{
 "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": "2026-11-12T07:41:16+08:00",
 "published_at": "2026-11-12T07:44:30+08:00",
 "fetch_window": {
  "start": "2026-11-11T07:28:48+08:00",
  "end": "2026-11-12T07:26:30+08:00"
 },
 "date": "2026-11-12",
 "title": "AI 要闻 · 2026-11-12",
 "url": "https://www.wangwu.ai/daily/2026-11-12/",
 "fetched": "2026-11-12 07:22–07:45（北京时间）",
 "sections": [
  {
   "id": "top3",
   "title": "今日最该知道的 3 条",
   "items": [
    {
     "id": "a70fe30e3911ab6c",
     "canonical_url": "https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/",
     "url": "https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/",
     "title": "Gemini Decisions API：快速、校准的多模态决策",
     "summary": "单次前向即可给最多 1,000 道是否/选择题或量表题打出概率和置信度，端到端延迟 150–250 毫秒，输入每百万 token 0.10 美元且不收取输出。上下文可到 256K，并直接接受文本、图片、音频、视频和 PDF。",
     "why": "这是给路由、合规检查和打分用的闭集接口，不必再让生成模型逐字写出答案。若你已经在分流模型或选工具，先看单次能带多少题、校准是否够稳，再决定要不要换掉现在的生成式打分。",
     "tags": [
      "Gemini Decision 1",
      "Decisions API",
      "路由",
      "打分"
     ],
     "section": "top3",
     "published_at": "2026-11-11T00:00:00+00:00",
     "first_seen_at": "2026-11-12",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/",
       "title": "Google DeepMind · 11-11；官方博客"
      },
      {
       "title": "文档与 AI Studio",
       "url": "https://simonwillison.net/2026/Nov/11/gemini-decisions-api/"
      },
      {
       "title": "@OfficialLoganK",
       "url": "https://x.com/OfficialLoganK/status/2120581433639198885"
      },
      {
       "title": "@osanseviero",
       "url": "https://x.com/osanseviero/status/2120582290976002177"
      }
     ],
     "signals": {
      "source_count": 4
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "gemini-decisions-api-2026-11",
     "source": "Google DeepMind · 11-11；官方博客"
    },
    {
     "id": "6337dd63b6b28dff",
     "canonical_url": "https://aws.amazon.com/blogs/aws/test-your-applications-against-aws-locally-with-aws-emulator/",
     "url": "https://aws.amazon.com/blogs/aws/test-your-applications-against-aws-locally-with-aws-emulator/",
     "title": "在本地用 AWS Emulator 测试你的应用",
     "summary": "在本机提供本地 HTTPS 接口，AWS CLI、SDK 和编码智能体用同一套云上 API 对它发请求，不必先部署到某个区域。Linux 用 RPM 或 Debian 安装，macOS 与 Windows 需自行用文档里的 Dockerfile 运行；源码公开，不另收模拟器服务费。",
     "why": "智能体改基础设施时，最怕每改一次都要等一次云上部署。若你用编码助手写 AWS 代码，先在自己电脑上把创建、写入、读回的闭环跑通，再决定哪些步骤必须上真实账号。",
     "tags": [
      "AWS Emulator",
      "编码智能体",
      "本地测试"
     ],
     "section": "top3",
     "published_at": "2026-11-11T22:30:00+08:00",
     "first_seen_at": "2026-11-12",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://aws.amazon.com/blogs/aws/test-your-applications-against-aws-locally-with-aws-emulator/",
       "title": "AWS 新闻博客 · 11-11"
      },
      {
       "title": "Werner Vogels 转 Swami 客座文",
       "url": "https://www.allthingsdistributed.com/2026/11/aws-emulator-and-the-engineers-job-when-agents-write-the-code.html"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "aws-emulator-2026-11",
     "source": "AWS 新闻博客 · 11-11"
    },
    {
     "id": "96936a6fd53c22fa",
     "canonical_url": "https://alignment.anthropic.com/2026/worse-together/",
     "url": "https://alignment.anthropic.com/2026/worse-together/",
     "title": "Worse Together：多用户多智能体团队中性能如何崩塌",
     "summary": "四个环境里，每人各派一个智能体都不如由一个智能体服务整组，Claude 两种型号上协调者比点对点高出约 10 到 45 个百分点。共享额度时指定队长并写明优先级会缩小差距；下单前强制读完待办消息，能挽回 Opus 5 失败案例中的 73%。",
     "why": "多人各带一个助手进同一套代码库或同一笔预算时，失败常常出在互不等待、互相覆盖和编造他人意图。若你已经在跑多会话协作，先把「谁拍板、信息有没有进上下文」写成硬门槛，再考虑加不加更多并行会话。",
     "tags": [
      "MAMUBench",
      "多智能体",
      "协作"
     ],
     "section": "top3",
     "published_at": "2026-11-11T08:00:00+08:00",
     "first_seen_at": "2026-11-12",
     "score": 0.645,
     "importance": 0.8,
     "sources": [
      {
       "url": "https://alignment.anthropic.com/2026/worse-together/",
       "title": "Anthropic Alignment Science · 11-11"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "mamubench-worse-together-2026-11",
     "source": "Anthropic Alignment Science · 11-11"
    }
   ]
  },
  {
   "id": "edge",
   "title": "专栏：端侧模型",
   "intro": "不到十亿参数的端侧助手，以及 llama.cpp 接上 CUDA FP8。",
   "items": [
    {
     "id": "5821cbc55937e13b",
     "canonical_url": "https://huggingface.co/blog/qvac/botpsy",
     "url": "https://huggingface.co/blog/qvac/botpsy",
     "title": "BotPsy：配备不到十亿参数智能体语言模型的端侧个人助手",
     "summary": "0.8B 参数、基于 Qwen3.5-0.8B，在作者测试的不超过 1.2B 的模型里工具调用和智能体任务全面领先，六个基准中拿下五个第一，只在 Claw-Eval 上输给大 2.5 倍的 Qwen3.5-2B-Think。",
     "why": "这是把「不一定要云端大模型」落到可下载的端侧助手。若你关心隐私和本地工具调用，先看它声明只适合中低难度任务，并对照自己的工具集，不要把它当成高风险决策模型。",
     "tags": [
      "BotPsy",
      "端侧",
      "工具调用",
      "Qwen3.5"
     ],
     "section": "edge",
     "published_at": "2026-11-11T00:00:00+00:00",
     "first_seen_at": "2026-11-12",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://huggingface.co/blog/qvac/botpsy",
       "title": "Hugging Face 博客 / Tether AI Research · 11-11"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "botpsy-0-8b-2026-11",
     "source": "Hugging Face 博客 / Tether AI Research · 11-11"
    },
    {
     "id": "10b540e38cb057f3",
     "canonical_url": "https://github.com/ggml-org/llama.cpp/releases/tag/b12284",
     "url": "https://github.com/ggml-org/llama.cpp/releases/tag/b12284",
     "title": "llama.cpp b12284",
     "summary": "CUDA 路径补上 OCP FP8，GEMV 走 W8A16。作者在 RTX PRO 4500 Blackwell 上用 Qwen3.5 27B NVFP4 测得预填充 2781.53 tok/s、解码 37.56 tok/s（ubatch 512）。",
     "why": "若你在 NVIDIA 新卡上跑 FP8 或 NVFP4 权重，这是运行时真正接上格式的一步。同窗口还有 b12274 的 FP8 KV cache，可以一并验证显存和速度，不必把中间的后端修补逐条跟。",
     "tags": [
      "llama.cpp",
      "FP8",
      "CUDA"
     ],
     "section": "edge",
     "published_at": "2026-11-11T13:13:19Z",
     "first_seen_at": "2026-11-12",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://github.com/ggml-org/llama.cpp/releases/tag/b12284",
       "title": "ggml-org/llama.cpp · 11-11 21:13（北京）"
      },
      {
       "title": "llama.cpp b12274：FP8 KV cache",
       "url": "https://github.com/ggml-org/llama.cpp/releases/tag/b12274"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "key": "llamacpp-b12284-ocp-fp8-2026-11",
     "source": "ggml-org/llama.cpp · 11-11 21:13（北京）"
    }
   ],
   "note": "llama.cpp 同窗口多为后端修补，仅列出新增 CUDA OCP FP8 的 b12284，并把 b12274 的 FP8 KV cache 放在相关链接。Unsloth 的 GGUF 量化帖为短讯，未单列。OpenRouter 的 fp8/fp4 对比见工程栏相邻主题，放在检索与量化相关的论文和实测里分别处理。"
  },
  {
   "id": "col",
   "title": "专栏：Agent 记忆与协作",
   "intro": "Google Labs 把共享计划做成可同时编辑的实验；Mastra 让长耗时 MCP 调用改走后台任务。",
   "items": [
    {
     "id": "f7bfba52d44a0494",
     "canonical_url": "https://blog.google/innovation-and-ai/models-and-research/google-labs/tabletop/",
     "url": "https://blog.google/innovation-and-ai/models-and-research/google-labs/tabletop/",
     "title": "试试 Tabletop：供小组一起规划、玩耍和协作的可定制空间",
     "summary": "目前只对美国 18 岁以上用户开放候补。空间建好后把邀请链接丢进群聊，成员可同时编辑行程、投票、分账或赛程；材料是你放进去的链接和文件，生成则靠 @tabletop 或模板。",
     "why": "这是面向朋友和家庭的共享画布，不是新的团队知识库。若你关心「多人同时改同一份计划」，先看候补范围，权限和年龄限制以官方说明为准。",
     "tags": [
      "Tabletop",
      "Google Labs",
      "协作"
     ],
     "section": "col",
     "published_at": "2026-11-11T00:00:00+00:00",
     "first_seen_at": "2026-11-12",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://blog.google/innovation-and-ai/models-and-research/google-labs/tabletop/",
       "title": "Google Labs / Roman Nurik · 11-11"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "google-labs-tabletop-2026-11",
     "source": "Google Labs / Roman Nurik · 11-11"
    },
    {
     "id": "23c33eec4c6401e4",
     "canonical_url": "https://mastra.ai/blog/introducing-mcp-tasks",
     "url": "https://mastra.ai/blog/introducing-mcp-tasks",
     "title": "Mastra 推出 MCP Tasks",
     "summary": "MCP 工具改为先返回任务 ID，客户端再轮询 working、input_required、completed、failed 或 cancelled，避免长连接被代理超时打断。任务完成后会在下一次模型请求时通知仍在运行的智能体。",
     "why": "若你已经用 Mastra 接了会跑很久的 MCP 工具，先看轮询和通知能不能让对话继续，而不要再把整段工作堵在一次 HTTP 请求里。这是接在他们 MCP 2.0 和事件机制后面的能力。",
     "tags": [
      "Mastra",
      "MCP",
      "后台任务"
     ],
     "section": "col",
     "published_at": "2026-11-11T08:00:00+08:00",
     "first_seen_at": "2026-11-12",
     "score": 0.582,
     "importance": 0.5,
     "sources": [
      {
       "url": "https://mastra.ai/blog/introducing-mcp-tasks",
       "title": "Mastra · 11-11"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "mastra-mcp-tasks-2026-11",
     "source": "Mastra · 11-11"
    }
   ],
   "note": "Cursor 多人 Projects、Claude 协作套件已在 11-11 报道，未重复。公开栏不收录内部协作产品发布。"
  },
  {
   "id": "cl",
   "title": "主流产品 Changelog 精读",
   "intro": "Qwen Code 开出 1.0 预览线；Codex 在 Vertex 上默认推送增量工具更新。",
   "items": [
    {
     "id": "431d7c7dc4b05ad7",
     "canonical_url": "https://github.com/QwenLM/qwen-code/releases/tag/v1.0.0-preview",
     "url": "https://github.com/QwenLM/qwen-code/releases/tag/v1.0.0-preview",
     "title": "QwenLM/qwen-code v1.0.0-preview",
     "summary": "把 Qoder 的 Harness 并进开源内核，提供 macOS、Linux、Windows 原生日包，不必再单独装 Node.js。预览分支沿统一事件模型往多节点扩展，推理编排和工具执行分开管理；现有稳定分支保留。",
     "why": "相对昨日刚报的 v0.27.2，这次是 1.0 预览线，接口还可能改。若你要把编程智能体嵌进自己的产品，先在预览分支看会话恢复和非交互执行，不要把它当成已经冻结的稳定版。",
     "tags": [
      "qwen-code",
      "Harness",
      "Qoder"
     ],
     "section": "cl",
     "published_at": "2026-11-11T03:55:44Z",
     "first_seen_at": "2026-11-12",
     "score": 0.466,
     "importance": 0.475,
     "sources": [
      {
       "url": "https://github.com/QwenLM/qwen-code/releases/tag/v1.0.0-preview",
       "title": "QwenLM/qwen-code · 11-11 11:55（北京）；预览版"
      },
      {
       "title": "IT之家报道",
       "url": "https://www.ithome.com/1/025/393.htm"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "qwen-code-v1-0-0-preview-2026-11",
     "source": "QwenLM/qwen-code · 11-11 11:55（北京）；预览版"
    },
    {
     "id": "2beeead61aca412d",
     "canonical_url": "https://github.com/openai/codex/releases/tag/rust-v0.167.2",
     "url": "https://github.com/openai/codex/releases/tag/rust-v0.167.2",
     "title": "openai/codex rust-v0.167.2",
     "summary": "",
     "why": "相对说明几乎为空的 0.167.1，这次只动了 Vertex 上的工具增量推送，而且默认打开。若你在 GCP 上跑 Codex，先确认工具更新是不是变成流式的，不需要时再按配置关掉。",
     "tags": [
      "Codex",
      "Vertex AI"
     ],
     "section": "cl",
     "published_at": "2026-11-11T22:44:46Z",
     "first_seen_at": "2026-11-12",
     "score": 0.622,
     "importance": 0.3,
     "sources": [
      {
       "url": "https://github.com/openai/codex/releases/tag/rust-v0.167.2",
       "title": "openai/codex · 11-12 06:44（北京）"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "codex-rust-v0-167-2-2026-11",
     "source": "openai/codex · 11-12 06:44（北京）"
    }
   ],
   "note": "openai/codex 多个 0.169/0.170 alpha 发布说明几乎为空，未收。qwen-code 稳定分支 v0.27.2 已在 11-11 报道，今天只收 1.0 预览。zed 补丁与预发布、opencode process-daemon 预览未单列。"
  },
  {
   "id": "gh",
   "title": "GitHub 热门",
   "intro": "OpenAI 把安全扫描做成 Codex 的 CLI；shadcn 用技能把审计和执行拆开。",
   "items": [
    {
     "id": "d761ebc8c45bfd95",
     "canonical_url": "https://github.com/openai/codex-security",
     "url": "https://github.com/openai/codex-security",
     "title": "openai/codex-security：发现、验证并修复安全漏洞的 CLI 与 TypeScript SDK",
     "summary": "",
     "why": "官方把安全扫描做成可在编码助手里调用的 CLI 和 SDK。若你已经用 Codex 写代码，值得看它如何发现、验证再给出修复，而不是另开一套扫描器；先在测试仓库跑，避免把令牌扫进日志。",
     "tags": [
      "Codex",
      "安全",
      "CLI"
     ],
     "section": "gh",
     "published_at": "2026-11-12T07:26:30+08:00",
     "first_seen_at": "2026-11-12",
     "score": 0.718,
     "importance": 0.475,
     "sources": [
      {
       "url": "https://github.com/openai/codex-security",
       "title": "GitHub · 日榜；11.7k★ / 今日 +20"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "openai-codex-security-2026-11",
     "source": "GitHub · 日榜；11.7k★ / 今日 +20"
    },
    {
     "id": "6fcc1eb439bc8c52",
     "canonical_url": "https://github.com/shadcn/improve",
     "url": "https://github.com/shadcn/improve",
     "title": "shadcn/improve：用最强模型审计代码库，并给更便宜的模型写执行计划",
     "summary": "强模型只负责审计代码库并写出可执行的 Markdown 计划，自己并不改代码；执行交给其他更便宜的模型。可用 npx skills add shadcn/improve 装进支持 Agent Skills 的助手。",
     "why": "这是把「谁来想、谁来做」拆开的一种技能。若你已经在用贵模型写代码，可以先让它只出计划，再把实现丢给便宜模型，核对计划是否真的自洽、会不会漏测试。",
     "tags": [
      "Agent Skills",
      "审计",
      "计划"
     ],
     "section": "gh",
     "published_at": "2026-11-12T07:26:30+08:00",
     "first_seen_at": "2026-11-12",
     "score": 0.428,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://github.com/shadcn/improve",
       "title": "GitHub · 日榜；10.8k★ / 今日 +333"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "shadcn-improve-2026-11",
     "source": "GitHub · 日榜；10.8k★ / 今日 +333"
    }
   ],
   "note": "herdr 已在 10-31 报道开源，周榜再出现但无新版本信号，未重复。个人理财应用 sure 与方向无关。ha-mcp 昨日已报。"
  },
  {
   "id": "eng",
   "title": "Agent 工程",
   "intro": "同一项 GitHub 任务，MCP、CLI 和 Skill 吃掉的上下文差了一个数量级。",
   "items": [
    {
     "id": "eb34ffb6e7fa5fd2",
     "canonical_url": "https://openrouter.ai/blog/insights/mcp-vs-cli-vs-skills/",
     "url": "https://openrouter.ai/blog/insights/mcp-vs-cli-vs-skills/",
     "title": "MCP、CLI 与 Skills 各自要消耗多少上下文",
     "summary": "在 Sonnet 5.5 上，默认加载的 GitHub MCP 完成同一项查询要用 59,417 个输入 token，大约是 CLI 的 42 倍；Skill 是 3,315，CLI 是 1,421。更贵的往往是工具返回：一次 release 对象约 20,400，已经超过 46 个默认定义的 19,436。",
     "why": "若你给助手默认挂了整套 MCP，先数每轮带上多少工具，再试把结果裁到需要的字段，或改成 CLI/Skill、打开 Tool Search。同一套优化在 Haiku 5.5 上并不一定更省，有时还会更贵。",
     "tags": [
      "MCP",
      "Skills",
      "上下文",
      "成本"
     ],
     "section": "eng",
     "published_at": "2026-11-11T08:00:00+08:00",
     "first_seen_at": "2026-11-12",
     "score": 0.551,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://openrouter.ai/blog/insights/mcp-vs-cli-vs-skills/",
       "title": "OpenRouter · 11-11"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "openrouter-mcp-cli-skills-2026-11",
     "source": "OpenRouter · 11-11"
    }
   ],
   "note": "Simon Willison 对 Gemini Decisions API 的短评已作为三条相关链接，未再占坑。视频生成与具身论文未收。"
  },
  {
   "id": "infra",
   "title": "检索与基础设施",
   "intro": "多向量检索在训练时就把 token pooling 算进目标，压缩后更不容易掉点。",
   "items": [
    {
     "id": "0e5234acf2e85460",
     "canonical_url": "https://huggingface.co/blog/token-pooling",
     "url": "https://huggingface.co/blog/token-pooling",
     "title": "面向多向量嵌入模型的 Token Pooling-Aware 训练",
     "summary": "在 pool_factor=4 时，按 pooling 目标训练把 NanoBEIR nDCG@10 的保持率从 94.6% 提到 97.8%，分数从 49.03 升到 51.23，文档向量大约少存 74.8%。",
     "why": "多向量检索的体积往往先爆掉。若你已经在用这类方案，先看压缩是写进训练目标还是只在索引时合并；同样的 4 倍合并，训练时就见到 pooling，比事后再压更不容易掉点。",
     "tags": [
      "Token Pooling",
      "多向量",
      "检索"
     ],
     "section": "infra",
     "published_at": "2026-11-11T08:00:00+08:00",
     "first_seen_at": "2026-11-12",
     "score": 0.426,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://huggingface.co/blog/token-pooling",
       "title": "Hugging Face / Tom Aarsen · 11-11"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "hf-token-pooling-aware-2026-11",
     "source": "Hugging Face / Tom Aarsen · 11-11"
    }
   ]
  },
  {
   "id": "papers",
   "title": "论文",
   "intro": "一层一层做的 W4A4 量化，以及专门考决策模型的新基准。",
   "items": [
    {
     "id": "6562f4d99163dbb2",
     "canonical_url": "https://arxiv.org/abs/2611.10649",
     "url": "https://arxiv.org/abs/2611.10649",
     "title": "MACaQ：面向低精度权重与激活的快速、模型无关校准量化",
     "summary": "122B 参数的模型可以在 32.5GB 显存里用 64 分钟完成量化，因为每一层单独处理、不必整网同时留在内存里。Llama-3-8B 的困惑度已齐平 DuQuant++ 与 MR-GPTQ 的最好结果，量化时间大约只要十一分之一；在 vLLM 里跑 Qwen3-32B，相对纯 MXFP4 只多 2.3% 端到端开销。",
     "why": "若你要在消费级显卡上做 W4A4，先看这 2.3% 的在线变换开销能不能接受，再和必须把变换跨层传播的方法比速度和显存。它声明不必按架构改网络，连图像扩散模型也用同一套层内流程。",
     "tags": [
      "MACaQ",
      "量化",
      "MXFP4",
      "W4A4"
     ],
     "section": "papers",
     "published_at": "2026-11-11T00:00:00.000Z",
     "first_seen_at": "2026-11-12",
     "score": 0.515,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2611.10649",
       "title": "arXiv · 2611.10649；HF Daily Papers 11-11"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "arxiv-2611-10649-macaq-2026-11",
     "source": "arXiv · 2611.10649；HF Daily Papers 11-11"
    },
    {
     "id": "6812748aa8aae3d5",
     "canonical_url": "https://arxiv.org/abs/2611.14309",
     "url": "https://arxiv.org/abs/2611.14309",
     "title": "MMDM：大规模多任务决策评测",
     "summary": "现有考卷很难同时覆盖长上下文、大选项空间和选项之间的逻辑约束，所以作者新做了 16 类、共 5,448 道题，选项最多 1,024 个，输入可以到 64,817 个 token。28 个模型里，常识题往往不差，但选项相互依赖、带约束和多选题明显变弱——最好的硬多选正确率只有 36.81%。",
     "why": "刚发布的决策 API 需要一份不容易背题的考卷。若你在评估路由或打分模型，先看它在大选项空间和长上下文里会不会塌，而不只看常识题分数。",
     "tags": [
      "MMDM",
      "决策模型",
      "评测"
     ],
     "section": "papers",
     "published_at": "2026-11-11T00:00:00.000Z",
     "first_seen_at": "2026-11-12",
     "score": 0.445,
     "importance": 0.225,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2611.14309",
       "title": "arXiv · 2611.14309；HF Daily Papers 11-11"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "arxiv-2611-14309-mmdm-2026-11",
     "source": "arXiv · 2611.14309；HF Daily Papers 11-11"
    }
   ],
   "note": "论文日期以 HF Daily Papers 11-11 列表日为准（部分 arXiv 标注日更早）。具身操作、视频生成与金融智能体评测未收。OpenRouter 的 fp8/fp4 实测是博客不是论文，已在工程栏主题外保留为量化背景，未单列。"
  },
  {
   "id": "biz",
   "title": "大厂与模型",
   "intro": "千问下一代旗舰开了预览窗；文心给出 Flash 的首个后训练检查点；DeepMind 更新了基因组模型。",
   "items": [
    {
     "id": "290cb5a25e73fb19",
     "canonical_url": "https://help.aliyun.com/zh/model-studio/qwen4-max-preview",
     "url": "https://help.aliyun.com/zh/model-studio/qwen4-max-preview",
     "title": "Qwen4-Max-Preview",
     "summary": "预览期为北京时间 11 月 11 日 10:00 至 25 日 24:00，只走 Token Plan、不提供按量计费，上下文 1M token。思考与最终回答共享最多 393,216 的输出额度，并兼容 OpenAI 与 Anthropic 的接口；千问办公、Qoder、千问 App 与 QwenStudio 也已接入。",
     "why": "这是下一代旗舰的预览窗，不是已经定稿的正式版，官方也写了期间不做变更、到期下线。若你要测编程和智能体任务，先用套餐专用地址和模型 ID qwen4-max-preview，并把思考占用的输出额度一起留出来。",
     "tags": [
      "Qwen4",
      "Token Plan",
      "预览"
     ],
     "section": "biz",
     "published_at": "2026-11-11T10:00:00+08:00",
     "first_seen_at": "2026-11-12",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://help.aliyun.com/zh/model-studio/qwen4-max-preview",
       "title": "阿里云百炼文档 · 11-11"
      },
      {
       "title": "英文文档",
       "url": "https://www.alibabacloud.com/help/en/model-studio/qwen4-max-preview"
      },
      {
       "title": "IT之家",
       "url": "https://www.ithome.com/1/025/301.htm"
      }
     ],
     "signals": {
      "source_count": 3
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "qwen4-max-preview-2026-11",
     "source": "阿里云百炼文档 · 11-11"
    },
    {
     "id": "c6109a7889525214",
     "canonical_url": "https://ernie.baidu.com/release/ernie-5-5-flash-preview/",
     "url": "https://ernie.baidu.com/release/ernie-5-5-flash-preview/",
     "title": "ERNIE 5.5 Flash Preview",
     "summary": "基于此前的 Flash Base 做了首次后训练，总参数 280B、激活 14B，上下文 1M token。API 已在千帆上线，权重将以 Apache 2.0 开源；官方同时写明这是后训练中的中间检查点，能力还可能不完整。",
     "why": "若你在看国产开源编程模型，先按预览版来试，不要把它当成最终的 5.5 Flash。后训练还在做，正式版才是对标点。",
     "tags": [
      "ERNIE",
      "文心",
      "开源预告"
     ],
     "section": "biz",
     "published_at": "2026-11-11",
     "first_seen_at": "2026-11-12",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://ernie.baidu.com/release/ernie-5-5-flash-preview/",
       "title": "百度文心 · 11-11；官方博客"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "ernie-5-5-flash-preview-2026-11",
     "source": "百度文心 · 11-11；官方博客"
    },
    {
     "id": "811f94d4044991fa",
     "canonical_url": "https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/",
     "url": "https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/",
     "title": "AlphaGenome 2：破解细胞如何解读 DNA，以便更好理解疾病",
     "summary": "在 68 种组织、逾 1 亿个单细胞图谱上训练，识别会改变基因表达的变异数量约为上一代最强模型的两倍，并能针对特定细胞类型做条件预测。非商业研究可走 AlphaGenome API，商业用途目前是私下预览。",
     "why": "这是生物医学模型，但对「同一段输入在不同条件下读法不同」的架构值得看一眼。若你只跟踪智能体工程，把它当一条科学发布即可，不必改你的工具链。",
     "tags": [
      "AlphaGenome 2",
      "DeepMind",
      "基因组"
     ],
     "section": "biz",
     "published_at": "2026-11-11T00:00:00+00:00",
     "first_seen_at": "2026-11-12",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/",
       "title": "Google DeepMind · 11-11；官方博客"
      },
      {
       "title": "@demishassabis",
       "url": "https://x.com/demishassabis/status/2120565316619284734"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "alphagenome-2-2026-11",
     "source": "Google DeepMind · 11-11；官方博客"
    }
   ],
   "note": "Gemini Decisions API、AWS Emulator 与多用户多智能体评测见三条；BotPsy、llama.cpp FP8、Tabletop、Mastra MCP Tasks、Qwen Code 1.0 预览、Codex 0.167.2、两个热门仓库、MCP 上下文成本、Token Pooling、两篇论文，以及 Qwen4-Max 预览、文心 5.5 Flash 预览和 AlphaGenome 2 见各栏。营收传闻、融资与图像模型未收。"
  },
  {
   "id": "x",
   "title": "X 精选",
   "items": []
  }
 ]
}
