{
 "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": "2026-11-17T08:22:45+08:00",
 "published_at": "2026-11-17T08:35:17+08:00",
 "fetch_window": {
  "start": "2026-11-16T13:45:39+08:00",
  "end": "2026-11-17T08:22:25+08:00"
 },
 "date": "2026-11-17",
 "title": "AI 要闻 · 2026-11-17",
 "url": "https://www.wangwu.ai/daily/2026-11-17/",
 "fetched": "2026-11-17 08:22（北京时间）",
 "sections": [
  {
   "id": "top3",
   "title": "今日最该知道的 3 条",
   "items": [
    {
     "id": "b25cc016d6d94348",
     "canonical_url": "https://www.marktechpost.com/2026/11/16/alibaba-qwen4-max-release/",
     "url": "https://www.marktechpost.com/2026/11/16/alibaba-qwen4-max-release/",
     "title": "阿里发布 Qwen4-Max：2.4T MoE、71B 激活、1M 上下文",
     "summary": "团队报告预训练全程无损失尖峰，基座每字节比特数较上代降 5%–13% 且激活量少 25%，Work360 单任务均价降 62%；采用 1 层 QSA 微块稀疏注意力配 3 层 GDN、4 分支门控残差与 328B N-gram 嵌入，支持 393K 输出且未开放权重。",
     "why": "每层 QSA 配 3 层 GDN，后者将历史压进定长状态，配合微块检索降低长序列注意力与索引开销；闭源托管意味着无法本地量化部署，接入前可用长代码库实测上下文召回遗漏与跨轮工具调用稳定性。",
     "tags": [
      "Qwen4-Max",
      "MoE",
      "稀疏注意力"
     ],
     "section": "top3",
     "published_at": "2026-11-16",
     "first_seen_at": "2026-11-17",
     "score": 0.704,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://www.marktechpost.com/2026/11/16/alibaba-qwen4-max-release/",
       "title": "MarkTechPost · 11-16"
      },
      {
       "title": "Qwen 博客",
       "url": "https://qwen.ai/blog?id=qwen4-max"
      },
      {
       "title": "量子位",
       "url": "https://www.qbitai.com/2026/11/510486.html"
      }
     ],
     "signals": {
      "source_count": 3
     },
     "via": [],
     "source": "MarkTechPost · 11-16"
    },
    {
     "id": "4f168122f29dbd3b",
     "canonical_url": "https://x.com/_philschmid/status/2122358498461691913",
     "url": "https://x.com/_philschmid/status/2122358498461691913",
     "title": "@_philschmid: Gemini 4 与 Decisions API",
     "summary": "",
     "why": "模型接入与沙箱部署需要分开验收：先用逐步增长的输入探测上下文上限和账单计量，再在沙箱中测试外网访问、文件挂载及超时终止；上线前保存所用模型标识与价格生效日期，避免限时价格被写成长期成本假设。",
     "tags": [
      "Gemini 4",
      "Decisions API",
      "Agent沙箱"
     ],
     "section": "top3",
     "published_at": "2026-11-16",
     "first_seen_at": "2026-11-17",
     "score": 0.662,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://x.com/_philschmid/status/2122358498461691913",
       "title": "X (@_philschmid) · 11-16"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "X (@_philschmid) · 11-16"
    },
    {
     "id": "8be4de084fcea68c",
     "canonical_url": "https://blogs.nvidia.com/blog/rubin-hpc/",
     "url": "https://blogs.nvidia.com/blog/rubin-hpc/",
     "title": "NVIDIA 发布 Rubin HPC：面向科学、工程与分析加速智能体的新 GPU",
     "summary": "预计 2028 年上市，配合 CUDA-X 库加速稀疏矩阵、模板计算、排序与归约等高频算子；隶属 Vera Rubin 平台，单卡集成超 5 万颗 CUDA 核心，结合稀疏数据加速与第七代 NVLink 同步扩展数据搬运与算力。",
     "why": "传统仿真受限于不规则访存导致的算力空转，Rubin HPC 将数据搬运与稀疏加速结合，旨在缩短智能体假设验证闭环；鉴于预计 2028 年交付，现阶段可先梳理现有仿真的访存瓶颈，并评估 CUDA-X 算子接口适配。",
     "tags": [
      "NVIDIA",
      "Rubin HPC",
      "科学计算"
     ],
     "section": "top3",
     "published_at": "2026-11-17",
     "first_seen_at": "2026-11-17",
     "score": 0.645,
     "importance": 0.8,
     "sources": [
      {
       "url": "https://blogs.nvidia.com/blog/rubin-hpc/",
       "title": "NVIDIA 官方博客 · 11-17"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "NVIDIA 官方博客 · 11-17"
    }
   ]
  },
  {
   "id": "edge",
   "title": "专栏：端侧模型",
   "intro": "MLX 多模态服务补齐分布式运行与恢复能力，llama.cpp 修正 Vulkan top-k 平局选择的确定性。",
   "items": [
    {
     "id": "7ec6f6c95354afde",
     "canonical_url": "https://github.com/Neywa-Labs/mlx-vlm/releases/tag/v0.8.0",
     "url": "https://github.com/Neywa-Labs/mlx-vlm/releases/tag/v0.8.0",
     "title": "Neywa-Labs/mlx-vlm v0.8.0",
     "summary": "服务端新增分布式流式推理、可恢复服务、RDMA 链路保留与集群投机草稿模型，重启后复用已验证检查点，并在 /v1/embeddings 后清理 MLX 缓存；训练端加入 QAT 与 3/6-bit DoRA 宽度保持，原生适配 DeepSeek V4.1 DSpark 等模型。",
     "why": "此次更新将分布式流式推理、可恢复服务与集群投机草稿模型整合进服务端；多机部署时需重点压测重启后的检查点复用逻辑，以及连续向量嵌入请求下的内存水位，确认缓存清理与恢复路径符合实际负载。",
     "tags": [
      "MLX",
      "多模态",
      "端侧推理"
     ],
     "section": "edge",
     "published_at": "2026-11-17",
     "first_seen_at": "2026-11-17",
     "score": 0.528,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://github.com/Neywa-Labs/mlx-vlm/releases/tag/v0.8.0",
       "title": "GitHub Release · 11-17"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "GitHub Release · 11-17"
    },
    {
     "id": "4980f3198d1192d8",
     "canonical_url": "https://github.com/ggml-org/llama.cpp/releases/tag/b12435",
     "url": "https://github.com/ggml-org/llama.cpp/releases/tag/b12435",
     "title": "ggml-org/llama.cpp b12435",
     "summary": "在 Vulkan 后端开启确定性 top-k 校验并针对 AMD、Intel 与 Nvidia 着色器做专门优化；改由各 subgroup 负责固定连续列块并经前缀扫描与 ballot 按最小列号裁决平局，消除原子计数受线程调度影响的非确定性。",
     "why": "平局元素的线程竞争会妨碍复现实验，该补丁让 Vulkan 的 top-k 平局选择规则与 CUDA 及 Metal 对齐；管线要求 subgroup arithmetic、ballot 和完整 subgroup，升级后应核验目标驱动支持并回归测试平局输入。",
     "tags": [
      "llama.cpp",
      "Vulkan",
      "采样确定性"
     ],
     "section": "edge",
     "published_at": "2026-11-16",
     "first_seen_at": "2026-11-17",
     "score": 0.683,
     "importance": 0.475,
     "sources": [
      {
       "url": "https://github.com/ggml-org/llama.cpp/releases/tag/b12435",
       "title": "GitHub Release · 11-16"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "GitHub Release · 11-16"
    }
   ],
   "note": "llama.cpp b12430 的 HIP MoE 修复、b12431 的 reasoning_effort 校验及 b12432 的 kv-cells 扫描优化合并说明，不另列卡片；LiteRT 因正文未抓全未单列。"
  },
  {
   "id": "col",
   "title": "专栏：Agent 记忆与协作",
   "items": [],
   "note": "stitch-skills、ffmpeg-skill、last30days-skill 与 anthropics/skills 仅有榜单及当前 README，缺少窗口内实质发布的日期证据，本期未单列。"
  },
  {
   "id": "cl",
   "title": "主流产品 Changelog 精读",
   "intro": "Claude Code 修复后台网关与进程生命周期问题，Qwen Code 夜间版加固会话文件校验及重连恢复。",
   "items": [
    {
     "id": "f788c225674a20a5",
     "canonical_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.326",
     "url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.326",
     "title": "anthropics/claude-code v2.1.326",
     "summary": "对嵌套超 64 层的配置值丢弃并告警，修复后台服务误杀复用 PID 的外部进程及后台会话丢失网关后继续请求的问题；支持开启团队模式时启用 Code Mode，修正仅源路径受限时的 cp 误拦与覆写命令摘要误标。",
     "why": "后台会话丢失网关环境变量后立即阻断请求，避免该路径下意外绕过指定网关；维护托管配置与自定义插件的团队升级后，可借助新增的启动告警排查残缺的工具白名单规则与深层嵌套字段。",
     "tags": [
      "Claude Code",
      "权限安全",
      "Agent Teams"
     ],
     "section": "cl",
     "published_at": "2026-11-17",
     "first_seen_at": "2026-11-17",
     "score": 0.455,
     "importance": 0.3,
     "sources": [
      {
       "url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.326",
       "title": "GitHub Release · 11-17"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "GitHub Release · 11-17"
    },
    {
     "id": "ffcbdd7db77aab71",
     "canonical_url": "https://github.com/QwenLM/qwen-code/releases/tag/v0.27.4-nightly.20261116.eb3a599ef3",
     "url": "https://github.com/QwenLM/qwen-code/releases/tag/v0.27.4-nightly.20261116.eb3a599ef3",
     "title": "QwenLM/qwen-code v0.27.4-nightly.20261116.eb3a599ef3",
     "summary": "在已打开文件描述符上二次校验会话记录类型以防御 TOCTOU 竞争，断线重连后自动刷新 SDK 回放能力，并将过期子任务恢复收敛为终态；新增钉钉托管通道适配器、Web Shell 工作区过滤及原生 REST 技能变更路由。",
     "why": "过期子任务恢复不再无限重试，fd 级类型复核则缩小了会话文件检查与使用之间的竞争窗口；自建 Agent Host 服务跟进此预发布构建时，应回归共享工具批次路由、重连后的回放能力刷新及能力协商超时行为。",
     "tags": [
      "Qwen Code",
      "ACP",
      "安全加固"
     ],
     "section": "cl",
     "published_at": "2026-11-17",
     "first_seen_at": "2026-11-17",
     "score": 0.405,
     "importance": 0.3,
     "sources": [
      {
       "url": "https://github.com/QwenLM/qwen-code/releases/tag/v0.27.4-nightly.20261116.eb3a599ef3",
       "title": "GitHub Release · 11-17"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "source": "GitHub Release · 11-17"
    }
   ],
   "note": "Codex alpha.8–alpha.10 与 claude-agent-sdk-python v0.2.183 的抓取正文未提供具体变更说明，不以安装信息或构建元数据填充卡片；Qwen Code 条目为预发布版本。"
  },
  {
   "id": "gh",
   "title": "GitHub 热门",
   "items": [],
   "note": "SkillEvaluator、paper-orchestra、Claude-BugHunter、MuJoCo 与 Sherlock 的材料未证明窗口内有实质发布，榜单收录不作为发布日期，故未单列；Qwen Code 更新见 Changelog。"
  },
  {
   "id": "eng",
   "title": "Agent 工程",
   "items": [],
   "note": "量子位的 Qwen4 报道与 Top 3 为同一发布事件，作为延伸链接保留，不重复列卡。"
  },
  {
   "id": "infra",
   "title": "检索与基础设施",
   "items": []
  },
  {
   "id": "papers",
   "title": "论文",
   "items": [],
   "note": "本期输入未提供可核实窗口内发表日期的独立论文候选。"
  },
  {
   "id": "biz",
   "title": "大厂与模型",
   "intro": "Google 推出多模态知识智能体基准，OpenAI 扩大科研模型使用计划并引入外部嵌入式评估。",
   "items": [
    {
     "id": "bbfbe843ec6e1402",
     "canonical_url": "https://x.com/OpenAI/status/2122359930628063248",
     "url": "https://x.com/OpenAI/status/2122359930628063248",
     "title": "@OpenAI: 联合四机构开展嵌入式安全评估",
     "summary": "",
     "why": "Apollo Research、METR、Transluce 与 Goodfire 可深入接触内部模型和系统，独立审查的价值仍需看评估范围与整改跟踪；Goodfire 另明确参与训练监控、调试与对齐，不能据此推定四家机构拥有相同访问权限。",
     "tags": [
      "OpenAI",
      "安全对齐",
      "模型可解释性"
     ],
     "section": "biz",
     "published_at": "2026-11-17",
     "first_seen_at": "2026-11-17",
     "score": 0.588,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://x.com/OpenAI/status/2122359930628063248",
       "title": "X (@OpenAI) · 11-17"
      },
      {
       "title": "@OpenAI: Goodfire 可解释性合作",
       "url": "https://x.com/OpenAI/status/2122359937649242197"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "source": "X (@OpenAI) · 11-17"
    }
   ],
   "note": "Qwen4-Max、Gemini 4 与 NVIDIA Rubin HPC 见 Top 3；Goodfire 可解释性合作并入嵌入式评估条目，避免同一合作重复列卡。"
  },
  {
   "id": "x",
   "title": "X 精选",
   "items": []
  }
 ]
}
