{
 "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": "2026-11-07T07:40:00+08:00",
 "published_at": "2026-11-07T07:40:07+08:00",
 "fetch_window": {
  "start": "2026-11-06T07:40:00+08:00",
  "end": "2026-11-07T07:30:00+08:00"
 },
 "date": "2026-11-07",
 "title": "AI 要闻 · 2026-11-07",
 "url": "https://www.wangwu.ai/daily/2026-11-07/",
 "fetched": "2026-11-07 07:24–07:40（北京时间）",
 "sections": [
  {
   "id": "top3",
   "title": "今日最该知道的 3 条",
   "items": [
    {
     "id": "4cbb759bf2c3b9fe",
     "canonical_url": "https://openai.com/index/measuring-how-ai-agents-handle-wet-lab-science",
     "url": "https://openai.com/index/measuring-how-ai-agents-handle-wet-lab-science",
     "title": "衡量 AI 智能体如何处理日常湿实验室科学",
     "summary": "与 Ligo 合做的 BioBasicsBench 共 100 题，覆盖蛋白/成像流式/定量/核酸/克隆，用原始凝胶与显微读数按量规给分，并测「证据不足则不下结论」。七模型 37 档共 1.85 万次：xhigh 下 Astra 66.7%、GPT-5.6 Sol 43.8%；对照误判与菌落计数仍易翻车。",
     "why": "把 agent 评测从论文精选图拉回日常湿实验的失败与对照陷阱。做科研 agent / lab copilot 时，用这 100 题量规对读图、查对照、拒下结论三项，别只盯 Astra 相对 Sol 的分差。",
     "tags": [
      "BioBasicsBench",
      "湿实验室",
      "GPT-6 Astra",
      "Ligo"
     ],
     "section": "top3",
     "published_at": "2026-11-06T08:00:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": 0.645,
     "importance": 0.8,
     "sources": [
      {
       "url": "https://openai.com/index/measuring-how-ai-agents-handle-wet-lab-science",
       "title": "OpenAI · 11-06；与 Ligo 合做"
      },
      {
       "title": "OpenAI 原帖",
       "url": "https://x.com/OpenAI/status/2118782708767965238"
      },
      {
       "title": "Rohan Paul 转述",
       "url": "https://x.com/rohanpaul_ai/status/2118794080885051683"
      }
     ],
     "signals": {
      "source_count": 3
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "openai-biobasicsbench-2026-11",
     "source": "OpenAI · 11-06；与 Ligo 合做"
    },
    {
     "id": "a35091b25996e50b",
     "canonical_url": "https://x.com/rohanpaul_ai/status/2118813225336873181",
     "url": "https://x.com/rohanpaul_ai/status/2118813225336873181",
     "title": "@rohanpaul_ai: 融资首批到账",
     "summary": "",
     "why": "把「首笔到账」和「整轮上限」拆开读：帖文援引 Bloomberg，称 110 亿美元已在当日打入、对应 1.25 万亿美元估值一轮的首 tranche，其余最高至合计 200 亿美元「数周内」才到——跟踪时先核已入账金额与估值口径，别把首批当成整轮交割完毕。",
     "tags": [
      "OpenAI",
      "融资",
      "Bloomberg",
      "1.25万亿"
     ],
     "section": "top3",
     "published_at": "2026-11-06T21:10:13.000Z",
     "first_seen_at": "2026-11-07",
     "score": 0.659,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://x.com/rohanpaul_ai/status/2118813225336873181",
       "title": "X · rohanpaul_ai 援引 Bloomberg · 11-07 05:10"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "openai-funding-11b-2026-11",
     "source": "X · rohanpaul_ai 援引 Bloomberg · 11-07 05:10"
    },
    {
     "id": "ea0bb811a5a96163",
     "canonical_url": "https://techcrunch.com/2026/11/06/meta-introduces-spaces-a-new-way-to-collaborate-with-your-muse-agent/",
     "url": "https://techcrunch.com/2026/11/06/meta-introduces-spaces-a-new-way-to-collaborate-with-your-muse-agent/",
     "title": "为你定制 Muse 应用",
     "summary": "Muse 上线 spaces：用自然语言描述想法即可生成常驻自定义体验（礼物追踪、日程看板、DIY 规划等），可放侧栏/主标签、按时段显隐；接上常用应用后会在后台自更新（演出档期、价格变动等）。即日起在 Muse 可用地区全开，并探索空间分享协作。",
     "why": "把「对话式 agent」推进到可钉在导航里的个人微应用。若做助手产品，重点核 spaces 的布局调度与后台连应用刷新，而不是又一个聊天窗；与同日 Active Memory 升级是同一产品族、不同层面。",
     "tags": [
      "Meta",
      "Muse",
      "Spaces",
      "项目协作"
     ],
     "section": "top3",
     "published_at": "2026-11-07T02:44:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": 0.645,
     "importance": 0.8,
     "sources": [
      {
       "url": "https://techcrunch.com/2026/11/06/meta-introduces-spaces-a-new-way-to-collaborate-with-your-muse-agent/",
       "title": "TechCrunch · 11-07；Meta Newsroom 同步"
      },
      {
       "title": "TechCrunch",
       "url": "https://techcrunch.com/2026/11/06/meta-introduces-spaces-a-new-way-to-collaborate-with-your-muse-agent/"
      },
      {
       "title": "Active Memory 研究博文",
       "url": "https://research.meta.ai/blog/specialized-models-for-muse-and-metaai"
      },
      {
       "title": "@finkd",
       "url": "https://x.com/finkd/status/2118767522782626104"
      },
      {
       "title": "Meta Newsroom",
       "url": "https://about.fb.com/news/2026/11/introducing-spaces-customize-muse-apps-with-your-ai/"
      }
     ],
     "signals": {
      "source_count": 5
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "meta-muse-spaces-2026-11",
     "source": "TechCrunch · 11-07；Meta Newsroom 同步"
    }
   ]
  },
  {
   "id": "edge",
   "title": "专栏：端侧模型",
   "intro": "MNN-Pocket 手机原生小模型；Underdog 浏览器离线 TTS；llama.cpp 补 Kimi-K3.1-Pro。",
   "items": [
    {
     "id": "2cbc93711c824ade",
     "canonical_url": "https://www.ithome.com/1/023/212.htm",
     "url": "https://www.ithome.com/1/023/212.htm",
     "title": "淘宝 MNN 团队开源 0.3B 手机模型，苹果 iPhone 18 Pro 离线跑出 350 token/s",
     "summary": "约 3 亿参数、INT4 部署包 173MB；iPhone 18 Pro 经 MNN 离线解码约 350 tok/s。iPad Pro M5 上 Prefill 8721、Decode 466 tok/s（约为同条件 Qwen3-0.6B 的 1.63 倍），峰值 Metal 内存增量 228MB（对方 427MB/包体 340MB）；设计目标含峰值内存 <300MB、解码 >200 tok/s。",
     "why": "做端侧小模型选型时，对照目标机型核对 INT4/MNN 包体与峰值内存，以及指令遵循是否落到你的场景——别把 iPhone 18 Pro 的 350 tok/s 直接当成全机型现货吞吐。",
     "tags": [
      "MNN-Pocket",
      "端侧",
      "INT4",
      "开源"
     ],
     "section": "edge",
     "published_at": "2026-11-06T11:50:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": 0.599,
     "importance": 0.55,
     "sources": [
      {
       "url": "https://www.ithome.com/1/023/212.htm",
       "title": "IT之家 · 11-06 11:50"
      },
      {
       "title": "GitHub alibaba/MNN",
       "url": "https://github.com/alibaba/mnn"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "mnn-pocket-0-3b-2026-11",
     "source": "IT之家 · 11-06 11:50"
    },
    {
     "id": "57d8075e73bb07cb",
     "canonical_url": "https://techcrunch.com/2026/11/06/underdog-shows-off-a-32mb-speech-model-in-new-demo/",
     "url": "https://techcrunch.com/2026/11/06/underdog-shows-off-a-32mb-speech-model-in-new-demo/",
     "title": "Underdog 在新演示中展示 32MB 语音模型",
     "summary": "开源 TTS「Woof」约 31.7MB、3100 万参数，浏览器离线可跑并配复古主题打字出声演示；同厂此前有端侧 STT，另有排队中的本地助手应用叙事，强调设备端私有 AI。",
     "why": "端侧语音若要「可下载、可离线」，可把 Woof 体积与开源下载当作对照样本；落地前核浏览器实机延迟与许可证，别把演示游戏当成生产级助手。",
     "tags": [
      "Underdog",
      "Woof",
      "端侧语音",
      "TTS"
     ],
     "section": "edge",
     "published_at": "2026-11-06T05:02:00-08:00",
     "first_seen_at": "2026-11-07",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://techcrunch.com/2026/11/06/underdog-shows-off-a-32mb-speech-model-in-new-demo/",
       "title": "TechCrunch · 11-06"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "underdog-woof-32mb-2026-11",
     "source": "TechCrunch · 11-06"
    },
    {
     "id": "8c0e4276d9e76dbf",
     "canonical_url": "https://github.com/ggml-org/llama.cpp/releases/tag/b12180",
     "url": "https://github.com/ggml-org/llama.cpp/releases/tag/b12180",
     "title": "ggml-org/llama.cpp b12180",
     "summary": "Kimi-K3.1-Pro 文本权重可进本机构建：93 层里 24 层带 lightning indexer（DSA），其余 KDA/残差/MoE 仍走 K3 图与转换器；indexer 用整段 cache 上的加性掩码复用已有稀疏 flash，而不是另抽 top-k 行。",
     "why": "本地若跟 Kimi K3.1 系权重，可先核 llama.cpp 此 b 号的 indexer/掩码路径是否覆盖你的量化与后端，再谈吞吐。",
     "tags": [
      "llama.cpp",
      "Kimi-K3.1-Pro",
      "DSA"
     ],
     "section": "edge",
     "published_at": "2026-11-06T23:18:52Z",
     "first_seen_at": "2026-11-07",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://github.com/ggml-org/llama.cpp/releases/tag/b12180",
       "title": "GitHub · 11-07 07:18"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "llama-cpp-b12180",
     "source": "GitHub · 11-07 07:18"
    }
   ],
   "note": "日常 llama.cpp b 号仅保留含新模型支持的一条。"
  },
  {
   "id": "col",
   "title": "专栏：Agent 记忆与协作",
   "intro": "Claude dreaming 整理记忆；Amp / Muse 两侧 spaces；Inkling 2 Small 长程 agent。",
   "items": [
    {
     "id": "6ad54605491f6c6e",
     "canonical_url": "https://www.anthropic.com/research/components-of-agent-memory",
     "url": "https://www.anthropic.com/research/components-of-agent-memory",
     "title": "智能体记忆的组成",
     "summary": "记忆拆成会话内显著性笔记与离线 dreaming：对 transcript 纠错、建 wiki、提升预载层。Pokémon 例中纠错方位后才过 Mt. Moon；Stainless 语料 dreaming 2→128 轮，25 问约 13%→48%。Frontier 模型更宜自由判断显著性，而非「只记失败」。",
     "why": "自建 agent 记忆栈时可对标三件事：谁在会话里写笔记、谁在空闲时对 transcript 做 dreaming、预载层如何保持短小——别把「多存文件」当成记忆系统。",
     "tags": [
      "Anthropic",
      "dreaming",
      "记忆",
      "CL-Bench"
     ],
     "section": "col",
     "published_at": "2026-11-07T05:16:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://www.anthropic.com/research/components-of-agent-memory",
       "title": "Anthropic Engineering · 11-06"
      },
      {
       "title": "Anthropic 原帖",
       "url": "https://x.com/AnthropicAI/status/2118814753154064501"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "anthropic-agent-memory-dreaming-2026-11",
     "source": "Anthropic Engineering · 11-06"
    },
    {
     "id": "0f0928945a08b039",
     "canonical_url": "https://ampcode.com/news/spaces",
     "url": "https://ampcode.com/news/spaces",
     "title": "Amp 推出 spaces：用单个智能体协调多个线程的全新协作方式",
     "summary": "space 是与单个 agent 的会话，由它生成并协调其他线程，用户不必自己开帖组织。内部用同一 space 做功能共建、上线值守与安全审查（曾拆成 16 个区域并行审），团队讨论也常搬进 space 而非另开 Slack。",
     "why": "多 agent 产品若还停留在「人手工开线程」，可对照 Amp 的「船长椅」：一个协调 agent + 可下钻线程。落地前核权限边界与线程产物如何回写主会话。",
     "tags": [
      "Amp",
      "spaces",
      "多 Agent",
      "协作"
     ],
     "section": "col",
     "published_at": "2026-11-06T08:00:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": 0.54,
     "importance": 0.5,
     "sources": [
      {
       "url": "https://ampcode.com/news/spaces",
       "title": "Amp · 11-06"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "amp-spaces-2026-11",
     "source": "Amp · 11-06"
    },
    {
     "id": "f0bfa79925521f6a",
     "canonical_url": "https://thinkingmachines.ai/news/inkling-2-small/",
     "url": "https://thinkingmachines.ai/news/inkling-2-small/",
     "title": "Introducing Inkling 2 Small",
     "summary": "594B/激活 17B 的 MoE，支持图像与 variable effort；AA Intelligence Index 自报约 31。先经 Tinker 专有提供，安全外测后拟 Apache 2.0 放权重。强调长程 agentic coding / autoresearch；Code Arena WebDev 等榜单在窗内继续被引用。",
     "why": "若盯「西方开放权重」选型：先分清现货是 Tinker API、权重仍待安全测试；并把自报指数与条件榜位（权重放出后）放一起看，别把预告第一当成已可下载的现货位次。",
     "tags": [
      "Inkling 2 Small",
      "MoE",
      "开放权重",
      "Tinker"
     ],
     "section": "col",
     "published_at": "2026-11-06T12:25:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://thinkingmachines.ai/news/inkling-2-small/",
       "title": "MarkTechPost · 11-06"
      },
      {
       "title": "the-decoder",
       "url": "https://the-decoder.com/thinking-machines-inkling-2-small-claims-top-spot-among-western-open-weight-models/"
      },
      {
       "title": "MarkTechPost",
       "url": "https://www.marktechpost.com/2026/11/05/thinking-machines-inkling-2-small-release/"
      },
      {
       "title": "Code Arena WebDev",
       "url": "https://x.com/arena/status/2118492327887155680"
      }
     ],
     "signals": {
      "source_count": 4
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "inkling-2-small-2026-11",
     "source": "MarkTechPost · 11-06"
    }
   ],
   "note": "google/skills 昨日报过未重复；Muse Active Memory 并入 Spaces 相关链接。"
  },
  {
   "id": "cl",
   "title": "主流产品 Changelog 精读",
   "intro": "Claude Code 网关/auto mode 与 MCP task；Qwen Code Skills 安装激活。",
   "items": [
    {
     "id": "9ebed600c4957576",
     "canonical_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.319",
     "url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.319",
     "title": "v2.1.319",
     "summary": "MCP 工具一旦被服务端标成 task，终端会立刻丢到后台并在结束时回灌结果；同时在无遥测/LLM 网关场景，auto mode 对「服务端核不过」的调用只问或拒，不再掉回内置分类器（可用 `CLAUDE_CODE_AUTO_MODE_SERVER=0` 保旧路径）。",
     "why": "自建/代理 Claude Code 时，先核网关是否关掉了服务端 auto 校验，以及 MCP 长任务是后台还是阻塞——这两项会直接改变权限体验与超时行为。",
     "tags": [
      "Claude Code",
      "auto mode",
      "MCP",
      "gateway"
     ],
     "section": "cl",
     "published_at": "2026-11-06T17:04:01Z",
     "first_seen_at": "2026-11-07",
     "score": 0.455,
     "importance": 0.3,
     "sources": [
      {
       "url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.319",
       "title": "GitHub · anthropics/claude-code · 11-07 01:04"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "claude-code-2-1-319",
     "source": "GitHub · anthropics/claude-code · 11-07 01:04"
    },
    {
     "id": "3c76384994249a01",
     "canonical_url": "https://github.com/QwenLM/qwen-code/releases/tag/v0.26.9",
     "url": "https://github.com/QwenLM/qwen-code/releases/tag/v0.26.9",
     "title": "v0.26.9",
     "summary": "一次请求可启停 1–100 个用户 Skill（带偏好落盘与传播状态），大 ZIP 改走磁盘安装并带失败回执与备份保护；TypeScript SDK / WebShell 补上传入口。Web Shell 另修标题重建丢字、草稿被上下键清掉、批量删会话分页等，无宣称破坏性变更。",
     "why": "若在 Qwen Code 上推团队 Skill 包，优先对一下 ZIP 安装回执与批量 activation API，是否覆盖你的托管 daemon 限额。",
     "tags": [
      "Qwen Code",
      "Skills",
      "WebShell"
     ],
     "section": "cl",
     "published_at": "2026-11-06T16:07:57Z",
     "first_seen_at": "2026-11-07",
     "score": 0.405,
     "importance": 0.3,
     "sources": [
      {
       "url": "https://github.com/QwenLM/qwen-code/releases/tag/v0.26.9",
       "title": "GitHub · QwenLM/qwen-code · 11-07 00:07"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "qwen-code-0-26-9",
     "source": "GitHub · QwenLM/qwen-code · 11-07 00:07"
    }
   ],
   "note": "Codex rust-v0.169.0-alpha.2 发布说明几乎为空，未展开。"
  },
  {
   "id": "gh",
   "title": "GitHub 热门",
   "items": [],
   "note": "今日无重要更新（日榜非 AI 主线未收；security-audit-skill / hallmark / openbot 此前已报，无增量后续）。"
  },
  {
   "id": "eng",
   "title": "Agent 工程",
   "intro": "vLLM 对齐内核后的异步 REINFORCE Zero；Hy4 扩容与 APEX 后续。",
   "items": [
    {
     "id": "bc86e45131f3ef3b",
     "canonical_url": "https://vllm.ai/blog/2026-11-06-zero-mismatch-async-rl",
     "url": "https://vllm.ai/blog/2026-11-06-zero-mismatch-async-rl",
     "title": "零失配、简单异步 RL：REINFORCE Zero、DeepSeek-V4.1 与 Kimi K3.1",
     "summary": "对齐内核后的异步 REINFORCE Zero 在 16 卡上相对同步拿到 2.40× 训练步吞吐，并可在滞后 64、约 2 万次更新下保持稳定；同设定标准内核 IS 的 KL 中位涨约 83×、AIME24 从 34% 掉到 14%。DeepSeek-V4.1-Flash / Kimi K3.1（含 MI355X）已做到逐 token 零失配。",
     "why": "做 agent/tool RL 时先核 rollout-trainer 是否 bit 级 logprob 一致；文中把失配拆成可平滑校正的权重滞后与会炸方差的内核差，对齐后再敞开 IS，比先堆截断/拒样门控更对症。",
     "tags": [
      "vLLM",
      "REINFORCE Zero",
      "异步 RL",
      "零失配"
     ],
     "section": "eng",
     "published_at": "2026-11-06T00:00:00+00:00",
     "first_seen_at": "2026-11-07",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://vllm.ai/blog/2026-11-06-zero-mismatch-async-rl",
       "title": "vLLM Blog · 11-06"
      },
      {
       "title": "REINFORCE Zero 代码",
       "url": "https://github.com/hijkzzz/REINFORCE-Zero"
      },
      {
       "title": "零失配配方",
       "url": "https://github.com/Inferact/zero-mismatch-rl"
      }
     ],
     "signals": {
      "source_count": 3
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "vllm-reinforce-zero-2026-11",
     "source": "vLLM Blog · 11-06"
    },
    {
     "id": "e47a8ffbae8d6897",
     "canonical_url": "https://www.qbitai.com/2026/11/507967.html",
     "url": "https://www.qbitai.com/2026/11/507967.html",
     "title": "后续｜Hy4 十倍扩容并登顶 APEX-Agents 开源权重榜",
     "summary": "WorkBuddy 接入后首日大量触顶，腾讯称已对 Hy4 推理集群十倍扩容并提高免费额度。Agent Arena 总榜约第 13、开源模型第一；AA 智能指数称中国模型第二（次于 Kimi K3.1 Pro）、单任务约 0.78 美元。混元官号另称 Mercor APEX-Agents 开源权重榜 Pass@1 77.6%。",
     "why": "相对昨日额度故事，增量是「十倍扩容 + APEX 开源权重榜」。落地前分开核：Agent Arena / AA / APEX 不是同一榜，通稿由腾讯提供——对照正式榜与 API/产品面额度是否同步放开。",
     "tags": [
      "Hy4",
      "WorkBuddy",
      "APEX-Agents",
      "后续"
     ],
     "section": "eng",
     "published_at": "2026-11-06T10:00:11+08:00",
     "first_seen_at": "2026-11-07",
     "score": 0.537,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://www.qbitai.com/2026/11/507967.html",
       "title": "量子位（腾讯供稿）· 11-06；混元官号"
      },
      {
       "title": "混元官号 APEX",
       "url": "https://x.com/TencentHunyuan/status/2118634166107754690"
      },
      {
       "title": "Mercor APEX",
       "url": "https://www.mercor.com/apex/"
      }
     ],
     "signals": {
      "source_count": 3
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "hunyuan-hy4-capacity-apex-2026-11",
     "source": "量子位（腾讯供稿）· 11-06；混元官号"
    }
   ],
   "note": "Cursor Live Apps 仅见聚合转述、无官网原文，未展开。"
  },
  {
   "id": "infra",
   "title": "检索与基础设施",
   "intro": "OpenRouter 新一代 Auto / Auto Code；十月用量里 agent 占八成。",
   "items": [
    {
     "id": "175f7e77c927ce6f",
     "canonical_url": "https://openrouter.ai/data/october-2026",
     "url": "https://openrouter.ai/data/october-2026",
     "title": "Tokens rose 37% in October",
     "summary": "10 月跑了 790T token（环比 +37%、相对 1 月约 25×）；agentic 密钥约占全站 76% 流量。开源权重付费份额在闭源新模型当周承压后，主要靠 DeepSeek V4.1 Flash 等回升；自 5 月起 Z.ai 月度 token 约增 20×、DeepSeek 约 12×。",
     "why": "当路由/选型月报读：核对开源权重能否从闭源发布冲击里回血、以及 agent 流量是否仍抬缓存命中；别只转推文截图。",
     "tags": [
      "OpenRouter",
      "用量",
      "agent",
      "开源权重"
     ],
     "section": "infra",
     "published_at": "2026-11-07T04:28:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://openrouter.ai/data/october-2026",
       "title": "OpenRouter Data · 十月复盘"
      },
      {
       "title": "Z.ai / DeepSeek 增长帖",
       "url": "https://x.com/OpenRouter/status/2118802623059173829"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "openrouter-october-2026",
     "source": "OpenRouter Data · 十月复盘"
    }
   ],
   "note": "Nace Drex 仅见聚合短帖、无原技术报告页，未展开。"
  },
  {
   "id": "papers",
   "title": "论文",
   "intro": "检索嵌入干扰；视觉规划加速偏机器人未展开。",
   "items": [
    {
     "id": "4d59f2018b9a38ba",
     "canonical_url": "https://arxiv.org/abs/2611.07298",
     "url": "https://arxiv.org/abs/2611.07298",
     "title": "用可训练前缀 token 降低检索嵌入中的表征干扰",
     "summary": "标准检索与长上下文基准上相对微调基线有稳定提升；做法是给嵌入模型加可训练 `[DUMP]` 前缀位去吸低信息残差（含 attention sink 痕迹），架构、池化与对比损失不动，并观察到低信息 token 的激活/注意力分布随之变。",
     "why": "做 RAG/记忆检索底座时，若怀疑向量被 token 级噪声缠住，可把「可训练前缀」当轻量对照，而不是先上更复杂的解缠架构。",
     "tags": [
      "检索",
      "嵌入",
      "前缀 token"
     ],
     "section": "papers",
     "published_at": "2026-11-06T00:00:00.000Z",
     "first_seen_at": "2026-11-07",
     "score": 0.477,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2611.07298",
       "title": "arXiv · 2611.07298 · HF 日更 11-06"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "arxiv-2611-07298-dump-prefix",
     "source": "arXiv · 2611.07298 · HF 日更 11-06"
    }
   ],
   "note": "VFlash（2611.08449）偏具身视觉规划，与本栏主线弱相关，未收。"
  },
  {
   "id": "biz",
   "title": "大厂与模型",
   "intro": "METR 量化 Anthropic 研发加速；工信部智能体终端；Ling-3.2-flash；安全框架与 Remix。",
   "items": [
    {
     "id": "fbc4ee81d03f34c0",
     "canonical_url": "https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/",
     "url": "https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/",
     "title": "METR 独立评估：Anthropic 的 AI 研发加速约达 2 倍",
     "summary": "METR 在 Anthropic 做约 11 周高权限评估后估计：截至 2026-10，其对明确任务的能力改进速度约是 2024-03 至 2025-10 的 2 倍（AECI 自 2025-10 起约 1.9×，95% 区间 1.5–2.8×）；约八成现时加速归因于自用 AI，并排除训练算力增速加快，但多数关键证据被 Anthropic 遮蔽。",
     "why": "这是少见的实验室内加速量化：读时核 AECI 口径与被遮蔽证据占比，以及「近 12 个月未见明确瓶颈」是否高估——别把约 2 倍当成已验证的失控阈值穿越。",
     "tags": [
      "METR",
      "Anthropic",
      "AI R&D",
      "AECI"
     ],
     "section": "biz",
     "published_at": "2026-11-06T08:24:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/",
       "title": "METR · 博文标注 11-05；窗内讨论"
      },
      {
       "title": "Rohan Paul 转述",
       "url": "https://x.com/rohanpaul_ai/status/2118499733782384818"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "metr-anthropic-rd-2x-2026-11",
     "source": "METR · 博文标注 11-05；窗内讨论"
    },
    {
     "id": "958018c837f0086d",
     "canonical_url": "https://www.ithome.com/1/023/059.htm",
     "url": "https://www.ithome.com/1/023/059.htm",
     "title": "工信部发布通知加强智能体终端设备进网管理，提出四大工作任务",
     "summary": "四件事：上架前提交适配验证并公示机型；预置智能体的手机/平板/可穿戴纳入进网检测与唯一身份码；OTA 先报方案、结束后交报告经评估才可正式服务；分发平台公示适配列表，未签名同意不得被智能体强制调用。",
     "why": "做手机/穿戴预置 Agent 或商店分发时，把「适配列表 + 进网检测 + OTA 报备 + 调用声明」当硬门槛核对，别默认现有上架流程还能绕过。",
     "tags": [
      "工信部",
      "智能体终端",
      "进网",
      "OTA"
     ],
     "section": "biz",
     "published_at": "2026-11-06T09:13:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": 0.728,
     "importance": 0.8,
     "sources": [
      {
       "url": "https://www.ithome.com/1/023/059.htm",
       "title": "IT之家 · 11-06 09:13"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "miit-agent-terminal-2026-11",
     "source": "IT之家 · 11-06 09:13"
    },
    {
     "id": "68054d070974de17",
     "canonical_url": "https://www.ithome.com/1/023/043.htm",
     "url": "https://www.ithome.com/1/023/043.htm",
     "title": "蚂蚁百灵大模型 Ling 3.2 Flash 发布：三项基准超越 Claude Opus 5，前两周限时免费",
     "summary": "原生多模态 Ling-3.2-flash：560B 总参、约 25B 激活、1M 上下文，强化视觉与多模态长程 Agent。官方称 OmniDocBench1.5 / ZEROBench(Sub) / Vision2Web 三项超 Claude Opus 5（max）；API 输入 $0.30、输出 $0.90/百万词元，前两周限免。",
     "why": "若做多模态长程 agent 选型，先核这三项基准是否贴近你的文档/视觉任务，以及限免结束后的价与配额——别只看「超 Opus 5」标题。",
     "tags": [
      "蚂蚁百灵",
      "Ling-3.2-flash",
      "多模态",
      "MoE"
     ],
     "section": "biz",
     "published_at": "2026-11-06T08:00:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": 0.704,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://www.ithome.com/1/023/043.htm",
       "title": "IT之家 · 11-06 08:00"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "ling-3-2-flash-2026-11",
     "source": "IT之家 · 11-06 08:00"
    },
    {
     "id": "4ef7746362e46005",
     "canonical_url": "https://www.frontiermodelforum.org/updates/introducing-the-fmf-jailbreak-rating-framework-for-frontier-cyber-threats/",
     "url": "https://www.frontiermodelforum.org/updates/introducing-the-fmf-jailbreak-rating-framework-for-frontier-cyber-threats/",
     "title": "Introducing the FMF Jailbreak Rating Framework for Frontier Cyber Threats",
     "summary": "FMF 发布越狱评级框架 1.0：按能力增益、能力范围、部署难度、发现难度四维比较成功网络越狱的相对风险，并开放意见征集；OpenAI 联同 AWS、Anthropic、Google DeepMind、Microsoft 等成员背书，意在统一口径并优先加固高危绕过。",
     "why": "跨厂统一「哪类越狱更危险」的量尺：做红队/安全评测时按四维对照自家发现，别把可复现绕过一律当同等 P0。",
     "tags": [
      "FMF",
      "越狱评级",
      "网络安全"
     ],
     "section": "biz",
     "published_at": "2026-11-07T03:58:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://www.frontiermodelforum.org/updates/introducing-the-fmf-jailbreak-rating-framework-for-frontier-cyber-threats/",
       "title": "Frontier Model Forum · 框架 1.0；OpenAI 等同日转发"
      },
      {
       "title": "OpenAI 原帖",
       "url": "https://x.com/OpenAI/status/2118795218099253606"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "fmf-jailbreak-rating-2026-11",
     "source": "Frontier Model Forum · 框架 1.0；OpenAI 等同日转发"
    },
    {
     "id": "eb3b6803929b7cbc",
     "canonical_url": "https://openai.com/index/partnering-with-hacktron",
     "url": "https://openai.com/index/partnering-with-hacktron",
     "title": "与 Hacktron 合作排查 OpenAI 系统漏洞",
     "summary": "",
     "why": "合作焦点在「系统↔第三方服务」交界，并用 Daybreak Red 把前沿模型交给红队；同周 Cathedral 进冲刺、bug bounty 上调。自建 agent 若大量接外部工具，优先对标交界面怎么选测、frontier 访问怎么授给进攻方。",
     "tags": [
      "OpenAI",
      "Hacktron",
      "Daybreak Red",
      "安全"
     ],
     "section": "biz",
     "published_at": "2026-11-06T16:00:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": 0.537,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://openai.com/index/partnering-with-hacktron",
       "title": "OpenAI · 11-06"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "openai-hacktron-2026-11",
     "source": "OpenAI · 11-06"
    },
    {
     "id": "c954afa3a1d2d839",
     "canonical_url": "https://x.com/KateClarkTweets/status/2118563680694751405",
     "url": "https://x.com/KateClarkTweets/status/2118563680694751405",
     "title": "@KateClarkTweets: Anthropic 年化收入",
     "summary": "",
     "why": "把这条当 IPO 路演期的收入锚点：Kate Clark 援引 WSJ 称 10 月底 run rate 已到 800 亿美元，同链又写大型出资方对 1.5 万亿美元估值存疑——对照估值倍数前先核 run rate 口径，别只拿推文数字外推。",
     "tags": [
      "Anthropic",
      "ARR",
      "WSJ"
     ],
     "section": "biz",
     "published_at": "2026-11-06T12:38:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": 0.537,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://x.com/KateClarkTweets/status/2118563680694751405",
       "title": "X · Kate Clark 援引 WSJ · 11-06"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "anthropic-arr-80b-2026-11",
     "source": "X · Kate Clark 援引 WSJ · 11-06"
    },
    {
     "id": "693fd1cd2f18e56f",
     "canonical_url": "https://x.com/ChatGPT/status/2118827837239226387",
     "url": "https://x.com/ChatGPT/status/2118827837239226387",
     "title": "@ChatGPT: Remix 网站改版",
     "summary": "",
     "why": "把同事现成站当改版起点、用对话下达改动并继续修副本——适合 Business/Enterprise 里「有站可改」的协作流；官帖只写能力与逐步放量，落地前核账号是否已开 Remix。",
     "tags": [
      "ChatGPT",
      "Remix",
      "Business"
     ],
     "section": "biz",
     "published_at": "2026-11-07T06:08:00+08:00",
     "first_seen_at": "2026-11-07",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/ChatGPT/status/2118827837239226387",
       "title": "X · ChatGPT · 11-07 06:08"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "chatgpt-remix-2026-11",
     "source": "X · ChatGPT · 11-07 06:08"
    }
   ],
   "note": "BioBasicsBench / 融资首批 / Muse Spaces 见三条；记忆 dreaming、Amp spaces、Inkling、MNN-Pocket、vLLM 异步 RL、OpenRouter Auto 见专栏/工程/基建；METR≈2×、工信部智能体终端、Ling-3.2-flash、FMF/Hacktron、Remix 与 Hy4 后续见本栏。google/skills 与 Muse 语音昨日报过未重复。"
  },
  {
   "id": "x",
   "title": "X 精选",
   "items": [
    {
     "id": "c50b35f2b3707333",
     "canonical_url": "https://x.com/emollick/status/2118551754908356979",
     "url": "https://x.com/emollick/status/2118551754908356979",
     "title": "@emollick: 缺国家级 AI 速成项目",
     "summary": "",
     "why": "抱怨实验室与政策侧缺少「近期可完成、能做好事」的国家级 crash projects。若你在写公共向 agent 方案，可把它当需求信号：先交可落地的短周期目标，而不是等数据中心里的天才民族。",
     "tags": [
      "工程杂讯",
      "政策",
      "crash projects"
     ],
     "section": "x",
     "published_at": "2026-11-06",
     "first_seen_at": "2026-11-07",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/emollick/status/2118551754908356979",
       "title": "@emollick · 11-06 11:51 · ♥ 275"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 275
     },
     "via": [],
     "source": "@emollick · 11-06 11:51 · ♥ 275"
    },
    {
     "id": "a56f0c931a709ec9",
     "canonical_url": "https://x.com/tison1096/status/2118613982097580220",
     "url": "https://x.com/tison1096/status/2118613982097580220",
     "title": "@tison1096: Codex 回车改插入",
     "summary": "",
     "why": "Codex App：有进行中对话时，回车从「排队」改成「插入」，CMD+Enter 才排队；划线批注时 CMD+Enter 又是直接插入。日常热键肌肉记忆要按场景重训一遍。",
     "tags": [
      "工程杂讯",
      "Codex",
      "Vibe Coding"
     ],
     "section": "x",
     "published_at": "2026-11-06",
     "first_seen_at": "2026-11-07",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/tison1096/status/2118613982097580220",
       "title": "@tison1096 · 11-06 15:58 · ♥ 43"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 43
     },
     "via": [],
     "source": "@tison1096 · 11-06 15:58 · ♥ 43"
    },
    {
     "id": "9083a19f06ce8301",
     "canonical_url": "https://x.com/mattpocockuk/status/2118663407616606637",
     "url": "https://x.com/mattpocockuk/status/2118663407616606637",
     "title": "@mattpocockuk: React 更重要了",
     "summary": "",
     "why": "在引用讨论里强调「懂 React 比以往更重要」。agent 写前端代码量上升时，人侧框架直觉仍是验收与纠偏的底盘，别假想提示词能完全替代。",
     "tags": [
      "工程杂讯",
      "React",
      "工程"
     ],
     "section": "x",
     "published_at": "2026-11-06",
     "first_seen_at": "2026-11-07",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/mattpocockuk/status/2118663407616606637",
       "title": "@mattpocockuk · 11-06 19:14 · ♥ 690"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 690
     },
     "via": [],
     "source": "@mattpocockuk · 11-06 19:14 · ♥ 690"
    }
   ]
  }
 ]
}
