{
 "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": "2026-11-14T07:40:51+08:00",
 "published_at": "2026-11-14T07:43:59+08:00",
 "fetch_window": {
  "start": "2026-11-13T07:28:17+08:00",
  "end": "2026-11-14T07:27:04+08:00"
 },
 "date": "2026-11-14",
 "title": "AI 要闻 · 2026-11-14",
 "url": "https://www.wangwu.ai/daily/2026-11-14/",
 "fetched": "2026-11-14 07:21–07:27（北京时间）",
 "sections": [
  {
   "id": "top3",
   "title": "今日最该知道的 3 条",
   "items": [
    {
     "id": "feffbe6957a38896",
     "canonical_url": "https://techcrunch.com/2026/11/13/chatgpts-new-wallet-lets-ai-agents-shop-and-check-out-for-you/",
     "url": "https://techcrunch.com/2026/11/13/chatgpts-new-wallet-lets-ai-agents-shop-and-check-out-for-you/",
     "title": "ChatGPT 的新 Wallet 可让 AI 智能体代为购物并结账",
     "summary": "面向美国 18 岁以上的 ChatGPT Pro 用户，先把符合条件的 Visa、Mastercard 或 American Express 卡加入 Wallet，再让 Dots 代买。结账前要核对商品、总额和所用卡，并用 Face ID 或指纹确认；实际卡号不会交给智能体或商家，而是走支付代币，积分和拒付权益仍在。",
     "why": "这是把代购从建议推进到代付。目前还是小范围测试，而且要每单做生物识别；官方说以后可以按限额、品类或时间预先授权。有的零售商会拦截智能体访问，这条链路就走不通，他们也给商家准备了接入手册。",
     "tags": [
      "OpenAI",
      "ChatGPT",
      "Dots",
      "支付"
     ],
     "section": "top3",
     "published_at": "2026-11-14T00:00:00+08:00",
     "first_seen_at": "2026-11-14",
     "score": 0.678,
     "importance": 0.8,
     "sources": [
      {
       "url": "https://techcrunch.com/2026/11/13/chatgpts-new-wallet-lets-ai-agents-shop-and-check-out-for-you/",
       "title": "TechCrunch · 11-13 16:00（北京）；The Verge 稍后亦有报道。官方介绍页本次抓取返回 403，改用这两家原发报道"
      },
      {
       "title": "The Verge",
       "url": "https://www.theverge.com/ai-artificial-intelligence/1033425/openai-chatgpt-wallet"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "chatgpt-wallet-2026-11",
     "source": "TechCrunch · 11-13 16:00（北京）；The Verge 稍后亦有报道。官方介绍页本次抓取返回 403，改用这两家原发报道"
    },
    {
     "id": "7fd510d70ae47224",
     "canonical_url": "https://github.com/ggml-org/llama.cpp/releases/tag/v1.0.0",
     "url": "https://github.com/ggml-org/llama.cpp/releases/tag/v1.0.0",
     "title": "llama.cpp v1.0.0",
     "summary": "这是首个大版本，ggml 也一起到了 v1.0.0。公开 API 删掉一批已弃用接口（约 35 个函数），把 NextN/投机解码收进 llama.h；新增 OCP FP8 的 KV 缓存，以及可配置 CPU 份额的主机内存 MoE 专家缓存。服务端还补了 SSRF、CORS 与 DNS 重绑定防护。",
     "why": "本地部署如果还钉在 0.9.x，这是该对一下破坏性 API 变更的版本。日常构建还在出，但 v1.0.0 把接口和运行时能力收成了一个可依赖的基线；升级前先看发布说明里删掉的那 35 个废弃函数。",
     "tags": [
      "llama.cpp",
      "端侧运行时",
      "v1.0"
     ],
     "section": "top3",
     "published_at": "2026-11-13T10:20:19Z",
     "first_seen_at": "2026-11-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://github.com/ggml-org/llama.cpp/releases/tag/v1.0.0",
       "title": "ggml-org/llama.cpp · 11-13 18:20（北京）；官方同时发了介绍博文"
      },
      {
       "title": "介绍博文",
       "url": "https://huggingface.co/blog/ggml-org/llama-cpp-v1"
      },
      {
       "title": "@ggerganov 宣布帖",
       "url": "https://x.com/ggerganov/status/2121246103312212124"
      }
     ],
     "signals": {
      "source_count": 3
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "llamacpp-v1-0-0",
     "source": "ggml-org/llama.cpp · 11-13 18:20（北京）；官方同时发了介绍博文"
    },
    {
     "id": "2da16ca4bc5982f1",
     "canonical_url": "https://api-docs.deepseek.com/zh-cn/updates/",
     "url": "https://api-docs.deepseek.com/zh-cn/updates/",
     "title": "DeepSeek 发布全家桶应用 API",
     "summary": "把 App 里原来跟推理绑在一起的联网搜索、文件解析、阅读链接、语音和安全校验，做成可在 API 里打开的参数，工具调用和上下文拼接由服务端完成。deepseek-flash 在 Messages、Responses 和 Chat Completions 上都支持；搜索、抓取、OCR、ASR、TTS 也可单独调用。",
     "why": "以前要自己拼搜索和读网页，现在可以按官方 App 同一套能力来做对话。安全校验打开后，命中风险会返回拒答标识，对外提供服务时官方要求开启并自行承担审核义务。具体价格和用法要看文档，不要默认和聊天会员额度是同一套。",
     "tags": [
      "DeepSeek",
      "API",
      "搜索",
      "语音"
     ],
     "section": "top3",
     "published_at": "2026-11-13T22:24:44+08:00",
     "first_seen_at": "2026-11-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://api-docs.deepseek.com/zh-cn/updates/",
       "title": "DeepSeek API 更新日志 · 11-13；官方账号同步宣布"
      },
      {
       "title": "官方宣布帖",
       "url": "https://x.com/deepseek_ai/status/2121247896918253875"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "deepseek-app-api-2026-11",
     "source": "DeepSeek API 更新日志 · 11-13；官方账号同步宣布"
    }
   ]
  },
  {
   "id": "edge",
   "title": "专栏：端侧模型",
   "items": [],
   "note": "llama.cpp v1.0.0 见三条。窗内 b12336 之后多为 Hexagon、CUDA/HIP 与 Vulkan 等日常补丁，相对这个大版本没有单独列出的必要；仓库根目录再上榜亦不重复。Cohere 转发 v1.0.0 并提到 North Micro Vision 可跑 llama.cpp，信息不足以单列。"
  },
  {
   "id": "col",
   "title": "专栏：Agent 记忆与协作",
   "intro": "个人智能体协议 Poppy 出了可落地的 1.0；Google 把 Antigravity 的 /teamwork 开放给研究者。",
   "items": [
    {
     "id": "227df943fda4468c",
     "canonical_url": "https://sierra.ai/blog/releasing-poppy",
     "url": "https://sierra.ai/blog/releasing-poppy",
     "title": "后续｜发布个人智能体协议 1.0",
     "summary": "协议 1.0 可按现成的 MCP、OpenAPI 和 UCP 分步接入，认证走 OAuth。Shopify 全站商家会自动发出 Poppy 文件，Muse 访问站点时会先找它。Mercury、Synchrony、土耳其航空已上线；新加入 60 家设计伙伴，总数超过 100 家。",
     "why": "10 月公布的还是草案，这次是可落地的 1.0，而且 Muse 和 Shopify 已经接上。若你在做要登录、查单或代下单的个人智能体，可以先按官方实现包和校验器对一下自己的站点；支付和转账仍要用户确认，并走商家自己的审批。",
     "tags": [
      "Poppy",
      "协议",
      "Muse",
      "Shopify"
     ],
     "section": "col",
     "published_at": "2026-11-14T04:20:19+08:00",
     "first_seen_at": "2026-11-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://sierra.ai/blog/releasing-poppy",
       "title": "Sierra · 官方博文；10-11 已报过草案"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "poppy-1-0-2026-11",
     "source": "Sierra · 官方博文；10-11 已报过草案"
    },
    {
     "id": "0e40d692e8fc9ab5",
     "canonical_url": "https://research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature/",
     "url": "https://research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature/",
     "title": "用 Antigravity 的 /teamwork 推进科学与工程",
     "summary": "Google Research 向外部研究者开放了 Antigravity 中的 /teamwork 多智能体流程。它把并行探索和 Lean 4、测试套件一类的专门证伪器配在一起，在数学、化学、系统软件和机器学习工程的未决问题上与人类专家一起做；过程中会用到 Gemini 4 的 Argon、Carbon 和 Hydrogen。",
     "why": "这不是新模型发布，而是把内部用来做长程研究的多智能体流程对外开放。适合已经能把问题写成可验证约束的团队；没有自动判对错的环境，效果会差一截。申请走研究者准入，并不意味对所有 Antigravity 用户默认开放。",
     "tags": [
      "Antigravity",
      "多智能体",
      "Gemini 4"
     ],
     "section": "col",
     "published_at": "2026-11-13",
     "first_seen_at": "2026-11-14",
     "score": 0.656,
     "importance": 0.475,
     "sources": [
      {
       "url": "https://research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature/",
       "title": "Google Research · 11-13；Vahab Mirrokni"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "antigravity-teamwork-2026-11",
     "source": "Google Research · 11-13；Vahab Mirrokni"
    }
   ],
   "note": "Muse 向美国付费用户开放代打电话是产品功能更新，与 Poppy 1.0 相比增量有限，未另占坑。仅在日榜出现的交易类技能仓库未收。"
  },
  {
   "id": "cl",
   "title": "主流产品 Changelog 精读",
   "intro": "Codex 正式版加了侧栏会话；OpenCode 让子智能体默认在后台跑；Claude Code 修了一处安全问题。",
   "items": [
    {
     "id": "ef6f4b600bd3cc25",
     "canonical_url": "https://github.com/openai/codex/releases/tag/rust-v0.168.0",
     "url": "https://github.com/openai/codex/releases/tag/rust-v0.168.0",
     "title": "openai/codex rust-v0.168.0",
     "summary": "会话里可以浏览并选择默认模型供应商；打开 features.tui_workspace 后，主对话旁边能开一条侧栏会话，拖分割线或用快捷键放大。支持给能听音频的模型附加录音；终端通知可改用 osc777，批准请求会区分终端输入和联网。",
     "why": "这是 0.167 之后的正式版，侧栏会话和供应商切换是最直接能感到的变化。若还在用 features.multi_agent 开关，升级后会看到改用 agents.enabled 的提示。",
     "tags": [
      "Codex",
      "TUI",
      "模型供应商"
     ],
     "section": "cl",
     "published_at": "2026-11-13T03:54:29Z",
     "first_seen_at": "2026-11-14",
     "score": 0.683,
     "importance": 0.475,
     "sources": [
      {
       "url": "https://github.com/openai/codex/releases/tag/rust-v0.168.0",
       "title": "openai/codex · 11-13 11:54（北京）"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "codex-rust-v0-168-0",
     "source": "openai/codex · 11-13 11:54（北京）"
    },
    {
     "id": "3d316f1b8cf467f6",
     "canonical_url": "https://github.com/anomalyco/opencode/releases/tag/v2.3.0",
     "url": "https://github.com/anomalyco/opencode/releases/tag/v2.3.0",
     "title": "OpenCode v2.3.0",
     "summary": "升级后最明显的是默认行为：子智能体先交回会话 ID、在后台跑，父会话不用干等。完整逐步记录不再默认展开，需把 Verbosity 调到 medium 或高，已有设置会保留。此版本还加了实验性 MCP events 开关；桌面端可用命令面板跳回更早的提示。",
     "why": "这是 2.2 之后一次默认行为的变化：以前会在前台等子智能体，现在会先回到父会话。若你依赖完整逐步记录，升级后先检查详细程度设置，避免只看到折叠后的一行摘要。",
     "tags": [
      "OpenCode",
      "子智能体",
      "TUI"
     ],
     "section": "cl",
     "published_at": "2026-11-13T08:55:07Z",
     "first_seen_at": "2026-11-14",
     "score": 0.361,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://github.com/anomalyco/opencode/releases/tag/v2.3.0",
       "title": "anomalyco/opencode · 11-13 16:55（北京）"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "opencode-v2-3-0",
     "source": "anomalyco/opencode · 11-13 16:55（北京）"
    },
    {
     "id": "bdb92b385d1eb928",
     "canonical_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.325",
     "url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.325",
     "title": "Claude Code v2.1.325",
     "summary": "zstd 压缩回包的解码路径有一处安全缺陷，这次已经补上。被组织策略拦住的请求，错误信息下会多一行说明：用 /rewind 改提示，或执行 /clear。不少第三方网关和 Foundry 上，原先每轮都卡在上下文超长、走不了压缩，现在会正常压缩。",
     "why": "没有新的大功能，但安全修复和网关压缩问题值得尽快跟上。若你在用 Opus/Haiku 的 plan 模式或给子智能体单独设模型，这次也修了被主会话模型带跑的问题。",
     "tags": [
      "Claude Code",
      "安全",
      "修复"
     ],
     "section": "cl",
     "published_at": "2026-11-13T20:40:01Z",
     "first_seen_at": "2026-11-14",
     "score": 0.455,
     "importance": 0.3,
     "sources": [
      {
       "url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.325",
       "title": "anthropics/claude-code · 11-14 04:40（北京）"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "claude-code-v2-1-325",
     "source": "anthropics/claude-code · 11-14 04:40（北京）"
    }
   ],
   "note": "openai/codex 的 0.171.0-alpha 系列发布说明几乎为空，未收。qwen-code v0.27.3 及桌面/SDK 同步版改动分散，未单列。claude-agent-sdk-python v0.2.182 仅为常规补丁。"
  },
  {
   "id": "gh",
   "title": "GitHub 热门",
   "intro": "Google 开源了一套先写规格、再换一双眼睛评审的智能体插件。",
   "items": [
    {
     "id": "9c3337bc2e4f5f81",
     "canonical_url": "https://github.com/google/elephant-goldfish-model",
     "url": "https://github.com/google/elephant-goldfish-model",
     "title": "google/elephant-goldfish-model：用智能体设计并评审软件的开源插件",
     "summary": "这套开源插件把起草和评审拆开：审的一侧看不到前面的对话，只根据草稿和引用材料挑未写明的假设，架构取舍仍由人决定。起草的一侧（Elephant）守着项目上下文；评审一侧叫 Goldfish。已有 Claude Code、Codex、Antigravity、Cursor 和 Windsurf 的安装包。",
     "why": "仓库本周周榜约加了 5,555 星，是第一次进我们的 GitHub 栏。适合规格先写清楚再开工的团队；它不会替你做架构取舍，评审循环也要自己看输出是否真指出了盲点。",
     "tags": [
      "规格",
      "代码评审",
      "插件"
     ],
     "section": "gh",
     "published_at": "2026-11-14T07:27:04+08:00",
     "first_seen_at": "2026-11-14",
     "score": 0.566,
     "importance": 0.475,
     "sources": [
      {
       "url": "https://github.com/google/elephant-goldfish-model",
       "title": "GitHub · 周榜；18.6k★ / 本周 +5,555"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "elephant-goldfish-model-2026-11",
     "source": "GitHub · 周榜；18.6k★ / 本周 +5,555"
    }
   ],
   "note": "lobehub/lobehub、tailscale/tailcat 与视频相关仓库或与方向不符，或只是周榜涨星、未见新版本信号，未收。chrome-devtools-mcp 同样缺少可核的新进展，未单列。"
  },
  {
   "id": "eng",
   "title": "Agent 工程",
   "intro": "一份基准专门测助手会不会纠正问错的问题；另一份测模型会不会在没人叫时主动帮忙。",
   "items": [
    {
     "id": "ae6d85d1de45fe2b",
     "canonical_url": "https://allenai.org/blog/xybench",
     "url": "https://allenai.org/blog/xybench",
     "title": "XYBench：帮助解决提问背后的问题",
     "summary": "这是一份 8,115 道题的基准，用来测助手面对 XY 问题（人想做 X，却问成了做法 Y）时会不会纠正前提。题来自真实的编程和日常提问，并获得 EMNLP 2026 最佳资源论文。评过的 Claude Sonnet 4.6、GPT-5、Gemini 3.1 Pro 等模型，常常仍绕着原来那个错的做法给建议。",
     "why": "如果你在做会接到含糊或问错问题的助手，这份数据比普通问答基准更贴近真实误问。论文、数据和代码都已公开；榜上的模型代次与今天线上最新版不是一一对应，数字只说明这类纠偏仍然普遍偏弱。",
     "tags": [
      "基准",
      "XY 问题",
      "EMNLP"
     ],
     "section": "eng",
     "published_at": "2026-11-13",
     "first_seen_at": "2026-11-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://allenai.org/blog/xybench",
       "title": "Ai2 · 11-13；与卡内基梅隆、NVIDIA、约翰霍普金斯合作"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "xybench-2026-11",
     "source": "Ai2 · 11-13；与卡内基梅隆、NVIDIA、约翰霍普金斯合作"
    },
    {
     "id": "655eec489787f50f",
     "canonical_url": "https://deepnote.com/research/benchmark",
     "url": "https://deepnote.com/research/benchmark",
     "title": "Agent Initiative Benchmark",
     "summary": "他们搭了一个叫 PageMantle 的模拟公司（含 CRM、分析、账单、仓库、工单、文档和 Slack），用 50 个职场情景测 12 个前沿模型会不会在没人叫时主动帮忙。GPT-6 Astra 综合 8.97/10 居首，随后是 GPT-6.1 Sol 和 Claude Fable 5.2。评测环境已开源。",
     "why": "测的是会不会在该出手时出手、证据从哪来、结果交给谁，不是常规问答。样本只有 50 题，而且是合成公司，不能直接当成你自己业务里的排名；要选型的话，更有用的是他们公开的场景和工具环境，可以拿自己的数据复测。",
     "tags": [
      "基准",
      "主动智能体",
      "评测"
     ],
     "section": "eng",
     "published_at": "2026-11-13T22:58:55+08:00",
     "first_seen_at": "2026-11-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://deepnote.com/research/benchmark",
       "title": "Deepnote Research · 11-13"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "deepnote-agent-initiative-bench-2026-11",
     "source": "Deepnote Research · 11-13"
    }
   ],
   "note": "视频生成与驾驶世界模型论文未收。Reason Machines 的编程辅助框架目前主要是推广帖，未见可核的官方技术说明，未单列。"
  },
  {
   "id": "infra",
   "title": "检索与基础设施",
   "intro": "OpenRouter 用合成题把重排序的收益测清楚了；DeepMind 公开了天气模型背后的观测数据。",
   "items": [
    {
     "id": "4e711dbb1f520bf2",
     "canonical_url": "https://openrouter.ai/blog/tutorials/reranking-rag/",
     "url": "https://openrouter.ai/blog/tutorials/reranking-rag/",
     "title": "RAG 中的重排序：用重排序模型修正糟糕的检索",
     "summary": "先拉大约 30 条再重排、只把前 3 条送给生成模型。底检索用 google/gemini-embedding-2、重排用 voyageai/rerank-3-lite 时，274 道合成题上正确答案排第一的比例从 81% 升到 97%，大约每千次 0.10 美元。GPT-6 Luna 能到 99.6% 但更慢；作者强调这是干净合成语料。",
     "why": "这是一篇带完整评测和示例代码的教程，不是新模型发布。若检索经常取到相近但答错套餐或版本的片段，可以按他们的办法先拉 30 条再重排、只把前 3 条送给生成模型；也要注意重排模型的上下文长度，超长文档可能被静默截断。",
     "tags": [
      "RAG",
      "重排序",
      "评测"
     ],
     "section": "infra",
     "published_at": "2026-11-13T08:00:00+08:00",
     "first_seen_at": "2026-11-14",
     "score": 0.426,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://openrouter.ai/blog/tutorials/reranking-rag/",
       "title": "OpenRouter · 11-13"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "openrouter-rerank-rag-2026-11",
     "source": "OpenRouter · 11-13"
    },
    {
     "id": "b2cf4551f09dc08d",
     "canonical_url": "https://deepmind.google/blog/opening-the-datasets-behind-our-ai-weather-models/",
     "url": "https://deepmind.google/blog/opening-the-datasets-behind-our-ai-weather-models/",
     "title": "公开我们天气 AI 模型背后的数据集",
     "summary": "DeepMind 公开了 WeatherNext 3 用过的两套基础数据：SatWeave 是每小时更新的全球云图（17 个通道，含十年以上历史）；PARDIG 是约 10 公里、15 分钟分辨率的全球降水再分析，中小雨检测技能比 IMERG-Final 高 30% 以上。历史库和实时馈源放在 Google Cloud Storage。",
     "why": "天气模型 WeatherNext 3 本身是 9 月已经发布的，今天公开的是它用的观测和降水数据。做气象、水文或能源预测的人可以直接接实时馈源；这不是面向聊天应用的模型接口。",
     "tags": [
      "WeatherNext",
      "数据集",
      "气象"
     ],
     "section": "infra",
     "published_at": "2026-11-13",
     "first_seen_at": "2026-11-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://deepmind.google/blog/opening-the-datasets-behind-our-ai-weather-models/",
       "title": "Google DeepMind · 11-13"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "weathernext-pardig-satweave-2026-11",
     "source": "Google DeepMind · 11-13"
    }
   ]
  },
  {
   "id": "papers",
   "title": "论文",
   "intro": "决策模型也许当不好执行者，但仍可能当好老师。",
   "items": [
    {
     "id": "f9dab7ef6617d718",
     "canonical_url": "https://arxiv.org/abs/2611.17167",
     "url": "https://arxiv.org/abs/2611.17167",
     "title": "Actor、Judge 还是 Teacher？理解长时程智能体中的决策模型",
     "summary": "核心结论是：动手容易翻车的决策模型，事后仍可能当好裁判或老师。用 JEV 当老师，Qwen2.5-1.5B 在 GRPO 上的成功率从 ALFWorld 的 72.8% 提到 82.9%，WebShop 从 56.8% 提到 71.4%。执行时它们会被更简单的子目标带走，但看完整轨迹后能给必做的那一步记功。",
     "why": "结论很具体：执行时看不清后果的模型，事后当裁判或老师仍可能有用。代码已公开；这是在两个代理环境上的结果，换到你自己的任务前，先确认有没有同样能事后检验对错的轨迹。",
     "tags": [
      "决策模型",
      "智能体",
      "强化学习"
     ],
     "section": "papers",
     "published_at": "2026-11-13",
     "first_seen_at": "2026-11-14",
     "score": 0.448,
     "importance": 0.225,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2611.17167",
       "title": "arXiv 2611.17167 · HF Daily Papers 11-13 列表"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "actor-judge-teacher-2611-17167",
     "source": "arXiv 2611.17167 · HF Daily Papers 11-13 列表"
    }
   ],
   "note": "论文日期以 HF Daily Papers 11-13 列表日为准（这篇 arXiv 标注提交于 11-12）。同日的视频生成与驾驶世界模型论文未收。"
  },
  {
   "id": "biz",
   "title": "大厂与模型",
   "intro": "OpenAI 把网络安全做成两款智能体产品；腾讯混元的实时语音模型加了工具调用。",
   "items": [
    {
     "id": "8da4978302c57e7b",
     "canonical_url": "https://the-decoder.com/openai-enters-the-cybersecurity-market-with-two-daybreak-products/",
     "url": "https://the-decoder.com/openai-enters-the-cybersecurity-market-with-two-daybreak-products/",
     "title": "OpenAI 凭两款 Daybreak 产品进入网络安全市场",
     "summary": "目前还在早期试用：两款产品都跑在 GPT-6 Cyber 上。Defense Factory 扫描代码库并生成带测试的修复拉取请求，由开发者确认；Red Team 结合源码与上下文探网站漏洞，也可复核赏金或扫描器的报告、验证补丁。",
     "why": "这是 OpenAI 把网络安全做成可购买的智能体产品，还在早先试用和测试阶段。对有代码仓库、又需要持续渗透或补丁建议的团队有参考价值；实际检出率和误报还没有独立评测，不能只看官方说法。",
     "tags": [
      "OpenAI",
      "网络安全",
      "Daybreak"
     ],
     "section": "biz",
     "published_at": "2026-11-13T23:08:57+08:00",
     "first_seen_at": "2026-11-14",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://the-decoder.com/openai-enters-the-cybersecurity-market-with-two-daybreak-products/",
       "title": "THE DECODER · 11-13；官方产品页本次未作为正文源抓取"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "openai-daybreak-2026-11",
     "source": "THE DECODER · 11-13；官方产品页本次未作为正文源抓取"
    },
    {
     "id": "c11f1b5561d3284d",
     "canonical_url": "https://mp.weixin.qq.com/s/bFxAT04ohJs9hPcjjin3bg",
     "url": "https://mp.weixin.qq.com/s/bFxAT04ohJs9hPcjjin3bg",
     "title": "腾讯混元：发布全新实时语音模型 Hy Realtime",
     "summary": "",
     "why": "官方公众号原文这次没能抓到可读正文，只从第一财经转述核到：新模型在实时语音之外加了工具调用和唱歌，已进元宝语音通话和腾讯客服，并计划很快在腾讯云提供 API。是否真能在对话中稳定调工具，要等 API 文档和额度说明出来再看。",
     "tags": [
      "腾讯混元",
      "语音",
      "工具调用"
     ],
     "section": "biz",
     "published_at": "2026-11-13T16:32:00+08:00",
     "first_seen_at": "2026-11-14",
     "score": 0.537,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://mp.weixin.qq.com/s/bFxAT04ohJs9hPcjjin3bg",
       "title": "腾讯混元公众号 · 11-13 16:32；官方页本次遇到验证码，事实来自第一财经转述（东方财富转载）"
      },
      {
       "title": "第一财经（东方财富转载）",
       "url": "https://finance.eastmoney.com/a/202611133930722129.html"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "hunyuan-hy-realtime-2026-11",
     "source": "腾讯混元公众号 · 11-13 16:32；官方页本次遇到验证码，事实来自第一财经转述（东方财富转载）"
    }
   ],
   "note": "ChatGPT Wallet、llama.cpp v1.0.0 与 DeepSeek 全家桶 API 见三条；Poppy 1.0、Antigravity /teamwork、Codex 0.168.0、OpenCode 2.3.0、Claude Code 安全修复、Elephant-Goldfish 插件、XYBench、Agent Initiative Benchmark、OpenRouter 重排序实测、DeepMind 天气数据集，以及 Daybreak 与混元 Hy Realtime 见各栏。融资、图像与视频生成、已报的 Gemini 4 Hydrogen / Claude 自托管 Managed Agents / Inkling 权重未再占坑。"
  },
  {
   "id": "x",
   "title": "X 精选",
   "items": []
  }
 ]
}
