{
 "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": "2026-11-11T07:41:00+08:00",
 "published_at": "2026-11-11T07:54:43+08:00",
 "fetch_window": {
  "start": "2026-11-10T07:24:47+08:00",
  "end": "2026-11-11T07:28:48+08:00"
 },
 "date": "2026-11-11",
 "title": "AI 要闻 · 2026-11-11",
 "url": "https://www.wangwu.ai/daily/2026-11-11/",
 "fetched": "2026-11-11 07:25–07:41（北京时间）",
 "sections": [
  {
   "id": "top3",
   "title": "今日最该知道的 3 条",
   "items": [
    {
     "id": "402de9d749a9b5a1",
     "canonical_url": "https://claude.com/blog/sheets-whiteboards-data",
     "url": "https://claude.com/blog/sheets-whiteboards-data",
     "title": "Claude Whiteboards、Claude Sheets 和 Claude Data",
     "summary": "Whiteboards 与 Sheets 以测试版上线，原 Dashboards 转正并更名为 Claude Data，三者对所有付费方案开放。Data 里认证查询和实时共享仍是测试版，可对接 Redshift、BigQuery、ClickHouse。",
     "why": "文档、白板和表格可以留在同一个工作区里一起改，而不必先导出再同步。若你已经把数据定义交给团队认证，先看认证查询能否避免各写各的指标；连接器只是把应用嵌进来，权限边界还是要单独核。",
     "tags": [
      "Claude",
      "Whiteboards",
      "Sheets",
      "协作",
      "MCP"
     ],
     "section": "top3",
     "published_at": "2026-11-11T01:26:00+08:00",
     "first_seen_at": "2026-11-11",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://claude.com/blog/sheets-whiteboards-data",
       "title": "Anthropic · 11-11；官方博客"
      },
      {
       "title": "协作功能说明",
       "url": "https://claude.com/resources/articles/more-ways-to-work-with-claude-sheets-whiteboards-and-interactive-connectors"
      },
      {
       "title": "@claudeai 官宣",
       "url": "https://x.com/claudeai/status/2120206368619831597"
      },
      {
       "title": "MCP Apps 预览",
       "url": "https://x.com/ClaudeDevs/status/2120227982921097388"
      }
     ],
     "signals": {
      "source_count": 4
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/c1fodrvts3po1pprz70weooju"
      }
     ],
     "key": "claude-sheets-whiteboards-data-2026-11",
     "source": "Anthropic · 11-11；官方博客"
    },
    {
     "id": "c07e63fe47a8c275",
     "canonical_url": "https://cursor.com/blog/projects-multiplayer",
     "url": "https://cursor.com/blog/projects-multiplayer",
     "title": "Cursor Projects 现已支持多人协作",
     "summary": "9 月的 Projects 还是由单一协调者调度一组智能体；现在可邀请同事进入同一 Project，共享任务板、文档和上下文，各自仍保留一对一的协调者对话，不同成员的 agent 也能互相协调。已对所有 Cursor 用户以测试版推出。",
     "why": "这是把「一人一带一组 agent」扩成团队共享工作区。试用时先看共享层会不会把一个人聊天里的决定同步给全员 agent，再约定哪些上下文可共享、哪些应留在私人对话。",
     "tags": [
      "Cursor",
      "Projects",
      "多人协作",
      "共享上下文"
     ],
     "section": "top3",
     "published_at": "2026-11-10T20:00:00+08:00",
     "first_seen_at": "2026-11-11",
     "score": 0.426,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://cursor.com/blog/projects-multiplayer",
       "title": "Cursor · 11-10；官方博客"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/vouc5yc5intvgg7311kekkhsi"
      }
     ],
     "key": "cursor-projects-multiplayer-2026-11",
     "source": "Cursor · 11-10；官方博客"
    },
    {
     "id": "4149f8b4d44cbded",
     "canonical_url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/",
     "url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/",
     "title": "后续｜Gemini 4 Carbon 现已可用",
     "summary": "对付费开发者开放 API（AI Studio 与 Gemini for business），引入价与 Argon 同为每百万 token 输入 $2、输出 $10，11 月 30 日后改为 $4 / $20。CursorBench 4.0、WANDR 与 AutomationBench 均列第一，单任务成本不到次优外部模型的一半。",
     "why": "官方仍把 Argon 作为多数编程与知识工作的默认，Carbon 留给更长、更费 token 的难题。核账单时用引入期截止日和单任务成本，不要只看现在的 $2/$10；要试多代理编排再看 /swarm。",
     "tags": [
      "Gemini 4 Carbon",
      "定价",
      "Antigravity",
      "CursorBench"
     ],
     "section": "top3",
     "published_at": "2026-11-10T16:09:26.000Z",
     "first_seen_at": "2026-11-11",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/",
       "title": "Google DeepMind · 11-10；官方博客"
      },
      {
       "title": "@osanseviero",
       "url": "https://x.com/osanseviero/status/2120187081620418909"
      },
      {
       "title": "Cursor 上线",
       "url": "https://x.com/cursor_ai/status/2120193584670580857"
      }
     ],
     "signals": {
      "source_count": 3
     },
     "via": [],
     "key": "gemini-4-carbon-ga-2026-11",
     "source": "Google DeepMind · 11-10；官方博客"
    }
   ]
  },
  {
   "id": "edge",
   "title": "专栏：端侧模型",
   "intro": "浏览器里直接用 WebGPU 跑 PyTorch 模型；llama.cpp 补上 Kimi K3.1 视觉。",
   "items": [
    {
     "id": "c45365c736f1ab25",
     "canonical_url": "https://pytorch.org/blog/executorch-webgpu-pytorch-models-on-the-gpu-in-the-browser/",
     "url": "https://pytorch.org/blog/executorch-webgpu-pytorch-models-on-the-gpu-in-the-browser/",
     "title": "ExecuTorch WebGPU：在浏览器的 GPU 上运行 PyTorch 模型",
     "summary": "用 torch.export 得到的同一份 .pte，可在任意 WebGPU 浏览器里跑，不必先转成 GGUF、ONNX 或 TFLite。同一台机器上，相对 Wasm + XNNPACK 的 CPU 路径，预填充快 4–7 倍、解码快 2–10 倍；文中基准在 Apple M1 Pro 的 Chrome 上测得。",
     "why": "若你要把现有 PyTorch 模型送到浏览器而不换一套格式，这是一条原生路径。先在 Playground 对一下目标模型和浏览器，再看预填充与解码的加速是否覆盖你的上下文长度。",
     "tags": [
      "ExecuTorch",
      "WebGPU",
      "端侧",
      "浏览器"
     ],
     "section": "edge",
     "published_at": "2026-11-10T00:00:00+00:00",
     "first_seen_at": "2026-11-11",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://pytorch.org/blog/executorch-webgpu-pytorch-models-on-the-gpu-in-the-browser/",
       "title": "PyTorch 博客 · 11-10"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/yz9we1kzdchnco7hw065lbtr5"
      }
     ],
     "key": "executorch-webgpu-2026-11",
     "source": "PyTorch 博客 · 11-10"
    },
    {
     "id": "5569a5cafa3fb43b",
     "canonical_url": "https://github.com/ggml-org/llama.cpp/releases/tag/b12259",
     "url": "https://github.com/ggml-org/llama.cpp/releases/tag/b12259",
     "title": "llama.cpp b12259",
     "summary": "",
     "why": "这条值得看是因为补上了 Kimi K3.1 视觉（并修正 K2.5/K3.1 预处理和投影 MLP），不是又一次后端修补。若你在本地跑月之暗面的多模态权重，先升到这一档再谈画质和预处理差异。",
     "tags": [
      "llama.cpp",
      "Kimi K3.1",
      "视觉",
      "mtmd"
     ],
     "section": "edge",
     "published_at": "2026-11-10T22:11:56Z",
     "first_seen_at": "2026-11-11",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://github.com/ggml-org/llama.cpp/releases/tag/b12259",
       "title": "ggml-org/llama.cpp · 11-11 06:11（北京）"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "llamacpp-b12259-kimi-k31-vision-2026-11",
     "source": "ggml-org/llama.cpp · 11-11 06:11（北京）"
    }
   ],
   "note": "llama.cpp b12237–b12262 多为后端与 CI 修补，仅 b12259 新增 Kimi K3.1 视觉故单列；Muse Ember 已在 11-10 报道，未重复。Siri 调度规则为既有路线的再报道，未收。"
  },
  {
   "id": "col",
   "title": "专栏：Agent 记忆与协作",
   "intro": "Gemini 开始按类别展示可改可删的记忆摘要。",
   "items": [
    {
     "id": "8f2a4c04cc8dd37f",
     "canonical_url": "https://www.ithome.com/1/024/823.htm",
     "url": "https://www.ithome.com/1/024/823.htm",
     "title": "谷歌 Gemini 上线记忆摘要，可查看 / 编辑 / 删除 AI 记住的个人信息",
     "summary": "记忆按个人背景、人际关系、工作、兴趣和偏好等类别整理，可逐项编辑或删除并查看修改历史。入口在「个性化智能服务」下的「记忆摘要」；官方支持页写明目前主要面向美国、年满 18 岁且订阅 Google AI 的用户，IT之家在网页版测到已支持简体中文。",
     "why": "这是把零散记忆条目收成可审查的摘要。若你用 Gemini 做长期助手，先核对它记住了什么、改错了能不能回看历史；地域和订阅限制以支持页为准，不要假定所有账号都已开放。",
     "tags": [
      "Gemini",
      "记忆",
      "隐私"
     ],
     "section": "col",
     "published_at": "2026-11-10T14:39:31+08:00",
     "first_seen_at": "2026-11-11",
     "score": 0.77,
     "importance": 0.8,
     "sources": [
      {
       "url": "https://www.ithome.com/1/024/823.htm",
       "title": "IT之家 · 11-10 14:39；转述 Android Authority，并在网页版核实"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "gemini-memory-summary-2026-11",
     "source": "IT之家 · 11-10 14:39；转述 Android Authority，并在网页版核实"
    }
   ],
   "note": "Cursor 多人 Projects 与 Claude 协作套件见三条。Adobe 创意云本地 MCP 偏生成工具链，只在私有简报留一句。"
  },
  {
   "id": "cl",
   "title": "主流产品 Changelog 精读",
   "intro": "qwen-code 补上通道、Skill 导入和待审批通知；Claude Code 给 Vertex 加了快模式。",
   "items": [
    {
     "id": "5c5d6af7b3841f97",
     "canonical_url": "https://github.com/QwenLM/qwen-code/releases/tag/v0.27.2",
     "url": "https://github.com/QwenLM/qwen-code/releases/tag/v0.27.2",
     "title": "QwenLM/qwen-code v0.27.2",
     "summary": "可从所选 Session 的原 Runtime 工作区导入 Skill，不必重新上传；待审批会出现在实时面板，终端失焦时可响铃或发桌面通知。Web Shell 还能配置微信、QQ、邮箱通道。",
     "why": "相对已报的 v0.27.1 小补丁，这次动到了 Skill 安装路径和值守通知。若你用 Web Shell 或托管会话，先试一次从原工作区导入 Skill，并确认失焦通知不会漏掉待审批。",
     "tags": [
      "qwen-code",
      "Skill",
      "Web Shell"
     ],
     "section": "cl",
     "published_at": "2026-11-10T19:22:00Z",
     "first_seen_at": "2026-11-11",
     "score": 0.405,
     "importance": 0.3,
     "sources": [
      {
       "url": "https://github.com/QwenLM/qwen-code/releases/tag/v0.27.2",
       "title": "QwenLM/qwen-code · 11-11 03:22（北京）"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "qwen-code-v0-27-2-2026-11",
     "source": "QwenLM/qwen-code · 11-11 03:22（北京）"
    },
    {
     "id": "d4fe258087461eef",
     "canonical_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.322",
     "url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.322",
     "title": "后续｜Claude Code v2.1.322",
     "summary": "",
     "why": "增量很窄：只在 Vertex 的 global、us、eu 区域加了快模式，并在 Vertex 拒绝时回退到标准速度。若你走 Vertex 而不是 Anthropic 直连，升这一档后再测延迟；其他平台不必为它单独改流程。",
     "tags": [
      "Claude Code",
      "Vertex",
      "快模式"
     ],
     "section": "cl",
     "published_at": "2026-11-10T02:29:43Z",
     "first_seen_at": "2026-11-11",
     "score": 0.455,
     "importance": 0.3,
     "sources": [
      {
       "url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.322",
       "title": "anthropics/claude-code · 11-10 10:29（北京）"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "claude-code-v2-1-322-2026-11",
     "source": "anthropics/claude-code · 11-10 10:29（北京）"
    }
   ],
   "note": "openai/codex rust-v0.167.1 与 0.169.0-alpha.9 发布说明几乎为空，未收；gemini-cli v0.69.0-preview.0 为预发布，v0.68.0 多为补丁清单未单列。"
  },
  {
   "id": "gh",
   "title": "GitHub 热门",
   "intro": "给 Home Assistant 接上各家编码助手的 MCP 服务上了日榜。",
   "items": [
    {
     "id": "7158ba65d91ac23a",
     "canonical_url": "https://github.com/homeassistant-ai/ha-mcp",
     "url": "https://github.com/homeassistant-ai/ha-mcp",
     "title": "homeassistant-ai/ha-mcp：用各家 AI 助手配置和自动化 Home Assistant",
     "summary": "",
     "why": "仓库说明只有一句：让 Claude、ChatGPT 等助手去配置和自动化 Home Assistant。若你已经把智能家居当成 agent 的执行面，值得看它如何把 MCP 接到真实设备；权限和误操作面比网页自动化更大，先在测试实例上试。",
     "tags": [
      "MCP",
      "Home Assistant",
      "自动化"
     ],
     "section": "gh",
     "published_at": "2026-11-11T07:28:00+08:00",
     "first_seen_at": "2026-11-11",
     "score": 0.605,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://github.com/homeassistant-ai/ha-mcp",
       "title": "GitHub · 日榜；6.6k★ / 今日 +792"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "homeassistant-ai-ha-mcp-2026-11",
     "source": "GitHub · 日榜；6.6k★ / 今日 +792"
    }
   ],
   "note": "meta/astryx 已在 11-09 收录；claude-quickstarts、mattpocock/skills 再上榜无增量；mujoco、sherlock、个人理财应用 sure 与方向无关。billion-context 早前已报。周榜其余与本栏方向或公开规范不符的仓库未收。"
  },
  {
   "id": "eng",
   "title": "Agent 工程",
   "intro": "用 Lean 等价证明验收智能体写出来的加速代码。",
   "items": [
    {
     "id": "cc9d341e50df3aab",
     "canonical_url": "https://rdi.berkeley.edu/blog/aqueduct",
     "url": "https://rdi.berkeley.edu/blog/aqueduct",
     "title": "现实世界软件的智能体验证优化",
     "summary": "Verina 189 个程序里有 156 个写出带 Lean 等价证明的加速版，几何平均快 1.85 倍；最长递增子序列从指数搜索改成动态规划后快 4,468 倍。17 个迁到 Lean 的应用上，92 次被证明的改写让 51 个场景中的 47 个更快，但端到端只快 1.12 倍。",
     "why": "它把原函数当作规格，避免再写一份容易写错的形式化说明。若你用智能体做性能优化，先要求「每个输入都等价」而不是只过测试；应用级 1.12 倍也说明选哪些函数改，往往比单点加速更要紧。",
     "tags": [
      "Aqueduct",
      "Lean",
      "形式化验证",
      "智能体"
     ],
     "section": "eng",
     "published_at": "2026-11-10T08:00:00+08:00",
     "first_seen_at": "2026-11-11",
     "score": 0.645,
     "importance": 0.8,
     "sources": [
      {
       "url": "https://rdi.berkeley.edu/blog/aqueduct",
       "title": "Berkeley RDI / AWS · 11 月；官方博文"
      },
      {
       "title": "项目页",
       "url": "https://aqueduct.verina.io"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/twyiicje97259zbg43dllznnk"
      }
     ],
     "key": "aqueduct-verified-optimization-2026-11",
     "source": "Berkeley RDI / AWS · 11 月；官方博文"
    }
   ],
   "note": "Simon Willison 的排障随笔与 Yonkers 博文未收。Aqueduct 论文 arXiv 编号 2610.30188，公开栏以 11 月的 Berkeley 博文为锚。"
  },
  {
   "id": "infra",
   "title": "检索与基础设施",
   "items": [],
   "note": "今日无重要更新。llm-anthropic 0.31 仅为昨日 Fable 的客户端接入，未单列。"
  },
  {
   "id": "papers",
   "title": "论文",
   "intro": "元认知选课题、训练集群上的自改进，以及把果蝇哈希用到 agent 记忆。",
   "items": [
    {
     "id": "8ed46af327fa6d76",
     "canonical_url": "https://arxiv.org/abs/2611.12762",
     "url": "https://arxiv.org/abs/2611.12762",
     "title": "Sage-Taiyi：面向科学发现的元认知架构",
     "summary": "在含 10 个学科、50 篇论文、265 项任务的 Taiyi-Bench 上，8 项历史复现与开放构想指标全部第一，总分 54.0，对照最强基线 AI Scientist-v2 的 47.9。化学合作里，钴催化羰基化收率从痕量提到 84%，光催化烯丙基胺化从 13% 提到 69%。",
     "why": "它把「下一步研究什么、用哪套方法」写成可修订的监控与控制，而不只是多采样几个想法。若你在做科研助手，先看课题选择和可复用推理方法是否被分开评估，再把实验验证当成独立门槛。",
     "tags": [
      "Sage-Taiyi",
      "元认知",
      "科学发现",
      "智能体"
     ],
     "section": "papers",
     "published_at": "2026-11-10T00:00:00.000Z",
     "first_seen_at": "2026-11-11",
     "score": 0.449,
     "importance": 0.225,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2611.12762",
       "title": "arXiv · 2611.12762；HF Daily Papers 11-10"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "arxiv-2611-12762-sage-taiyi-2026-11",
     "source": "arXiv · 2611.12762；HF Daily Papers 11-10"
    },
    {
     "id": "29cef194a466479b",
     "canonical_url": "https://arxiv.org/abs/2611.11558",
     "url": "https://arxiv.org/abs/2611.11558",
     "title": "InfraRSI：用递归自改进优化训练基础设施",
     "summary": "五阶段流程先锁定可改范围和损失、时序不变量，再联合调整启动参数与代码补丁，并把经验写入跨硬件库。在 4 个单机与多机、稠密与 MoE 工作负载上最高加速 2.75 倍，平均 MFU 从 17.4% 升到 30.0%。",
     "why": "训练集群上的自改进很容易被 OOM 和数值漂移奖励黑客。若你要把智能体接到训练栈，先把工作负载、损失和时序不变量写成不可改的合同，再允许它改启动参数和内核。",
     "tags": [
      "InfraRSI",
      "RSI",
      "训练",
      "MFU"
     ],
     "section": "papers",
     "published_at": "2026-11-10T00:00:00.000Z",
     "first_seen_at": "2026-11-11",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2611.11558",
       "title": "arXiv · 2611.11558；HF Daily Papers 11-10"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "arxiv-2611-11558-infrarsi-2026-11",
     "source": "arXiv · 2611.11558；HF Daily Papers 11-10"
    },
    {
     "id": "9d60857e80536a09",
     "canonical_url": "https://arxiv.org/abs/2611.05791",
     "url": "https://arxiv.org/abs/2611.05791",
     "title": "FlyHash 用于大模型智能体记忆：实证与新应用",
     "summary": "在匹配比特、10 个种子和四个数据集上，稀疏随机投影加赢者通吃的质量优势成立；但同等比特下编码比 SimHash 更大更慢，常规模下 HNSW 仍全面占优。作者同时放了实现和可运行的记忆演示，并纠正了文献里一处未记录的对比陷阱。",
     "why": "质量更好不等于该替换你现在的向量索引。若你在评估 agent 记忆检索，把 HNSW 和延迟、编码体积放进同一张对照表，不要只看近邻保持率。",
     "tags": [
      "FlyHash",
      "记忆",
      "检索",
      "HNSW"
     ],
     "section": "papers",
     "published_at": "2026-11-10T00:00:00.000Z",
     "first_seen_at": "2026-11-11",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2611.05791",
       "title": "arXiv · 2611.05791；HF Daily Papers 11-10"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "arxiv-2611-05791-flyhash-2026-11",
     "source": "arXiv · 2611.05791；HF Daily Papers 11-10"
    }
   ],
   "note": "论文日期以 HF Daily Papers 11-10 列表日为准。具身操作、皮肤科 RAG 与触觉表示未收。"
  },
  {
   "id": "biz",
   "title": "大厂与模型",
   "intro": "ChatGPT 给商家开了认领与商务智能体；谷歌与 Artificial Analysis 交出多模态知识工作基准。",
   "items": [
    {
     "id": "b831185d24338583",
     "canonical_url": "https://openai.com/index/introducing-businesses-chatgpt/",
     "url": "https://openai.com/index/introducing-businesses-chatgpt/",
     "title": "介绍 ChatGPT 中的 Businesses",
     "summary": "测试版里企业可在 Business Hub 认领并更新资料、查看汇总的提及与点击洞察；美国合格电商与在线服务还可上线 Business Agent，用户从企业主页另开单独标注的对话询价、比产品。洞察不向企业提供用户原对话，Shopify 商户可从 Hub 或后台完成验证。",
     "why": "这是把「在 ChatGPT 里被提到」做成可认领的商家主页和可对话的商务智能体。若你关注 agent 如何面对最终用户，先分清它和模型独立回答是分开标注的会话，再看隐私阈值和广告赞助位怎么披露。",
     "tags": [
      "ChatGPT",
      "Business Agent",
      "Business Hub"
     ],
     "section": "biz",
     "published_at": "2026-11-10T00:00:00+00:00",
     "first_seen_at": "2026-11-11",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://openai.com/index/introducing-businesses-chatgpt/",
       "title": "OpenAI · 11-10；官方博客"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/",
       "item_url": "https://aihot.news/items/vhu2xtpasbm5mewe5r7qdqkgi"
      }
     ],
     "key": "openai-businesses-chatgpt-2026-11",
     "source": "OpenAI · 11-10；官方博客"
    },
    {
     "id": "a8d2283cd895ddd7",
     "canonical_url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/",
     "url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/",
     "title": "介绍 Multimodal Knowledge Agent 基准",
     "summary": "与 Artificial Analysis 合作的 200 道私有任务覆盖 10 个行业、83 种职业，每题约 70 条是非量表（事实 75%、视觉设计 25%），用开源 Stirrup harness 打分，与专家一致性 96%。Gemini 4 Carbon 质量第二，平均成本约为最高分模型的六分之一。",
     "why": "它针对的是看图纸、报表和版式才能做完的知识工作，而不是纯文本问答。若你要拿来做选型，先看自己的任务是否真的依赖视觉文件；官方同时强调 200 题集不公开，避免漏进训练集。",
     "tags": [
      "MKA",
      "基准",
      "多模态",
      "Gemini 4 Carbon"
     ],
     "section": "biz",
     "published_at": "2026-11-10T00:00:00+00:00",
     "first_seen_at": "2026-11-11",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/",
       "title": "Google DeepMind · 11-10；官方博客"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "mka-benchmark-2026-11",
     "source": "Google DeepMind · 11-10；官方博客"
    }
   ],
   "note": "Claude 协作套件、Cursor 多人 Projects、Gemini 4 Carbon 见三条；ExecuTorch WebGPU、llama.cpp Kimi K3.1 视觉、Gemini 记忆摘要、qwen-code v0.27.2、Claude Code Vertex 快模式、ha-mcp、Aqueduct、三篇论文、ChatGPT Businesses 与 MKA 见各栏。Fable/Mythos、Muse Ember、Control Room、TPU、千问办公、gbrain、LongCat 已报未重复；融资八卦、图像生成、空发布说明与预览版、机器人与皮肤科论文未收。"
  },
  {
   "id": "x",
   "title": "X 精选",
   "items": [
    {
     "id": "e8d6eb506e7c5ff3",
     "canonical_url": "https://x.com/emollick/status/2120000888974766501",
     "url": "https://x.com/emollick/status/2120000888974766501",
     "title": "@emollick: Fable 在未饱和基准高于趋势线",
     "summary": "",
     "why": "他指出 Fable 5.2 在 ECI、Remote Labor、RoboDojo 这三条仍未饱和的基准上高于既有陡峭趋势线。这是对昨日官方分数的独立对照，适合用来看长程任务，而不是再读一遍发布稿。",
     "tags": [
      "工程杂讯",
      "Fable 5.2",
      "评测"
     ],
     "section": "x",
     "published_at": "2026-11-10",
     "first_seen_at": "2026-11-11",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://x.com/emollick/status/2120000888974766501",
       "title": "@emollick · 11-10 11:49 · ♥ 1436"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 1436
     },
     "via": [],
     "source": "@emollick · 11-10 11:49 · ♥ 1436"
    }
   ]
  }
 ]
}
