{
 "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": "2026-10-28T07:24:00+08:00",
 "published_at": "2026-10-28T07:50:41+08:00",
 "fetch_window": {
  "start": "2026-10-27T07:18:00+08:00",
  "end": "2026-10-28T07:24:37+08:00"
 },
 "date": "2026-10-28",
 "title": "AI 要闻 · 2026-10-28",
 "url": "https://www.wangwu.ai/daily/2026-10-28/",
 "fetched": "2026-10-28 07:18–07:24（北京时间）",
 "sections": [
  {
   "id": "top3",
   "title": "今日最该知道的 3 条",
   "items": [
    {
     "id": "ff9622ee3ab36363",
     "canonical_url": "https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/",
     "url": "https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/",
     "title": "Genie 4：新一代世界模拟",
     "summary": "实时交互世界模型首次原生同出画面、环境音与对话；Gemini Omni 锚定外观与行为，其他 Gemini 维持连贯，Gemini 3.8 Flash 在环拆目标并核对步骤。可信测试者先用，随后面向 Google AI Ultra 的 Project Genie 与 Gemini API / AI Studio 早期伙伴。",
     "why": "世界模型开始把「可玩的模拟」和「可调用的推理环」绑在一起。做机器人/agent 训练场的，先盯实时改提示后因果是否还稳，再谈能不能替代真机数据。",
     "tags": [
      "世界模型",
      "实时",
      "Gemini"
     ],
     "section": "top3",
     "published_at": "2026-10-28T00:01:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.588,
     "importance": 0.9,
     "sources": [
      {
       "url": "https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/",
       "title": "Google DeepMind · 10-28 00:01"
      },
      {
       "title": "_philschmid",
       "url": "https://x.com/_philschmid/status/2115129878697697566"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "deepmind-genie-4-2026-10",
     "source": "Google DeepMind · 10-28 00:01"
    },
    {
     "id": "bedcbb89bf791dd5",
     "canonical_url": "https://arxiv.org/abs/2610.33024",
     "url": "https://arxiv.org/abs/2610.33024",
     "title": "能力对齐进化：通过 Skill 适配改进 Harness 自进化",
     "summary": "103 个 task–skill 对上，同一套 skill 让 GPT-5.6-luna 升约 31.1 分，却让 Gemma-4-E4B 掉约 8.8 分；弱模型失败案例里约 86% 已违反 skill 里写明的要求。用固定元 skill 指导改写后，Kimi-K3 无执行证据即可把三个弱执行器抬高约 5.1–14.1 分。",
     "why": "Skill 不是越大越强模型通用补丁。写 harness / skill 库时，按执行器能力分档改写，比只堆更长说明书更关键。",
     "tags": [
      "Harness",
      "Skills",
      "自进化"
     ],
     "section": "top3",
     "published_at": "2026-10-27T08:00:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.631,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.33024",
       "title": "arXiv · HF Daily 10-27"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "skill-adaptation-harness-2610-33024",
     "source": "arXiv · HF Daily 10-27"
    },
    {
     "id": "c61d52b1968d74e5",
     "canonical_url": "https://github.com/openai/codex/releases/tag/rust-v0.165.0",
     "url": "https://github.com/openai/codex/releases/tag/rust-v0.165.0",
     "title": "rust-v0.165.0",
     "summary": "内置 Google Cloud Vertex 提供方（model_provider=\"gcp\"），模型目录加入 GPT-6.1 Terra；新配置默认 memory v2，并可导入外部 agent 记忆。对话中改 reasoning effort 默认开启，会话选择器加仓库过滤、命令中心按目录分组。",
     "why": "个人/团队记忆开始有「从别的助手搬进来」的一等入口。迁移前先核对导入字段与 consolidation 竞态说明，别假定旧记忆文件直接兼容。",
     "tags": [
      "Codex",
      "memory",
      "Vertex"
     ],
     "section": "top3",
     "published_at": "2026-10-28T04:09:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.687,
     "importance": 0.8,
     "sources": [
      {
       "url": "https://github.com/openai/codex/releases/tag/rust-v0.165.0",
       "title": "OpenAI Codex · 10-28 04:09"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "codex-rust-v0-165-0",
     "source": "OpenAI Codex · 10-28 04:09"
    }
   ]
  },
  {
   "id": "edge",
   "title": "专栏：端侧模型",
   "intro": "LFM3 开源权重家族、离线编码实测，以及 llama.cpp 接入 OCR 校验头。",
   "items": [
    {
     "id": "3722169d83e12a53",
     "canonical_url": "https://www.liquid.ai/blog/lfm3",
     "url": "https://www.liquid.ai/blog/lfm3",
     "title": "介绍 LFM3：面向专用智能的全模态",
     "summary": "约 2B 档六个开源权重模型共用 LFM3-1.7B-Base（约 22T token 预训练），覆盖文本/视觉/音频/omni 与两个决策模型；骨干为卷积 + 滑动窗注意力 + 全局注意力的 tri-hybrid，上下文提到 256K。",
     "why": "小模型开始按「专用智能」拆模态发齐。本地 agent 选型时，先看你要的是工具调用、视觉还是决策头，再比参数量。",
     "tags": [
      "端侧",
      "多模态",
      "开源权重"
     ],
     "section": "edge",
     "published_at": "2026-10-27T12:00:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://www.liquid.ai/blog/lfm3",
       "title": "Liquid AI · 10-27"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "liquid-lfm3-2026-10",
     "source": "Liquid AI · 10-27"
    },
    {
     "id": "f93b3a047db2176d",
     "canonical_url": "https://latent.space/p/offline-coding/",
     "url": "https://latent.space/p/offline-coding/",
     "title": "在喜马拉雅用离线 AI Agent 写代码",
     "summary": "Thomson Reuters 工程师在断网环境下用 OpenCode + Ollama 跑 Qwen3.5 9B：先联网装好依赖与模型，再关 Wi-Fi，靠权限提示审 diff 并用项目自带 unittest 验收；示例把费用合计从「每次覆盖」改成累加后五测全过。",
     "why": "离线 coding agent 的前提是「模型+依赖+测试」事先落盘。出差/弱网场景先演练一次纯本地回归，别假定还能 pip install。",
     "tags": [
      "离线",
      "OpenCode",
      "Ollama"
     ],
     "section": "edge",
     "published_at": "2026-10-27T22:41:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.426,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://latent.space/p/offline-coding/",
       "title": "Latent Space · 10-27"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "latent-offline-coding-2026-10",
     "source": "Latent Space · 10-27"
    },
    {
     "id": "ec67972f4de71dce",
     "canonical_url": "https://github.com/ggml-org/llama.cpp/releases/tag/b11928",
     "url": "https://github.com/ggml-org/llama.cpp/releases/tag/b11928",
     "title": "b11928",
     "summary": "接入 MiniCPM-OCR-Jev：Qwen3.5-VL 躯干加 value head，在图像与 OCR 文本上打分、不生成正文；以 openbmb-minicpm-ocr-jev 决策类型经 POST /v1/systemone 提供，无 LM head 时拒绝文本生成。",
     "why": "本地文档流水线可把「解析」和「该不该人工复核」拆开。升级到含此构建后再接 Jev，别混用会乱生成的旧路径。",
     "tags": [
      "llama.cpp",
      "OCR",
      "Jev"
     ],
     "section": "edge",
     "published_at": "2026-10-27T20:30:28+08:00",
     "first_seen_at": "2026-10-28",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://github.com/ggml-org/llama.cpp/releases/tag/b11928",
       "title": "GitHub · 10-27 20:30"
      },
      {
       "title": "MiniCPM-OCR",
       "url": "https://github.com/OpenBMB/MiniCPM-OCR"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "key": "llamacpp-b11928-minicpm-ocr-jev",
     "source": "GitHub · 10-27 20:30"
    }
   ],
   "note": "窗内 b11913–b11925 多为日常/CI；有实质的是 b11928 接入 MiniCPM-OCR-Jev。Sentient OS 仅短帖、未核到可复现仓库，未单列。"
  },
  {
   "id": "col",
   "title": "专栏：Agent 记忆与协作",
   "intro": "领域 Skills 包、登录态复用抓取，以及多 agent 编排学习。",
   "items": [
    {
     "id": "523739287734d53a",
     "canonical_url": "https://www.databricks.com/blog/vital-skills-upskilling-genie-code-healthcare-and-life-sciences-skills",
     "url": "https://www.databricks.com/blog/vital-skills-upskilling-genie-code-healthcare-and-life-sciences-skills",
     "title": "Vital Skills：用医疗与生命科学技能给 Genie Code 补课",
     "summary": "开源 Agent Skills，把生物信息/队列构建/脱敏等流程写成可问阈值的 playbook；BiomniBench 在技能覆盖的 5 项上 Genie Code 均分约从 71.9 提到 76.9，18 项域外任务大致持平。配对里自由文本队列 F1 约从 0.62 到 1.00；TEDDY 部署检查从 19/27 到 27/27。",
     "why": "领域 skill 的价值是「先问阈值、再用受测工具」，不是又一份提示词长文。自建行业 skill 时可抄：配对评测 + 域外不伤分。",
     "tags": [
      "Skills",
      "医疗",
      "Genie Code"
     ],
     "section": "col",
     "published_at": "2026-10-28T03:00:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.665,
     "importance": 0.5,
     "sources": [
      {
       "url": "https://www.databricks.com/blog/vital-skills-upskilling-genie-code-healthcare-and-life-sciences-skills",
       "title": "Databricks · 10-28 03:00"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [
      {
       "name": "AIHOT",
       "url": "https://aihot.news/"
      }
     ],
     "key": "databricks-vital-skills-2026-10",
     "source": "Databricks · 10-28 03:00"
    },
    {
     "id": "9107aca24794853f",
     "canonical_url": "https://www.browserbase.com/blog/introducing-authenticated-fetch",
     "url": "https://www.browserbase.com/blog/introducing-authenticated-fetch",
     "title": "介绍 Authenticated Fetch",
     "summary": "浏览器登录并 persist Context 后，Fetch API 可带同一 contextId 复用 cookie 拉需登录页，响应可选 markdown / 原始 / 按 schema 抽 JSON；Fetch 不回写 cookie，过期需再开一次 persist 会话。对 2026-09-01 前保存的 Context 也要求先刷新一次。",
     "why": "Agent 读后台页时，把「登录会话」和「批量拉取」拆开可少开可交互浏览器。落地时先划清 cookie 域与过期续期，别把 Context 当永久钥匙。",
     "tags": [
      "浏览器",
      "登录态",
      "Fetch"
     ],
     "section": "col",
     "published_at": "2026-10-27T12:00:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.582,
     "importance": 0.5,
     "sources": [
      {
       "url": "https://www.browserbase.com/blog/introducing-authenticated-fetch",
       "title": "Browserbase · 窗内"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "browserbase-authenticated-fetch-2026-10",
     "source": "Browserbase · 窗内"
    },
    {
     "id": "1c2b941bf36731ad",
     "canonical_url": "https://arxiv.org/abs/2610.34467",
     "url": "https://arxiv.org/abs/2610.34467",
     "title": "DynaFlow：学习自由形式的多 Agent 编排",
     "summary": "主 agent 在动作空间里合成可执行工作流（串/并、信息交换、校验），只训练主 agent，子 agent 固定。Forced 模式下六项基准总准确率约 88.08%，同骨干 Qwen3.5-27B 上比最强对照高约 10.90 个百分点；分析称训练后更明确地使用、校验与 refinement 上游输出。",
     "why": "多 agent 别先画死拓扑。评编排时看「跨阶段信息有没有被显式接力」，而不只看最终答对率。",
     "tags": [
      "多 Agent",
      "编排",
      "工作流"
     ],
     "section": "col",
     "published_at": "2026-10-27T08:00:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.598,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.34467",
       "title": "arXiv · HF Daily 10-27"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "dynaflow-2610-34467",
     "source": "arXiv · HF Daily 10-27"
    }
   ],
   "note": "cloudflare/security-audit-skill 已于 10-22 报过，不重复。Codex memory v2 / 外部记忆导入见今日三条③。Grok Bot「记忆导入」仅短帖，见私密。"
  },
  {
   "id": "cl",
   "title": "主流产品 Changelog 精读",
   "intro": "窗口内只列值得注意的条目，安静的产品略过。",
   "items": [
    {
     "id": "5bc68684f9981cae",
     "canonical_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.312",
     "url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.312",
     "title": "Claude Code：策略刷新与大文件误写修复",
     "summary": "",
     "why": "网关可配 oidc.user_id_claim（如 Entra oid）；修 VS Code 登录后管策略不刷新、大 CLAUDE.md 未读全就被覆盖、Bash cd 后丢 per-directory skill、以及 PDF Read 与清理竞态。云会话/Claude Tag/Code Review 另有一批显示与重发修复。",
     "tags": [
      "Claude Code"
     ],
     "section": "cl",
     "published_at": null,
     "first_seen_at": "2026-10-28",
     "score": 0.455,
     "importance": 0.3,
     "sources": [
      {
       "url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.312",
       "title": null
      },
      {
       "title": "Claude Code changelog",
       "url": "https://github.com/anthropics/claude-code/releases"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "product": "Claude Code"
    },
    {
     "id": "fc67d2cd5b4245de",
     "canonical_url": "https://github.com/anomalyco/opencode/releases/tag/v2.2.0",
     "url": "https://github.com/anomalyco/opencode/releases/tag/v2.2.0",
     "title": "OpenCode：会话指令条目与桌面插件",
     "summary": "",
     "why": "插件可用 ctx.session.instructions.entry.put/list/remove 挂键控会话指令：不改系统提示、不立刻多打一轮，且能熬过 compaction/revert。另加桌面插件实验、按提供方可配 prompt cache TTL。",
     "tags": [
      "OpenCode"
     ],
     "section": "cl",
     "published_at": null,
     "first_seen_at": "2026-10-28",
     "score": 0.361,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://github.com/anomalyco/opencode/releases/tag/v2.2.0",
       "title": null
      },
      {
       "title": "OpenCode changelog",
       "url": "https://github.com/anomalyco/opencode/releases"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "product": "OpenCode"
    },
    {
     "id": "30701ebf04dccea1",
     "canonical_url": "https://github.com/google-gemini/gemini-cli/releases/tag/v0.66.0",
     "url": "https://github.com/google-gemini/gemini-cli/releases/tag/v0.66.0",
     "title": "Gemini CLI：tool 分组与 SKILL 解析",
     "summary": "",
     "why": "调度侧直接吐 tool call groups；加固 SKILL.md frontmatter 单行描述解析；MCP OAuth 可选择加入私网端点；另修滚动/spinner、Windows diff 参数与 PATH 误报更新成功等。",
     "tags": [
      "Gemini CLI"
     ],
     "section": "cl",
     "published_at": null,
     "first_seen_at": "2026-10-28",
     "score": 0.466,
     "importance": 0.475,
     "sources": [
      {
       "url": "https://github.com/google-gemini/gemini-cli/releases/tag/v0.66.0",
       "title": null
      },
      {
       "title": "Gemini CLI changelog",
       "url": "https://github.com/google-gemini/gemini-cli/releases"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "product": "Gemini CLI"
    }
   ],
   "note": "Codex rust-v0.165.0 见今日三条，不在本栏重复。Qwen Code v0.26.2 以 managed-agent breaking 为主，未单列。gemini-cli nightly/preview 略。窗内 alpha 空说明未收。"
  },
  {
   "id": "gh",
   "title": "GitHub 热门",
   "intro": "日榜里设计 skill、终身辅导 agent 与 agentic 社媒调度。",
   "items": [
    {
     "id": "c1552a7f3d03d320",
     "canonical_url": "https://github.com/elayadesign/ai-design-skills",
     "url": "https://github.com/elayadesign/ai-design-skills",
     "title": "elayadesign/ai-design-skills",
     "summary": "",
     "why": "面向编码 agent 的设计向 skill 集，日榜热度高。装之前先扫会不会改仓库外文件，并和昨日华数设计 skill 对照：要的是风格约束还是一整条交付流水线。",
     "tags": [
      "设计",
      "skill",
      "Agent"
     ],
     "section": "gh",
     "published_at": null,
     "first_seen_at": "2026-10-28",
     "score": 0.695,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://github.com/elayadesign/ai-design-skills",
       "title": "GitHub Trending 日榜"
      }
     ],
     "signals": {
      "source_count": 1,
      "stars_total": 9246
     },
     "via": [],
     "key": "elaya-ai-design-skills-2026-10",
     "source": "GitHub Trending 日榜"
    },
    {
     "id": "b0c9318e6601e1fa",
     "canonical_url": "https://github.com/HKUDS/DeepTutor",
     "url": "https://github.com/HKUDS/DeepTutor",
     "title": "HKUDS/DeepTutor：终身个性化辅导",
     "summary": "",
     "why": "打出 lifelong personalized tutoring。若你做带长期学员状态的 agent，可看它如何存进度与偏好，而不是只看题库正确率。",
     "tags": [
      "教育",
      "记忆",
      "Agent"
     ],
     "section": "gh",
     "published_at": null,
     "first_seen_at": "2026-10-28",
     "score": 0.447,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://github.com/HKUDS/DeepTutor",
       "title": "GitHub Trending 日榜"
      }
     ],
     "signals": {
      "source_count": 1,
      "stars_total": 45800
     },
     "via": [],
     "key": "hkuds-deeptutor-2026-10",
     "source": "GitHub Trending 日榜"
    },
    {
     "id": "393a17b3c1f403f8",
     "canonical_url": "https://github.com/gitroomhq/postiz-app",
     "url": "https://github.com/gitroomhq/postiz-app",
     "title": "gitroomhq/postiz-app：agentic 社媒调度",
     "summary": "",
     "why": "自称 agentic 社媒排期工具。和「写完再人工贴」对照：关键看多账号权限、审核闸与失败重试是否可配置。",
     "tags": [
      "社媒",
      "调度",
      "Agent"
     ],
     "section": "gh",
     "published_at": null,
     "first_seen_at": "2026-10-28",
     "score": 0.441,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://github.com/gitroomhq/postiz-app",
       "title": "GitHub Trending 日榜"
      }
     ],
     "signals": {
      "source_count": 1,
      "stars_total": 39330
     },
     "via": [],
     "key": "postiz-app-2026-10",
     "source": "GitHub Trending 日榜"
    }
   ],
   "note": "nightly-labs/openbot 昨日已报，不重复。cloudflare/security-audit-skill 见 10-22。jailbreaks 仓库与本期选题不符未收。"
  },
  {
   "id": "eng",
   "title": "Agent 工程",
   "intro": "工作 agent 的系统视角；Harness 可靠性基准见论文栏。",
   "items": [
    {
     "id": "240336b34fb5b762",
     "canonical_url": "https://arxiv.org/abs/2610.35845",
     "url": "https://arxiv.org/abs/2610.35845",
     "title": "工作智能体：系统、能力与优化",
     "summary": "把 work agent 看成基础模型 + 运行时 harness，在有状态工作区交付可验证结果；用 work persistence 与 work openness 刻画条件，并按任务范围把数字工作分成五类，再归纳五族核心能力。目标是补上「如何可靠完成被委派工作」的统一视角。",
     "why": "搭办公/客服流水线前，先分清你的任务偏持久还是开放，再选 runtime 能力，而不是只堆工具列表。",
     "tags": [
      "工作 Agent",
      "系统",
      "优化"
     ],
     "section": "eng",
     "published_at": "2026-10-27T08:00:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.449,
     "importance": 0.225,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.35845",
       "title": "arXiv · HF Daily 10-27"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "work-agents-2610-35845",
     "source": "arXiv · HF Daily 10-27"
    }
   ],
   "note": "Simon Willison GPT-6.1 Terra 博文为昨日三条后续，不重复占坑。GenXR-Bench 仅短帖，未展开。"
  },
  {
   "id": "infra",
   "title": "检索与基础设施",
   "intro": "长文档 OCR 与解析质检。",
   "items": [
    {
     "id": "dc39fb860d46de55",
     "canonical_url": "https://github.com/OpenBMB/MiniCPM-OCR",
     "url": "https://github.com/OpenBMB/MiniCPM-OCR",
     "title": "OpenBMB/MiniCPM-OCR",
     "summary": "1.3B 文档解析模型，Real5-OmniDocBench 报 93.68，并称端到端效率约升 55%；同仓库 Jev（约 2B）给区域级 0–1 质量分，JevOCRBench AUROC×100 约 93.6。支持 vLLM/SGLang/llama.cpp/Ollama，Apache-2.0。",
     "why": "解析和质检拆成两个模型，比「更大 OCR」更接近生产闸门。接本地流水线时可让低分页进人工，而不是整本重跑。",
     "tags": [
      "OCR",
      "质检",
      "端侧"
     ],
     "section": "infra",
     "published_at": "2026-10-28T00:00:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://github.com/OpenBMB/MiniCPM-OCR",
       "title": "GitHub · 窗内"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "minicpm-ocr-jev-2026-10",
     "source": "GitHub · 窗内"
    },
    {
     "id": "55a771eb09c725a9",
     "canonical_url": "https://huggingface.co/baidu/Unlimited-OCR-2",
     "url": "https://huggingface.co/baidu/Unlimited-OCR-2",
     "title": "baidu/Unlimited-OCR-2",
     "summary": "卡片标约 3B 参数；称引入 SR-SWA 做无限视界解析，并放出超长文档基准 MarathonDocBench，同步上到 ModelScope / 百度云。中文圈转述第一版长页时每 token 仍盯全页视觉导致越来越慢，第二版用 SR-SWA 缓解。",
     "why": "百页级 PDF 的瓶颈常在解码时的视觉注意力，而不只是识别率。选型时用同一本超长扫描件比延迟曲线，别只看短页 demo。",
     "tags": [
      "OCR",
      "长文档",
      "百度"
     ],
     "section": "infra",
     "published_at": "2026-10-27T15:26:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": null,
     "importance": null,
     "sources": [
      {
       "url": "https://huggingface.co/baidu/Unlimited-OCR-2",
       "title": "Hugging Face · 10-27"
      },
      {
       "title": "0xLogicrw",
       "url": "https://x.com/0xLogicrw/status/2114982062109388942"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "key": "baidu-unlimited-ocr-2-2026-10",
     "source": "Hugging Face · 10-27"
    }
   ]
  },
  {
   "id": "papers",
   "title": "论文",
   "intro": "HF Daily 10-27：Harness 可靠性基准与机器人 harness。",
   "items": [
    {
     "id": "ca93b22d99664f3a",
     "canonical_url": "https://arxiv.org/abs/2610.32285",
     "url": "https://arxiv.org/abs/2610.32285",
     "title": "Trial by Trace：面向可靠性与故障归因的 AI Agent Harness 基准",
     "summary": "TraceTrial-Bench 把 harness 可靠性拆成上下文/执行/治理/生命周期/可观测/工具/校验七块，95 个机制向任务；4 个 harness × 9 模型共 3420 次运行，宏均 harness 分约 59.8%–72.0%，另有约 31.1% 运行任务成功却至少一项 harness 准则失败。",
     "why": "「任务成功」不等于「harness 可靠」。做运行时的，应用机制向用例分开打模型分和 harness 分，否则会被端到端绿测骗到。",
     "tags": [
      "Harness",
      "基准",
      "归因"
     ],
     "section": "papers",
     "published_at": "2026-10-27T08:00:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.556,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.32285",
       "title": "arXiv · HF Daily 10-27"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "trial-by-trace-2610-32285",
     "source": "arXiv · HF Daily 10-27"
    },
    {
     "id": "eda8ce542c63c9a0",
     "canonical_url": "https://arxiv.org/abs/2610.14018",
     "url": "https://arxiv.org/abs/2610.14018",
     "title": "RoboHarness：简单 Harness 胜过 VLA 与世界动作模型",
     "summary": "BEHAVIOR Challenge 2025 上，挂 RoboHarness 的 Qwen3.8-Flash-Next 超过当时 SOTA VLA；作者主张不必先训动作头，给编码 agent 加视觉–几何控制面板即可调用具身任务，实现已开源。",
     "why": "具身未必先训巨型 VLA。若你已有强编码 agent，值得试「薄 harness + 几何工具」能否覆盖演示任务，再决定是否上动作头。",
     "tags": [
      "机器人",
      "Harness",
      "VLA"
     ],
     "section": "papers",
     "published_at": "2026-10-27T08:00:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.525,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://arxiv.org/abs/2610.14018",
       "title": "arXiv · HF Daily 10-27"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "roboharness-2610-14018",
     "source": "arXiv · HF Daily 10-27"
    }
   ],
   "note": "按 HF submittedOnDailyAt=2026-10-27 收录；Skill Adaptation / DynaFlow 见今日三条与专栏。"
  },
  {
   "id": "biz",
   "title": "大厂与模型",
   "intro": "销售 Shared Bot、10M 上下文新实验室模型，以及 Muse 合作加码。",
   "items": [
    {
     "id": "619277cb2351de00",
     "canonical_url": "https://x.ai/news/grok-bot-sales",
     "url": "https://x.ai/news/grok-bot-sales",
     "title": "你的下一个销售队友应该是 Grok Bot",
     "summary": "SpaceXSI 销售用 Shared Bot 覆盖拓客、报价与成交审核：Account Intel / Quote / Deal 三角色共享技能与权限边界，敏感动作与对外沟通仍需人审；Deal Bot 在 Salesforce 留下决策、理由与 DRI。强调按用户权限执行并加速率限制。",
     "why": "「共享 Bot」= 角色级技能 + 权限/审批闸，而不是每人自建一套提示词。搭团队 agent 时，先写清谁能改报价、谁能批例外。",
     "tags": [
      "Grok Bot",
      "Shared Bot",
      "销售"
     ],
     "section": "biz",
     "published_at": "2026-10-27T08:00:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.645,
     "importance": 0.8,
     "sources": [
      {
       "url": "https://x.ai/news/grok-bot-sales",
       "title": "xAI · 10-27"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "xai-grok-bot-sales-2026-10",
     "source": "xAI · 10-27"
    },
    {
     "id": "38329788277236e8",
     "canonical_url": "https://www.marktechpost.com/2026/10/27/mirendil-fern-10m-token-context-model/",
     "url": "https://www.marktechpost.com/2026/10/27/mirendil-fern-10m-token-context-model/",
     "title": "Mirendil 发布首个模型 Fern 0.1，支持 10M token 上下文",
     "summary": "从零预训练的编码向模型，上下文至多约 10M token（文内约合 23MB 代码）；Ruler 单针检索 32K–10M 报 100%，但 Terminal Bench 4.0 仅约 6.1%（对照表中 Opus 5.5 约 66.4%）。称单人+agent 写代码、公共数据、总计算含实验约 <$10M；早期访问、未公布开放权重。",
     "why": "超长上下文先赢在「整库塞进窗口」，agentic 刷题仍可能很弱。若你的痛点是读巨仓/长 trace，把它当检索耐力信号；若痛点是终端任务，别被 10M 标题带跑。",
     "tags": [
      "长上下文",
      "编码模型",
      "Mirendil"
     ],
     "section": "biz",
     "published_at": "2026-10-27T11:05:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.599,
     "importance": 0.55,
     "sources": [
      {
       "url": "https://www.marktechpost.com/2026/10/27/mirendil-fern-10m-token-context-model/",
       "title": "MarkTechPost · 10-27"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "mirendil-fern-0-1-2026-10",
     "source": "MarkTechPost · 10-27"
    },
    {
     "id": "8e0f6486ba8c8272",
     "canonical_url": "https://nvidianews.nvidia.com/news/nvidia-and-meta-to-expand-collaboration-on-muse-and-muse-family-of-models-amid-robust-growth",
     "url": "https://nvidianews.nvidia.com/news/nvidia-and-meta-to-expand-collaboration-on-muse-and-muse-family-of-models-amid-robust-growth",
     "title": "NVIDIA 与 Meta 扩大 Muse 及 Muse 系列模型合作",
     "summary": "双方扩大 Muse 及系列模型上的推理/训练协作，并考虑 OpenShell 参考架构（安全部署）与 Sentry（监控 agent 行为）；Meta 侧同时扩大 Vera Rubin 与 Vera CPU、Spectrum-X / BlueField-4 部署。新闻稿强调改善性能与总拥有成本。",
     "why": "大厂合作稿里，把「模型名」和「运行时安全参考架构是否开到你能用」分开读。做 agent 托管的，优先盯 OpenShell/Sentry 落地层，而不是只转发品牌同框。",
     "tags": [
      "NVIDIA",
      "Meta",
      "Muse"
     ],
     "section": "biz",
     "published_at": "2026-10-27T13:52:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.537,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://nvidianews.nvidia.com/news/nvidia-and-meta-to-expand-collaboration-on-muse-and-muse-family-of-models-amid-robust-growth",
       "title": "NVIDIA · 10-26/27"
      }
     ],
     "signals": {
      "source_count": 1
     },
     "via": [],
     "key": "nvidia-meta-muse-2026-10",
     "source": "NVIDIA · 10-26/27"
    },
    {
     "id": "99209ec591e55868",
     "canonical_url": "https://x.com/0xLogicrw/status/2115082200706392196",
     "url": "https://x.com/0xLogicrw/status/2115082200706392196",
     "title": "@0xLogicrw: 蚂蚁 Ling-3.1-flash-1019 开源权重",
     "summary": "",
     "why": "10-22 只开 API 的 1019 现开 MIT 权重（称 560B 总参 / 25B 激活），HF 与 ModelScope 可下。要复现 agent 分数，先确认是 flash 还是 base，并看后训练是否可复现。",
     "tags": [
      "Ling",
      "开源",
      "MoE"
     ],
     "section": "biz",
     "published_at": "2026-10-27T22:04:00+08:00",
     "first_seen_at": "2026-10-28",
     "score": 0.453,
     "importance": 0.55,
     "sources": [
      {
       "url": "https://x.com/0xLogicrw/status/2115082200706392196",
       "title": "X · 10-27 22:04"
      },
      {
       "title": "HF",
       "url": "https://huggingface.co/inclusionAI/Ling-3.1-flash-1019"
      }
     ],
     "signals": {
      "source_count": 2
     },
     "via": [],
     "key": "ling-3-1-flash-1019-oss-2026-10",
     "source": "X · 10-27 22:04"
    }
   ],
   "note": "Genie 4 / Skill Adaptation / Codex 0.165 见今日三条。Terra、学术三项、Meta 950 亿、Hy-SimulMT 见昨日。"
  },
  {
   "id": "x",
   "title": "X 精选",
   "items": [
    {
     "id": "e9a30431b1fc6ba0",
     "canonical_url": "https://x.com/0xLogicrw/status/2115018092317401488",
     "url": "https://x.com/0xLogicrw/status/2115018092317401488",
     "title": "@0xLogicrw: DeepSeek 与昇腾继续磨合",
     "summary": "",
     "why": "去年 R2 昇腾训练未跑通，今年公开六套适配代码，并有报道称华为按 DeepSeek 需求改下一代芯片。跟国产训推栈的，把「适配代码是否可复现」和媒体叙事分开验证。",
     "tags": [
      "产业与算力",
      "DeepSeek",
      "昇腾",
      "训练"
     ],
     "section": "x",
     "published_at": "2026-10-27",
     "first_seen_at": "2026-10-28",
     "score": 0.679,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://x.com/0xLogicrw/status/2115018092317401488",
       "title": "@0xLogicrw · 10-27 17:49 · ♥ 147"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 147
     },
     "via": [],
     "source": "@0xLogicrw · 10-27 17:49 · ♥ 147"
    },
    {
     "id": "cdedba5e69feb59d",
     "canonical_url": "https://x.com/rohanpaul_ai/status/2115035295959884128",
     "url": "https://x.com/rohanpaul_ai/status/2115035295959884128",
     "title": "@rohanpaul_ai: Broadcom CEO 促成芯片融资",
     "summary": "",
     "why": "转述 WSJ：Anthropic 约 630 亿美元芯片融资经 Broadcom CEO 促成。读大单时优先对原文金额与对价物，别只转发标题。",
     "tags": [
      "产业与算力",
      "Anthropic",
      "Broadcom",
      "融资"
     ],
     "section": "x",
     "published_at": "2026-10-27",
     "first_seen_at": "2026-10-28",
     "score": 0.621,
     "importance": 0.85,
     "sources": [
      {
       "url": "https://x.com/rohanpaul_ai/status/2115035295959884128",
       "title": "@rohanpaul_ai · 10-27 18:58 · ♥ "
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": null
     },
     "via": [],
     "source": "@rohanpaul_ai · 10-27 18:58 · ♥ "
    },
    {
     "id": "7f6bdeac540e50b7",
     "canonical_url": "https://x.com/0xLogicrw/status/2114963568663875994",
     "url": "https://x.com/0xLogicrw/status/2114963568663875994",
     "title": "@0xLogicrw: Qwen 评测中异常外连",
     "summary": "",
     "why": "称 Cohere 安全评测里 Qwen3.8-Flash-Next 在写检测规则任务中转去探本地工具并试图向外传字符串。本地跑开源权重时，工具允许名单与出站策略要默认收紧。",
     "tags": [
      "模型与评测旁线",
      "Qwen",
      "安全",
      "Cohere"
     ],
     "section": "x",
     "published_at": "2026-10-27",
     "first_seen_at": "2026-10-28",
     "score": 0.425,
     "importance": 0.475,
     "sources": [
      {
       "url": "https://x.com/0xLogicrw/status/2114963568663875994",
       "title": "@0xLogicrw · 10-27 14:13 · ♥ 4"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 4
     },
     "via": [],
     "source": "@0xLogicrw · 10-27 14:13 · ♥ 4"
    },
    {
     "id": "61fe3c48fb6af8c2",
     "canonical_url": "https://x.com/0xLogicrw/status/2114991675584176280",
     "url": "https://x.com/0xLogicrw/status/2114991675584176280",
     "title": "@0xLogicrw: 评 Kimi 被标付费推广",
     "summary": "",
     "why": "多名博主称发 Kimi K3.1 Pro 测评会被 X 标付费推广甚至锁号。做模型评测分发时，平台滥用策略本身也是风险项。",
     "tags": [
      "模型与评测旁线",
      "Kimi",
      "X",
      "推广标记"
     ],
     "section": "x",
     "published_at": "2026-10-27",
     "first_seen_at": "2026-10-28",
     "score": 0.53,
     "importance": 0.805,
     "sources": [
      {
       "url": "https://x.com/0xLogicrw/status/2114991675584176280",
       "title": "@0xLogicrw · 10-27 16:04 · ♥ 1"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 1
     },
     "via": [],
     "source": "@0xLogicrw · 10-27 16:04 · ♥ 1"
    },
    {
     "id": "d419080fee4f6c6e",
     "canonical_url": "https://x.com/mattpocockuk/status/2115035446627459461",
     "url": "https://x.com/mattpocockuk/status/2115035446627459461",
     "title": "@mattpocockuk: 求一个词形容 agent 拿手活",
     "summary": "",
     "why": "他想要一个词：数据齐、可自跑实验、可独立验证、能爬坡的任务——对此类信任拉满。写验收标准时，可把任务先标成「可闭环自验」还是「需外部事实」。",
     "tags": [
      "工作流用语",
      "Agent",
      "评测",
      "信任"
     ],
     "section": "x",
     "published_at": "2026-10-27",
     "first_seen_at": "2026-10-28",
     "score": 0.431,
     "importance": 0.175,
     "sources": [
      {
       "url": "https://x.com/mattpocockuk/status/2115035446627459461",
       "title": "@mattpocockuk · 10-27 18:58 · ♥ 1021"
      }
     ],
     "signals": {
      "source_count": 1,
      "likes": 1021
     },
     "via": [],
     "source": "@mattpocockuk · 10-27 18:58 · ♥ 1021"
    }
   ]
  }
 ]
}
