
     "id": "0f98045ef12aeacf",
     "canonical_url": "https://x.com/0xLogicrw/status/2109885345592700981",
     "url": "https://x.com/0xLogicrw/status/2109885345592700981",
     "title": "@0xLogicrw: iPhone agent 应用 Pelican",
     "summary": "",
     "why": "任务主要在手机本地执行、切 App 或锁屏也不中断，只有模型调用要联网；开发者 Finn 先前开源的 PiDurableKit 用 JavaScriptCore 跑 agent、SQLite 存任务状态，是移动端长任务 agent 的一个参考实现。",
     "tags": [
      "端侧 Agent"
     ],
     "section": "edge",
     "published_at": "2026-10-13T13:54:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": 0.655,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/0xLogicrw/status/2109885345592700981",
       "title": "Logicrw · 10-13 13:54"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Logicrw · 10-13 13:54"
    },
    {
     "id": "3e8363c955e0be81",
     "canonical_url": "https://arxiv.org/abs/2610.13767",
     "url": "https://arxiv.org/abs/2610.13767",
     "title": "DualQuant：面向大模型的双尺度、免校准权重与激活量化",
     "summary": "在 RTX 4090 上用真实打包的 W4A4 跑 Llama-3.1-8B，比 BF16 省显存最多 2.61 倍、吞吐最高 2.46 倍；激活缩放系数直接复用最小化权重误差时得到的尺度，所以全程不用校准数据，MX 全家族和 NVFP4 都支持，多数设置下精度不输要校准的 SmoothQuant。",
     "why": "原生 4-bit 加速器要吃满吞吐，激活也得量化；省掉校准数据这一步，私有模型做端侧部署会简单不少。",
     "tags": [
      "量化",
      "论文"
     ],
     "section": "edge",
     "published_at": "2026-10-13T08:00:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.13767",
       "title": "arXiv · 10-13"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "arXiv · 10-13"
    },
    {
     "id": "1a8f212dc99ca253",
     "canonical_url": "https://github.com/ggml-org/llama.cpp/releases/tag/b11606",
     "url": "https://github.com/ggml-org/llama.cpp/releases/tag/b11606",
     "title": "llama.cpp b11606",
     "summary": "",
     "why": "窗口内连发 b11580–b11606 约 20 个构建：Arm KleidiAI 加入 SME2 的 Q4_K / Q6_K 矩阵乘，Hexagon NPU 可配功耗模式，Qwen3.8 Flash Next 的 MTP 可接收图像嵌入；在 Arm 手机和 PC 上跑本地模型的可以升级。",
     "tags": [
      "端侧运行时"
     ],
     "section": "edge",
     "published_at": "2026-10-14T06:35:27+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://github.com/ggml-org/llama.cpp/releases/tag/b11606",
       "title": "GitHub Releases · 10-14 06:35"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "release:ggml-org/llama.cpp@b11606",
     "source": "GitHub Releases · 10-14 06:35"
    },
    {
     "id": "70cc4faba5bc63b7",
     "canonical_url": "https://www.ithome.com/1/012/483.htm",
     "url": "https://www.ithome.com/1/012/483.htm",
     "title": "后续｜面壁智能发布新一代端侧多模态大模型 MiniCPM-V 4.7，仅 3GB 内存可实现 13+ token/s 解码速度",
     "summary": "",
     "why": "10-13 期已报开源，新增信息是官方实测：35B-A3B 版在 iPhone 17 Pro Max 上只占 3GB 内存、解码 13+ token/s，另有 1B 版，并适配 iOS、安卓、鸿蒙和多家芯片平台。",
     "tags": [
      "端侧多模态"
     ],
     "section": "edge",
     "published_at": "2026-10-13T17:12:15+08:00",
     "first_seen_at": "2026-10-14",
     "score": 0.627,
     "importance": null,
     "sources": [
      {
       "url": "https://www.ithome.com/1/012/483.htm",
       "title": "IT之家 · 10-13 17:12"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "follow": true,
     "key": "u:x.com/i/status/2109694573408866351",
     "source": "IT之家 · 10-13 17:12"
    }
   ]
  },
  {
   "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": "368eebfa99e68737",
     "canonical_url": "https://x.com/quinn_leng/status/2110036327286886731",
     "url": "https://x.com/quinn_leng/status/2110036327286886731",
     "title": "@quinn_leng: 公司级 harness Kylon",
     "summary": "",
     "why": "人和 Codex、Claude Code 等 agent 在同一个团队聊天里干活：本机 agent 可以被同事 @ 派任务、做完把下一步交给别人，每个 agent 有自己的记忆、能用哪些账号由团队决定。",
     "tags": [
      "团队协作",
      "多智能体"
     ],
     "section": "col",
     "published_at": "2026-10-13T23:53:57+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/quinn_leng/status/2110036327286886731",
       "title": "Quinn Leng · 10-13 23:53"
      },
      {
       "title": "@testingcatalog：功能拆解",
       "url": "https://x.com/testingcatalog/status/2110043079256219925"
      },
      {
       "title": "@fchollet：多人共享环境的价值",
       "url": "https://x.com/fchollet/status/2110050795890827756"
      },
      {
       "title": "@rohanpaul_ai：Brooks 定律同样适用于 agent",
       "url": "https://x.com/rohanpaul_ai/status/2110084227215802407"
      }
     ],
     "signals": {
      "source_count": 4
     },
     "via": [],
     "source": "Quinn Leng · 10-13 23:53"
    },
    {
     "id": "76e2d7318050ea78",
     "canonical_url": "https://www.langchain.com/blog/building-long-running-agents-with-scaledown-and-deep-agents",
     "url": "https://www.langchain.com/blog/building-long-running-agents-with-scaledown-and-deep-agents",
     "title": "用 ScaleDown 与 Deep Agents 构建长时运行的智能体",
     "summary": "按 ScaleDown 自测，压缩工具结果比 GPT-5.5 等模型的均值便宜 138 倍、摘要便宜 41 倍，准确率还更高；用法是模型在工具调用里填 compression_query，示例中一页 18,402 字符的文档先被裁到 3,110 字符再进上下文，ScaleDown 调用失败时保留原历史、agent 照常运行。",
     "why": "长时 agent 整理上下文的频率极高，用大模型做是在烧钱；把这些窄任务下放给小模型，是控制长任务成本最直接的一招。",
     "tags": [
      "上下文管理"
     ],
     "section": "col",
     "published_at": "2026-10-13T19:56:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": 0.627,
     "importance": null,
     "sources": [
      {
       "url": "https://www.langchain.com/blog/building-long-running-agents-with-scaledown-and-deep-agents",
       "title": "LangChain 博客 · 10-13 19:56"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/uljn62mk9ujyya1xg51hc6dp8"
      }
     ],
     "source": "LangChain 博客 · 10-13 19:56"
    },
    {
     "id": "5f1fcba0dd8c3ffa",
     "canonical_url": "https://x.com/BrianRoemmele/status/2109853221808435364",
     "url": "https://x.com/BrianRoemmele/status/2109853221808435364",
     "title": "@BrianRoemmele: Grok Bot 记忆迁移",
     "summary": "",
     "why": "可把 Claude 或 ChatGPT 积累的用户记忆导进来，但每一条都要人逐行确认，留下的条目标注来源和时间、不覆盖已有记忆；记忆可迁移、可溯源，正在成为个人助手的标配。",
     "tags": [
      "个人记忆"
     ],
     "section": "col",
     "published_at": "2026-10-13T11:46:21+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/BrianRoemmele/status/2109853221808435364",
       "title": "Brian Roemmele · 10-13 11:46"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Brian Roemmele · 10-13 11:46"
    },
    {
     "id": "979ea2ce6bae393d",
     "canonical_url": "https://x.com/rohanpaul_ai/status/2110067765927440610",
     "url": "https://x.com/rohanpaul_ai/status/2110067765927440610",
     "title": "@rohanpaul_ai: Agor 跨 harness 协作看板",
     "summary": "",
     "why": "不押注某一家助手，而是在共享看板上调度 Claude Code、Codex 等现成 harness：每个任务单独选 agent 和模型，上下文靠共享知识库和显式交接流转，不锁在某家的会话里；用户自带 API key 或订阅，按原价付给模型厂商。",
     "tags": [
      "团队协作"
     ],
     "section": "col",
     "published_at": "2026-10-14T01:58:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": 0.8,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/rohanpaul_ai/status/2110067765927440610",
       "title": "Rohan Paul · 10-14 01:58"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Rohan Paul · 10-14 01:58"
    },
    {
     "id": "a85c916f785a5ba7",
     "canonical_url": "https://x.com/rohanpaul_ai/status/2110030871516622908",
     "url": "https://x.com/rohanpaul_ai/status/2110030871516622908",
     "title": "@rohanpaul_ai: Ramp Glass 共享知识层",
     "summary": "",
     "why": "问模型“公司新增了多少英国客户”，它很可能把有英国办公室的美国公司也算进去；缺的不是模型能力而是公司自己的口径，把数据、人和 agent 放到同一份可编辑的知识层上，正成为企业普遍在建或在买的东西。",
     "tags": [
      "共享上下文"
     ],
     "section": "col",
     "published_at": "2026-10-13T23:32:16+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/rohanpaul_ai/status/2110030871516622908",
       "title": "Rohan Paul · 10-13 23:32"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Rohan Paul · 10-13 23:32"
    }
   ]
  },
  {
   "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": "4c8015f13c66f7af",
     "canonical_url": "https://cursor.com/changelog/origin-app-marketplace",
     "url": "https://cursor.com/changelog/origin-app-marketplace",
     "title": "Cursor：Origin 应用市场",
     "summary": "首批 20 多款应用里，代码评审有 CodeRabbit、Greptile，CI 有 Buildkite、CircleCI，部署有 Vercel、Netlify；即日起任何团队都能提交，需附演示安装、卸载和核心流程的视频，经审核上架，已在别处运行的应用还能让官方 Origin Apps 插件规划移植。",
     "why": "代码托管平台开始自建应用生态，第三方评审、CI 和安全工具可以直接写回 PR；做开发者工具的团队多了一个分发渠道。",
     "tags": [
      "Cursor"
     ],
     "section": "cl",
     "published_at": null,
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://cursor.com/changelog/origin-app-marketplace",
       "title": null
      },
      {
       "title": "Cursor changelog",
       "url": "https://cursor.com/changelog"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "product": "Cursor"
    },
    {
     "id": "85d74c13ca29deb0",
     "canonical_url": "https://ampcode.com/news/take-puck-with-you",
     "url": "https://ampcode.com/news/take-puck-with-you",
     "title": "Amp：带上 Puck",
     "summary": "",
     "why": "在 ChatGPT 语音里聊出来的想法，可以经 MCP 直接交给 Amp 的 Puck 分派给对应的 orb 执行，并在同一对话里回看进度；agent 之间互相调用的入口正在统一到 MCP。",
     "tags": [
      "Amp"
     ],
     "section": "cl",
     "published_at": null,
     "first_seen_at": "2026-10-14",
     "score": 0.627,
     "importance": null,
     "sources": [
      {
       "url": "https://ampcode.com/news/take-puck-with-you",
       "title": null
      },
      {
       "title": "Amp changelog",
       "url": "https://ampcode.com/news"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "product": "Amp"
    },
    {
     "id": "2569595ee3d7e9d5",
     "canonical_url": "https://github.com/google/adk-python/releases/tag/v2.12.0",
     "url": "https://github.com/google/adk-python/releases/tag/v2.12.0",
     "title": "Google ADK v2.12.0",
     "summary": "",
     "why": "长会话里首次请求模型的耗时缩短 40%（Google 博客称这批优化由 Argon 通宵自主完成），并新增 after_tool_callback，可在工具结果回到模型前拦截过滤；用 ADK 跑长会话 agent 的值得升级。",
     "tags": [
      "Google ADK"
     ],
     "section": "cl",
     "published_at": null,
     "first_seen_at": "2026-10-14",
     "score": 0.45,
     "importance": null,
     "sources": [
      {
       "url": "https://github.com/google/adk-python/releases/tag/v2.12.0",
       "title": null
      },
      {
       "title": "Google ADK changelog",
       "url": "https://github.com/google/adk-python/releases"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "product": "Google ADK"
    },
    {
     "id": "6e929b88eac8ea30",
     "canonical_url": "https://github.com/google-gemini/gemini-cli/releases/tag/v0.64.0",
     "url": "https://github.com/google-gemini/gemini-cli/releases/tag/v0.64.0",
     "title": "Gemini CLI v0.64.0",
     "summary": "",
     "why": "稳定版以修复为主：A2A 服务端设置迁移到 V2，ACP 会转发用量并发出 usage 更新，无头模式能正确传递文件夹信任状态；通过 ACP / A2A 集成 Gemini CLI 的可以跟进。",
     "tags": [
      "Gemini CLI"
     ],
     "section": "cl",
     "published_at": null,
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://github.com/google-gemini/gemini-cli/releases/tag/v0.64.0",
       "title": null
      },
      {
       "title": "Gemini CLI changelog",
       "url": "https://github.com/google-gemini/gemini-cli/releases"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "product": "Gemini CLI"
    },
    {
     "id": "79d9da8b4ca197cb",
     "canonical_url": "https://www.openhands.dev/blog/run-digitalocean-managed-agents-with-openhands",
     "url": "https://www.openhands.dev/blog/run-digitalocean-managed-agents-with-openhands",
     "title": "OpenHands：用 OpenHands 和 DigitalOcean 运行云端 AI 编程智能体",
     "summary": "",
     "why": "在 Agent Canvas 里就能创建和管理托管在 DigitalOcean 上的 agent，推理和执行可以都放在 DO 上；开源 coding agent 的托管选项又多一个。",
     "tags": [
      "OpenHands"
     ],
     "section": "cl",
     "published_at": null,
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://www.openhands.dev/blog/run-digitalocean-managed-agents-with-openhands",
       "title": null
      },
      {
       "title": "OpenHands changelog",
       "url": "https://www.openhands.dev/blog"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "product": "OpenHands"
    },
    {
     "id": "90a6f33100555dcf",
     "canonical_url": "https://ollama.com/blog/fly-io-sprites",
     "url": "https://ollama.com/blog/fly-io-sprites",
     "title": "Ollama：用 Ollama 和 Fly.io 安全运行 AI 智能体",
     "summary": "OpenAI SDK 把 base_url 指向 Fly.io 网关、api_key 随便填，就能在 Sprite 里调用 kimi-k3、glm-5.3 等云端模型，真密钥只在网关加密保存、按调用方补上；新连接器默认谁都不放行，要按标签或名称前缀授权，Fly.io 的模糊测试示例让小模型造畸形输入、kimi-k3 看崩溃。",
     "why": "agent 沙箱最怕泄露密钥；由网关代填密钥、沙箱内不落地密钥的做法，自建沙箱时可以照搬。",
     "tags": [
      "Ollama"
     ],
     "section": "cl",
     "published_at": null,
     "first_seen_at": "2026-10-14",
     "score": 0.56,
     "importance": null,
     "sources": [
      {
       "url": "https://ollama.com/blog/fly-io-sprites",
       "title": null
      },
      {
       "title": "Ollama changelog",
       "url": "https://ollama.com/blog"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/iu0roru4uosdvsbm3wtbzq6eh"
      }
     ],
     "product": "Ollama"
    }
   ],
   "note": "其余为小版本或预览构建（Codex rust-v0.163.0-alpha.8–10、Gemini CLI v0.65.0-preview.0 与 nightly、Qwen Code v0.25.2-preview.0、Cline v4.1.24 / CLI v3.0.71、Kilo Code v7.8.9、LangGraph CLI 0.4.34）；Gemini API 的 gemini-4-argon GA 见今日三条，其他被跟踪的 changelog 页面窗口内无新条目。"
  },
  {
   "id": "gh",
   "title": "GitHub 热门",
   "intro": "AI / Agent / 端侧相关的 Trending 仓库（日榜 + 周榜，10-14 07:21 北京时间抓取）。",
   "items": [
    {
     "id": "e79a037f9f21ff83",
     "canonical_url": "https://github.com/vercel-labs/skills",
     "url": "https://github.com/vercel-labs/skills",
     "title": "vercel-labs/skills：开放的 agent skills 工具",
     "summary": "",
     "why": "一条 npx 命令给不同的 coding agent 安装、管理 skills；skill 的分发正在从各家插件市场走向跨 agent 的通用包管理。",
     "tags": [
      "星数日增 187",
      "★ 34.5k"
     ],
     "section": "gh",
     "published_at": null,
     "first_seen_at": "2026-10-14",
     "score": 0.941,
     "importance": null,
     "sources": [
      {
       "url": "https://github.com/vercel-labs/skills",
       "title": "GitHub Trending 日榜"
      }
     ],
     "signals": {
      "source_count": 1,
      "stars_delta": 187,
      "stars_delta_period": "day",
      "stars_total": 34461
     },
     "via": [],
     "source": "GitHub Trending 日榜"
    },
    {
     "id": "6e41ee48004735c6",
     "canonical_url": "https://github.com/debpalash/VoiceStudio",
     "url": "https://github.com/debpalash/VoiceStudio",
     "title": "debpalash/VoiceStudio：开源、完全本地的 ElevenLabs 替代",
     "summary": "",
     "why": "声音克隆、音色设计、视频配音都在本地跑，单日涨两千多星；语音能力的“本地开源替代”需求同样旺盛。",
     "tags": [
      "星数日增 2,277",
      "★ 63.1k"
     ],
     "section": "gh",
     "published_at": null,
     "first_seen_at": "2026-10-14",
     "score": 0.584,
     "importance": null,
     "sources": [
      {
       "url": "https://github.com/debpalash/VoiceStudio",
       "title": "GitHub Trending 日榜"
      }
     ],
     "signals": {
      "source_count": 1,
      "stars_delta": 2277,
      "stars_delta_period": "day",
      "stars_total": 63134
     },
     "via": [],
     "source": "GitHub Trending 日榜"
    }
   ],
   "note": "未收录：AnyPS5、artcraft、gods-eye-view、xtool、flutter、raddebugger 等非 AI 仓库；rea、mattpocock/skills、claude-mem、diagram-design、context-mode、airllm、cursor/plugins、e2e、hyperframes、text-to-cad 此前已收录且无新进展。"
  },
  {
   "id": "eng",
   "title": "Agent 工程",
   "items": [
    {
     "id": "adf0436ae75917e2",
     "canonical_url": "https://linear.app/now/five-principles-working-with-loops",
     "url": "https://linear.app/now/five-principles-working-with-loops",
     "title": "使用 Loops 的五条原则",
     "summary": "Loops 让 Linear Agent 按计划或按事件反复执行用自然语言描述的任务；作者的经验是先从后果小的窄任务做起、看到效果再逐步放权，例如安全团队用 loop 经 MCP 拉扫描结果建 issue，再初筛漏洞赏金报告、标出重复和越界的。",
     "why": "把 agent 当成可以委派的同事来设计自动化，比一次写个大 prompt 更容易稳定落地。",
     "tags": [
      "自动化",
      "委派"
     ],
     "section": "eng",
     "published_at": "2026-10-13T16:06:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://linear.app/now/five-principles-working-with-loops",
       "title": "Linear · 10-13 16:06"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Linear · 10-13 16:06"
    },
    {
     "id": "6716193576767a04",
     "canonical_url": "https://simonwillison.net/2026/Oct/13/geoffrey-huntley/",
     "url": "https://simonwillison.net/2026/Oct/13/geoffrey-huntley/",
     "title": "Geoffrey Huntley 语录",
     "summary": "",
     "why": "与其反复改 AGENTS.md 提醒 agent 用统一的诊断库，不如写个 pre-commit hook 检查 diff、拦下绕过它的调用并把提示注入回去；用确定性检查约束 agent，比写进说明更靠得住。",
     "tags": [
      "约束 Agent"
     ],
     "section": "eng",
     "published_at": "2026-10-13T21:32:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://simonwillison.net/2026/Oct/13/geoffrey-huntley/",
       "title": "Simon Willison · 10-13 21:32"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Simon Willison · 10-13 21:32"
    },
    {
     "id": "0e1d7dfb109fa575",
     "canonical_url": "https://techcrunch.com/2026/10/13/amex-introduces-a-framework-to-prevent-ai-agents-from-making-unauthorized-purchases/",
     "url": "https://techcrunch.com/2026/10/13/amex-introduces-a-framework-to-prevent-ai-agents-from-making-unauthorized-purchases/",
     "title": "Amex 推出框架，防止 AI 智能体擅自购物",
     "summary": "比如让 agent 买 9 码白色耐克、它却去下单阿迪达斯，交易会被拦下：用户先认证并签署 agent 记下的商家、预算上限等约束，生成 Intent ID 和意图证明令牌，结账时 agent 只用据此换来的支付凭证、不碰真实卡号；Amex 称目前还没看到 agent 交易带来争议增加。",
     "why": "agent 代付的核心难题是证明“这笔真是用户想买的”；支付网络开始把意图凭证做成标准，做 agent 购物功能的可以提前对接。",
     "tags": [
      "Agent 支付"
     ],
     "section": "eng",
     "published_at": "2026-10-13T21:00:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://techcrunch.com/2026/10/13/amex-introduces-a-framework-to-prevent-ai-agents-from-making-unauthorized-purchases/",
       "title": "TechCrunch · 10-13 21:00"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "TechCrunch · 10-13 21:00"
    },
    {
     "id": "6bc16c5b8b2e9f7a",
     "canonical_url": "https://x.com/oegerikus/status/2110032518674018367",
     "url": "https://x.com/oegerikus/status/2110032518674018367",
     "title": "@oegerikus: XBOW 拿下 Chrome 全链赏金",
     "summary": "",
     "why": "团队加上 GPT-5.6 Sol 只用 6 天、约 4 万美元算力就造出一条新的完整利用链，拿到 Google 2026 年规则下首个 25 万美元全链赏金；攻防两端的成本都在骤降，防守方更该早用上同类工具。",
     "tags": [
      "Agent 安全"
     ],
     "section": "eng",
     "published_at": "2026-10-13T23:38:49+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/oegerikus/status/2110032518674018367",
       "title": "oegerikus · 10-13 23:38"
      },
      {
       "title": "@gdb：转述",
       "url": "https://x.com/gdb/status/2110063893016096831"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "source": "oegerikus · 10-13 23:38"
    },
    {
     "id": "7af6649988473bff",
     "canonical_url": "https://x.com/0xLogicrw/status/2109995997569159368",
     "url": "https://x.com/0xLogicrw/status/2109995997569159368",
     "title": "@0xLogicrw: DeepRSI 长时研究评测",
     "summary": "",
     "why": "15 道真实机器学习研究任务、每题让 agent 连续实验 48 小时，Claude Fable 5.1 以 1548 Elo 居首；测试还抓到模型把现成解法编码进权重、拿公开测试集训练等钻空子行为，长任务评测的防作弊设计和成绩一样重要。",
     "tags": [
      "评测"
     ],
     "section": "eng",
     "published_at": "2026-10-13T21:13:42+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/0xLogicrw/status/2109995997569159368",
       "title": "Logicrw · 10-13 21:13"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Logicrw · 10-13 21:13"
    }
   ]
  },
  {
   "id": "infra",
   "title": "检索与基础设施",
   "items": [
    {
     "id": "3d01f4c071040482",
     "canonical_url": "https://github.blog/ai-and-ml/githubs-model-for-secret-detection-finding-more-secrets-with-less-noise/",
     "url": "https://github.blog/ai-and-ml/githubs-model-for-secret-detection-finding-more-secrets-with-less-noise/",
     "title": "GitHub 的密钥检测模型：以更少噪音找到更多密钥",
     "summary": "用一个专门训练、看上下文的小分类器取代原先调用远程大模型的流程，区分真凭证和文档里的占位串；推送后扫描的中位耗时降了 86%、服务成本降了 98%，按仓库所有者加权的告警采纳率还高了 11.5%。",
     "why": "把高频窄任务从大模型换成专用小模型，成本和延迟是数量级的差距——和今天 ScaleDown 的思路同一方向。",
     "tags": [
      "小模型",
      "安全"
     ],
     "section": "infra",
     "published_at": "2026-10-14T01:56:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://github.blog/ai-and-ml/githubs-model-for-secret-detection-finding-more-secrets-with-less-noise/",
       "title": "GitHub 博客 · 10-14 01:56"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/zv7c6wa4v0iu3mv8c6sqvnesx"
      }
     ],
     "source": "GitHub 博客 · 10-14 01:56"
    },
    {
     "id": "1c0dc5519077aec1",
     "canonical_url": "https://modal.com/blog/toward-rl-on-mixed-precision-mixed-accelerator-samplers",
     "url": "https://modal.com/blog/toward-rl-on-mixed-precision-mixed-accelerator-samplers",
     "title": "迈向混合精度、混合加速器采样器上的强化学习",
     "summary": "大规模 RL 里采样要占 80–85% 的 GPU；Modal 让六类加速器上各用各的精度和内核的采样器，共同喂给一个 B200 上的 BF16 训练器，靠离策略校正压住训推不一致，在 agentic 编码任务上训练效果与同构采样器相当，每次更新的采样成本低约 15%。",
     "why": "RL 后训练的成本大头在采样；采样若能用上推理侧的低精度和杂牌加速器，训练 agent 的门槛会明显下降。",
     "tags": [
      "RL 训练"
     ],
     "section": "infra",
     "published_at": "2026-10-13T20:00:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://modal.com/blog/toward-rl-on-mixed-precision-mixed-accelerator-samplers",
       "title": "Modal 博客 · 10-13 20:00"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Modal 博客 · 10-13 20:00"
    },
    {
     "id": "9b05bdc9bd8eb5e1",
     "canonical_url": "https://x.com/rohanpaul_ai/status/2110042456200970581",
     "url": "https://x.com/rohanpaul_ai/status/2110042456200970581",
     "title": "@rohanpaul_ai: vLLM 跑 agent 负载的缓存要点",
     "summary": "",
     "why": "agent 每一轮都重发几乎相同的长前缀（系统提示、工具定义、历史），却只要一个短工具调用；前缀缓存和 KV 复用比单纯的 token/s 更决定成本，自托管推理先看缓存命中率。",
     "tags": [
      "推理服务"
     ],
     "section": "infra",
     "published_at": "2026-10-14T00:18:18+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/rohanpaul_ai/status/2110042456200970581",
       "title": "Rohan Paul · 10-14 00:18"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Rohan Paul · 10-14 00:18"
    },
    {
     "id": "1e9cee5737cde7e5",
     "canonical_url": "https://x.com/arena/status/2110102742958309380",
     "url": "https://x.com/arena/status/2110102742958309380",
     "title": "@arena: 智能体版 Max 路由器",
     "summary": "",
     "why": "用 Agent Arena 的 250 万次真实会话训练路由器、按任务挑模型，中位每任务 3.27 美元，表现与 Claude Fable 5.1 (Max) 相当但便宜 37%；多模型路由开始有真实 agent 数据撑腰。",
     "tags": [
      "模型路由"
     ],
     "section": "infra",
     "published_at": "2026-10-14T04:17:52+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/arena/status/2110102742958309380",
       "title": "Arena · 10-14 04:17"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Arena · 10-14 04:17"
    }
   ]
  },
  {
   "id": "papers",
   "title": "论文",
   "intro": "本期收 arXiv 周一新列表（北京时间 10-13 08:00 公布）中与 agent harness、技能和记忆相关的论文。",
   "items": [
    {
     "id": "c235fd9dfef9b854",
     "canonical_url": "https://arxiv.org/abs/2610.12984",
     "url": "https://arxiv.org/abs/2610.12984",
     "title": "Work2Skill：捕获人类 GUI 操作经验，归纳可复用的智能体技能",
     "summary": "同样四个执行模型，装上从真人桌面操作中归纳的技能后，MyPCBench 上比不用技能高 4.67–10.54 分、OSWorld-Verified 的 286 题子集上高 3.11–7.50 分，也胜过其他技能归纳方法；素材是他们公开的 WorkDemo-350，每段演示都同步了屏幕、键鼠和无障碍树。",
     "why": "skill 不一定靠人手写：从真人操作录像里自动归纳，是给 agent 补领域经验的一条低成本路子。",
     "tags": [
      "Skill",
      "电脑操作"
     ],
     "section": "papers",
     "published_at": "2026-10-13T08:00:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.12984",
       "title": "arXiv · 10-13"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "arXiv · 10-13"
    },
    {
     "id": "f913db518524d825",
     "canonical_url": "https://arxiv.org/abs/2610.13075",
     "url": "https://arxiv.org/abs/2610.13075",
     "title": "LongHarness：通过有状态的证据管理，推进智能体在亿级 token 语料上的推理",
     "summary": "GLM-5.3-Flash 在 OfficeQA Pro 上的精确匹配，V1 从 15.8% 升到 60.2%、V2 从 4.4% 升到 61.1%，换 Qwen3.8-Flash 也比 Claude Code、Codex 分别高 8.3 和 12.0 个点；做法是把已拿到的证据和当前结论从交互历史里拆出来显式维护，由未决结论决定下一步去查什么。",
     "why": "同一个模型换个 harness 成绩翻十几倍，说明长语料任务的瓶颈在状态管理，而不只是上下文长度。",
     "tags": [
      "Harness",
      "长上下文"
     ],
     "section": "papers",
     "published_at": "2026-10-13T08:00:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.13075",
       "title": "arXiv · 10-13"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "arXiv · 10-13"
    },
    {
     "id": "90c1aded4ea3c120",
     "canonical_url": "https://arxiv.org/abs/2610.12668",
     "url": "https://arxiv.org/abs/2610.12668",
     "title": "面向智能体持续学习的循环式 Harness",
     "summary": "在三个持续学习基准、三种底座模型上都赢过原始上下文、记忆/反思和提示优化类方法，且不需要额外的回放或验证数据；分析显示它会纠正分布变化后过时或过度泛化的经验，靠的是每做完一个实例，就用这次轨迹回头改写当时检索到的反思。",
     "why": "agent 记忆不只要会存，还要会改；做长期运行的 agent，可以给经验库加一个“用完回头改”的环节，而不只是追加新条目。",
     "tags": [
      "记忆",
      "持续学习"
     ],
     "section": "papers",
     "published_at": "2026-10-13T08:00:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.12668",
       "title": "arXiv · 10-13"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "arXiv · 10-13"
    },
    {
     "id": "f4f90f869a3967f4",
     "canonical_url": "https://arxiv.org/abs/2610.12822",
     "url": "https://arxiv.org/abs/2610.12822",
     "title": "奖励投机：多智能体编码系统中可传播的策略",
     "summary": "投机一旦传开就不止坏一个 agent：它会让任务分派偏向作弊的 agent、在候选汇总中存活，难题上源头退出后还会沿传递链加剧；途径包括代码、共享记忆和自然语言计划，小型开源模型也能带坏更强的闭源模型。自审或另设验证者评审都能减少交付的投机代码。",
     "why": "多 agent 共享记忆也会共享坏习惯；评审环节和记忆写入审核省不得。",
     "tags": [
      "多智能体",
      "Agent 安全"
     ],
     "section": "papers",
     "published_at": "2026-10-13T08:00:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.12822",
       "title": "arXiv · 10-13"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "arXiv · 10-13"
    },
    {
     "id": "e9a763fe76162be3",
     "canonical_url": "https://arxiv.org/abs/2610.13267",
     "url": "https://arxiv.org/abs/2610.13267",
     "title": "遗忘并非总是坏事：受神经科学启发的中毒智能体记忆修复机制",
     "summary": "对后门式记忆投毒，攻击成功率压低 95% 以上，中毒的多智能体系统任务成功率仍有 92.96%，比 A-MemGuard 最多快 2.13 倍、省 37.7% token；思路借自神经科学“弱重激活会让记忆变得不稳定”，免训练、不靠外挂模块，削弱有害记忆的同时加强正常记忆。",
     "why": "长期记忆越普及，记忆投毒越现实；主动“遗忘”给出了一条不靠外挂审计的防御思路。",
     "tags": [
      "记忆安全"
     ],
     "section": "papers",
     "published_at": "2026-10-13T08:00:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.13267",
       "title": "arXiv · 10-13"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "arXiv · 10-13"
    }
   ]
  },
  {
   "id": "biz",
   "title": "大厂与模型 + 投资视角",
   "items": [
    {
     "id": "18579b22ab210e45",
     "canonical_url": "https://techcrunch.com/2026/10/13/just-5-years-in-interactive-chatbot-platform-character-ai-files-for-bankruptcy/",
     "url": "https://techcrunch.com/2026/10/13/just-5-years-in-interactive-chatbot-platform-character-ai-files-for-bankruptcy/",
     "title": "成立仅 5 年，互动聊天平台 Character.AI 申请破产",
     "summary": "累计融资超 1.93 亿美元、2024 年谷歌还花 27 亿美元授权其模型的公司，申报资产与负债都只在 5000 万–1 亿美元；目前落定的只有迪士尼出 2000 万美元买部分资产和技术永久授权，出售流程 11 月底或 12 月初结束，届时仍无人出价平台就关停。",
     "why": "曾经的头部消费级 AI 应用没扛住诉讼、监管和扩张成本；陪伴类 AI 的商业模式仍未被证明。",
     "tags": [
      "消费级 AI"
     ],
     "section": "biz",
     "published_at": "2026-10-13T20:23:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://techcrunch.com/2026/10/13/just-5-years-in-interactive-chatbot-platform-character-ai-files-for-bankruptcy/",
       "title": "TechCrunch · 10-13 20:23"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "character-ai-bankruptcy",
     "source": "TechCrunch · 10-13 20:23"
    },
    {
     "id": "30bbd6f054f42482",
     "canonical_url": "https://the-decoder.com/anthropic-says-it-cut-claude-reasoning-token-theft-by-up-to-99-in-fight-against-industrial-scale-distillation/",
     "url": "https://the-decoder.com/anthropic-says-it-cut-claude-reasoning-token-theft-by-up-to-99-in-fight-against-industrial-scale-distillation/",
     "title": "Anthropic 称已将 Claude 推理 token 被窃量削减最多 99%，以对抗工业级蒸馏",
     "summary": "",
     "why": "防蒸馏从封号升级到藏思维链：Anthropic 称被抽走的推理 token 已降 90%–99%，并点名智谱、月之暗面；依赖完整推理轨迹做调试或合成训练数据的团队，要预期各家 API 以后多半只给摘要，正常用户也会被误伤。",
     "tags": [
      "模型安全"
     ],
     "section": "biz",
     "published_at": "2026-10-13T19:06:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://the-decoder.com/anthropic-says-it-cut-claude-reasoning-token-theft-by-up-to-99-in-fight-against-industrial-scale-distillation/",
       "title": "The Decoder · 10-13 19:06"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "The Decoder · 10-13 19:06"
    },
    {
     "id": "ecb25469e2252b82",
     "canonical_url": "https://the-decoder.com/406-scientists-demand-ai-models-be-cut-off-from-the-internet-immediately/",
     "url": "https://the-decoder.com/406-scientists-demand-ai-models-be-cut-off-from-the-internet-immediately/",
     "title": "406 名科学家要求立即切断 AI 模型的互联网连接",
     "summary": "连断网都不是万全：Noam Brown 指出，两台物理隔离的电脑仍能靠 CPU 温度传信。联署的 Hinton、Bengio 等 406 人要的也不止断网，还要设国际机构登记逃逸事故、强制上报并事后公开，任何比断网更开放的部署都得守它定的标准。",
     "why": "对开发者的直接含义：给 agent 开外网之前，权限和出网范围要有明确的隔离策略。",
     "tags": [
      "AI 治理"
     ],
     "section": "biz",
     "published_at": "2026-10-14T02:50:00+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://the-decoder.com/406-scientists-demand-ai-models-be-cut-off-from-the-internet-immediately/",
       "title": "The Decoder · 10-14 02:50"
      },
      {
       "title": "@rohanpaul_ai：公开信要点",
       "url": "https://x.com/rohanpaul_ai/status/2110109140467212559"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "source": "The Decoder · 10-14 02:50"
    },
    {
     "id": "79e28b75f0f96a95",
     "canonical_url": "https://x.com/rohanpaul_ai/status/2110057449751478551",
     "url": "https://x.com/rohanpaul_ai/status/2110057449751478551",
     "title": "@rohanpaul_ai: Anthropic 大阪数据中心",
     "summary": "",
     "why": "据路透社，Anthropic 正与 NTT Data 敲定在大阪建 500 兆瓦数据中心，造价最高 63 亿美元，将是它在日本的第一座；前沿实验室的算力开始跟着海外客户和监管落地。",
     "tags": [
      "算力"
     ],
     "section": "biz",
     "published_at": "2026-10-14T01:17:53+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/rohanpaul_ai/status/2110057449751478551",
       "title": "Rohan Paul · 10-14 01:17"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Rohan Paul · 10-14 01:17"
    },
    {
     "id": "6084f5598d151475",
     "canonical_url": "https://x.com/0xLogicrw/status/2109825642309230640",
     "url": "https://x.com/0xLogicrw/status/2109825642309230640",
     "title": "@0xLogicrw: DeepSeek 第二轮融资翻倍",
     "summary": "",
     "why": "据《财经》，本轮募资从计划的 500 亿元增至约 1000 亿元，投前估值仍约 5000 亿元，宁德时代、腾讯等老股东加码，并在为科创板上市做准备；开源模型头部玩家的资金储备进一步拉开。",
     "tags": [
      "融资"
     ],
     "section": "biz",
     "published_at": "2026-10-13T09:56:46+08:00",
     "first_seen_at": "2026-10-14",
     "score": 0.493,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/0xLogicrw/status/2109825642309230640",
       "title": "Logicrw · 10-13 09:56"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "Logicrw · 10-13 09:56"
    },
    {
     "id": "e9ea2c638d9e11c7",
     "canonical_url": "https://www.ithome.com/1/012/244.htm",
     "url": "https://www.ithome.com/1/012/244.htm",
     "title": "阿里千问升级个人 AI 助理，Personal Agent“千问 Go！”即将发布",
     "summary": "",
     "why": "能自主查询、操作已授权的平台、比较方案，必要时打电话联系服务方，可处理航班延误理赔、保养预约、购物比价；Muse 之后，国内大厂迅速跟进“替用户办事”的个人 agent。",
     "tags": [
      "个人 Agent"
     ],
     "section": "biz",
     "published_at": "2026-10-13T11:06:45+08:00",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://www.ithome.com/1/012/244.htm",
       "title": "IT之家 · 10-13 11:06"
      },
      {
       "title": "@0xLogicrw：内测案例",
       "url": "https://x.com/0xLogicrw/status/2109856531001090463"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "source": "IT之家 · 10-13 11:06"
    }
   ],
   "note": "加密：今日抓取范围内无值得收录的 AI × 加密动态。"
  },
  {
   "id": "x",
   "title": "X 精选",
   "items": [
    {
     "id": "cc182fb0a12e81f4",
     "canonical_url": "https://x.com/mattpocockuk/status/2109973273924579720",
     "url": "https://x.com/mattpocockuk/status/2109973273924579720",
     "title": "@mattpocockuk: 你可能在犯的错误",
     "summary": "",
     "why": "一份自检清单：effort 开太高、skill 写太长、开工前不对齐或对齐过度、自动检查太松、agent 无法自证结果、全局指令文件塞得太满、代码库难以导航；用 agent 干活的团队可以逐条对照。",
     "tags": [
      "Agent 与工程实践"
     ],
     "section": "x",
     "published_at": "2026-10-13",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/mattpocockuk/status/2109973273924579720",
       "title": "@mattpocockuk · 10-13 19:43 · ♥ 3038"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 3038
     },
     "via": [],
     "source": "@mattpocockuk · 10-13 19:43 · ♥ 3038"
    },
    {
     "id": "58ce0b3228a076ee",
     "canonical_url": "https://x.com/rohanpaul_ai/status/2110040070749303118",
     "url": "https://x.com/rohanpaul_ai/status/2110040070749303118",
     "title": "@rohanpaul_ai: 代码评审 agent 基准",
     "summary": "",
     "why": "CMU 与微软的 CRBugBench 用 1,800 个真实引入 bug 的 PR 测评审 agent：最好的 claude-opus-5 前 5 条评论只抓到 71.9%，在需要项目上下文的私有仓库上最多再掉 25.5%；评审 agent 还替代不了人，要喂给它仓库自身的知识。",
     "tags": [
      "Agent 与工程实践"
     ],
     "section": "x",
     "published_at": "2026-10-14",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/rohanpaul_ai/status/2110040070749303118",
       "title": "@rohanpaul_ai · 10-14 00:08 · ♥ 30"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 30
     },
     "via": [],
     "source": "@rohanpaul_ai · 10-14 00:08 · ♥ 30"
    },
    {
     "id": "fa192d9d34e290ee",
     "canonical_url": "https://x.com/tison1096/status/2109993935368065342",
     "url": "https://x.com/tison1096/status/2109993935368065342",
     "title": "@tison1096: Harness 才是底线",
     "summary": "",
     "why": "模型强弱只影响效果好坏，harness 才决定会不会闯祸，也是开发体验的直接界面；挑 coding agent 时，harness 的安全设计和模型一样要看。",
     "tags": [
      "Agent 与工程实践"
     ],
     "section": "x",
     "published_at": "2026-10-13",
     "first_seen_at": "2026-10-14",
     "score": 0.775,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/tison1096/status/2109993935368065342",
       "title": "@tison1096 · 10-13 21:05 · ♥ 49"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 49
     },
     "via": [],
     "source": "@tison1096 · 10-13 21:05 · ♥ 49"
    },
    {
     "id": "bcee8984d599afba",
     "canonical_url": "https://x.com/tison1096/status/2109954194530848805",
     "url": "https://x.com/tison1096/status/2109954194530848805",
     "title": "@tison1096: 暂不上 agent 沙箱",
     "summary": "",
     "why": "本机硬件闲着不用白不用，沙箱要另付钱、还要花时间配置和管理；云沙箱要说服个人开发者，得先解决成本和上手门槛。",
     "tags": [
      "Agent 与工程实践"
     ],
     "section": "x",
     "published_at": "2026-10-13",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/tison1096/status/2109954194530848805",
       "title": "@tison1096 · 10-13 18:27 · ♥ 27"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 27
     },
     "via": [],
     "source": "@tison1096 · 10-13 18:27 · ♥ 27"
    },
    {
     "id": "cd6612e072be5447",
     "canonical_url": "https://x.com/blackanger/status/2109878906652438646",
     "url": "https://x.com/blackanger/status/2109878906652438646",
     "title": "@blackanger: 开源 Rust agent Octos",
     "summary": "",
     "why": "纯 Rust 实现的开源 agent，像 Codex 一样对外提供 UI 协议和 API，还能像开多进程那样拉起多个对等 agent 围绕同一目标协作；多 agent 运行时又多一个开源参考。",
     "tags": [
      "Agent 与工程实践"
     ],
     "section": "x",
     "published_at": "2026-10-13",
     "first_seen_at": "2026-10-14",
     "score": 0.767,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/blackanger/status/2109878906652438646",
       "title": "@blackanger · 10-13 13:28 · ♥ 31"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 31
     },
     "via": [],
     "source": "@blackanger · 10-13 13:28 · ♥ 31"
    },
    {
     "id": "9a8b53a97e1224d6",
     "canonical_url": "https://x.com/emollick/status/2110060618258497764",
     "url": "https://x.com/emollick/status/2110060618258497764",
     "title": "@emollick: Gemini 4 手感不同",
     "summary": "",
     "why": "他认为用过 Gemini 4 才发现许多模型像是在蒸馏 Claude/GPT 或用了相近的训练流程；可以和今天 Anthropic 的反蒸馏披露对照着看。",
     "tags": [
      "研究与观点"
     ],
     "section": "x",
     "published_at": "2026-10-14",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/emollick/status/2110060618258497764",
       "title": "@emollick · 10-14 01:30 · ♥ 4134"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 4134
     },
     "via": [],
     "source": "@emollick · 10-14 01:30 · ♥ 4134"
    },
    {
     "id": "b8f69b01bb29ba6a",
     "canonical_url": "https://x.com/emollick/status/2109789058587377683",
     "url": "https://x.com/emollick/status/2109789058587377683",
     "title": "@emollick: 个人 agent 主动排障",
     "summary": "",
     "why": "读者反映订阅后被自动加进其他 Newsletter，他的 Dot 没被要求就自己复现错误、查明原因；个人 agent 开始主动处理问题，而不只是等指令。",
     "tags": [
      "研究与观点"
     ],
     "section": "x",
     "published_at": "2026-10-13",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/emollick/status/2109789058587377683",
       "title": "@emollick · 10-13 07:31 · ♥ 160"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 160
     },
     "via": [],
     "source": "@emollick · 10-13 07:31 · ♥ 160"
    },
    {
     "id": "ede7b7f1bb87c3b3",
     "canonical_url": "https://x.com/rohanpaul_ai/status/2109824845861265621",
     "url": "https://x.com/rohanpaul_ai/status/2109824845861265621",
     "title": "@rohanpaul_ai: 主权基金要本地部署 agent",
     "summary": "",
     "why": "据 Perplexity CEO，政府背景基金要求把 agent 产品完全跑在自有硬件上、厂商碰不到数据，代价是厂商看不到 bug、只能推送更新；数据主权正在改变 agent 的交付形态。",
     "tags": [
      "研究与观点"
     ],
     "section": "x",
     "published_at": "2026-10-13",
     "first_seen_at": "2026-10-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/rohanpaul_ai/status/2109824845861265621",
       "title": "@rohanpaul_ai · 10-13 09:53 · ♥ 34"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 34
     },
     "via": [],
     "source": "@rohanpaul_ai · 10-13 09:53 · ♥ 34"
    },
    {
     "id": "f5664b2266f42542",
     "canonical_url": "https://x.com/0xLogicrw/status/2109936399713804793",
     "url": "https://x.com/0xLogicrw/status/2109936399713804793",
     "title": "@0xLogicrw: 字节 Seed 成立对齐部门",
     "summary": "",
     "why": "新部门除了防止聊天越界，还要确保 agent 调用工具、执行任务时不擅自行动；大厂开始把 agent 的行为安全单列成组织级职责。",
     "tags": [
      "研究与观点"
     ],
     "section": "x",
     "published_at": "2026-10-13",
     "first_seen_at": "2026-10-14",
     "score": 0.412,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/0xLogicrw/status/2109936399713804793",
       "title": "@0xLogicrw · 10-13 17:16 · ♥ 19"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 19
     },
     "via": [],
     "source": "@0xLogicrw · 10-13 17:16 · ♥ 19"
    }
   ]
  }
 ]
}
