{
  "eyebrow": "DAILY AI BRIEFING",
  "title": "AI 前沿日报",
  "date": "2026-10-09 周五",
  "tldr": [
    {
      "title": "办公 Agent 混战",
      "text": "Google 发布通用 Gemini agent，可按任务调用 Claude；Claude 同日上线 Dashboards 与 Motion。"
    },
    {
      "title": "OpenAI 收入降温",
      "text": "FT/CNBC：OpenAI 9 月底年化收入约 500 亿美元，低于此前流传的 680 亿，AI 股应声下跌。"
    },
    {
      "title": "AI 数学纠错",
      "text": "OpenAI 撤回 3 篇霍奇猜想相关手稿、修订 14 篇；陶哲轩主持的 AHM 呼吁抵制。"
    }
  ],
  "sections": [
    {
      "heading": "🔥 模型发布与评测",
      "items": [
        {
          "title": "GPT-6.1 Sol Ultrafast：速度最高 8 倍",
          "tag": "模型发布",
          "detail": "OpenAI 为 GPT-6.1 Sol 推出 Ultrafast 模式，覆盖 API、Codex 与 ChatGPT Work，官方称智能接近 Astra、速度最高为 Sol Standard 的 8 倍。据 TestingCatalog，定价 $12/$60（每百万 token）。Epoch 实测 6.1 Sol 缓存输入价减半、长提示处理更快。",
          "why": "对延迟敏感的编码与 Agent 循环可以直接换档",
          "source": "OpenAI Developers（X）· 10/9 02:26",
          "url": "https://x.com/OpenAIDevs/status/2108262812489531498",
          "related": [
            "Codex Day 4：即时 steering",
            "Epoch 实测"
          ],
          "related_urls": [
            "https://x.com/thsottiaux/status/2108275041276420573",
            "https://x.com/EpochAIResearch/status/2108306213943586979"
          ]
        },
        {
          "title": "Epoch 自动化报告：前沿模型还替代不了研究员",
          "tag": "评测",
          "detail": "Epoch AI 从自家真实研究工作中抽取开放式任务，并按内部标准人工评分。Claude Fable 5.1 与 GPT-6 Astra 领先，但离完全自动化 Epoch 的工作还很远：明确定义的任务做得稳，开放式环节仍会失败。",
          "why": "比静态 benchmark 更接近「能否替代知识工作」",
          "source": "Epoch AI · 10/9 01:36",
          "url": "https://epoch.ai/publications/can-ai-automate-epoch",
          "related_urls": [
            "https://epoch.ai/benchmarks/epoch-automation-reports"
          ]
        },
        {
          "title": "JetBrains Mellum2.1：12B MoE 编码小模型开源",
          "tag": "开源",
          "detail": "12B 总参、2.5B 激活的思考型 MoE，Apache 2.0，131K 上下文，架构不变，提升主要来自真实软件环境中的 RL。模型卡自报 LiveCodeBench v6 82.0、SWE-bench Verified 47.0、Terminal-Bench 2.1 17.4；GGUF 7GB 起。",
          "why": "可本地部署的编码子 Agent 底座",
          "source": "Hugging Face 模型卡（自报）· 10/8 02:43",
          "url": "https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking",
          "related": [
            "MarkTechPost 解读"
          ],
          "related_urls": [
            "https://www.marktechpost.com/2026/10/08/jetbrains-releases-mellum2-1-a-12b-moe-open-model-for-coding-agents/"
          ]
        }
      ]
    },
    {
      "heading": "🤖 Agent 与 AI 编程",
      "items": [
        {
          "title": "Google 发布通用 Gemini agent，可调用 Claude",
          "tag": "Agent",
          "detail": "Gemini at Work 大会上，Google Cloud CEO Thomas Kurian 发布面向企业的单一通用 Agent，入口在 Gemini Enterprise：一个输入框规划任务、调用工具、写并运行代码，并直接嵌入 Gmail/Docs/Sheets 等。它按任务自动挑选底层模型，目前覆盖 Gemini 与 Anthropic Claude。「同事 Agent」拥有独立邮箱、存储和身份。金融、法律行业版进入预览。",
          "why": "大厂首次在自家 Agent 里接入竞品模型，多模型路由成默认",
          "source": "Google Cloud 博客 · 10/8 22:28",
          "url": "https://cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026/",
          "related": [
            "The Verge",
            "量子位"
          ],
          "related_urls": [
            "https://www.theverge.com/tech/1007904/google-gemini-ai-agent-enterprise",
            "https://www.qbitai.com/2026/10/502083.html"
          ]
        },
        {
          "title": "Hermes Agent 上架微软商店",
          "tag": "Agent",
          "detail": "Nous Research 的开源 Agent Hermes Agent 上架 Microsoft Store（获官方推荐），支持一键安装；团队称更多 Windows 生态更新即将推出。",
          "why": "开源个人 Agent 开始走向普通 Windows 用户",
          "source": "Nous Research（X）· 10/9 00:22",
          "url": "https://x.com/NousResearch/status/2108231536772596193"
        }
      ]
    },
    {
      "heading": "📱 产品与应用",
      "items": [
        {
          "title": "Claude Dashboards 与 Motion 开启 beta",
          "tag": "产品",
          "detail": "Dashboards 连接 BigQuery、Snowflake、Databricks、Salesforce 等，用自然语言生成随数据自动更新的看板，每个数字都可点开查看背后的查询，也可以转到 Hex、Grafana 等工具继续分析。Motion 用代码生成可编辑的动画讲解，可导出 MP4，不使用视频生成模型。Dashboards 面向付费用户，Motion 面向 Team/Enterprise；Docs、Slides、Design 结束 beta，对所有用户（含免费）开放。",
          "why": "「问数据」和「做讲解动画」不再需要分析师",
          "source": "Claude 官方 · 10/9 03:01",
          "url": "https://claude.com/resources/articles/dashboards-and-motion",
          "related": [
            "X 主帖 2.5 万赞",
            "The Decoder"
          ],
          "related_urls": [
            "https://x.com/claudeai/status/2108271552991252810",
            "https://the-decoder.com/claude-can-now-generate-animated-explainer-videos-and-live-data-dashboards-from-text-prompts/"
          ]
        },
        {
          "title": "Google AI Edge Foresight：完全离线的会议笔记",
          "tag": "产品",
          "detail": "Google 推出实验性 macOS 笔记应用，用端侧 EmbeddingGemma 2 在本地转写和总结会议与音频，不连云端，免费使用，对标 Granola 和 Wispr Flow。",
          "why": "隐私敏感场景的本地 AI 笔记方案",
          "source": "The Verge · 10/8 23:27",
          "url": "https://www.theverge.com/tech/1007985/google-ai-notetaking-app-transcribe-offline",
          "related": [
            "TechCrunch"
          ],
          "related_urls": [
            "https://techcrunch.com/2026/10/08/google-releases-a-new-local-first-granola-competitor/"
          ]
        }
      ]
    },
    {
      "heading": "📄 论文与研究",
      "items": [
        {
          "title": "更新：OpenAI 撤回 3 篇 AI 数学手稿",
          "tag": "论文",
          "detail": "openai/math 仓库发布首份更新日志：撤回 3 篇、修订 14 篇，手稿从 722 篇降为 719 篇。撤回的 3 篇都与霍奇猜想（K3 曲面）相关，起因是一处关键论证的符号把 -1 写成了 +1。同时，陶哲轩主持的人类数学协会（AHM）呼吁数学家抵制与 OpenAI 合作。",
          "why": "AI 证明的验证成本，成了数学界的新瓶颈",
          "source": "GitHub openai/math · 10/8 18:27（机器之心报道）",
          "url": "https://github.com/openai/math/blob/main/history.md",
          "related": [
            "The Decoder：AHM 抵制",
            "机器之心"
          ],
          "related_urls": [
            "https://the-decoder.com/some-mathematicians-call-for-openai-boycott-after-ai-generated-proofs-flood-their-field/",
            "https://www.jiqizhixin.com/articles/2026-10-08-9"
          ]
        },
        {
          "title": "AMIE 首次真实门诊前瞻研究登上《柳叶刀》",
          "tag": "研究",
          "detail": "Google 与 BIDMC 在真实急诊/门诊中评估研究型医疗对话系统 AMIE（患者就诊前与它聊天）。100 次患者交互中没有触发任何安全叫停，90% 的病例与医生诊断一致。这是首个面向患者的对话式诊断前瞻性研究。",
          "why": "医疗 AI 从模拟考走向真实临床",
          "source": "Google Research / Sundar Pichai（X）· 10/9 06:50",
          "url": "https://x.com/sundarpichai/status/2108329198955958339",
          "related_urls": [
            "https://x.com/GoogleResearch/status/2108326125676192100"
          ]
        },
        {
          "title": "世界模型物理大考：最高只有 57.76 分",
          "tag": "评测",
          "detail": "Einsia 发布 World Models' Last Exam in Physics：用 40 个受控任务、9 类物理规律测 8 个 SOTA 视频世界模型，最高分 57.76/100。通过一致性筛查的自由落体视频，平均综合分也只有 26.61。MiniMax 称其开源 H3 在此榜接近闭源 SOTA（声称）。",
          "why": "「看起来真」不等于「符合物理」",
          "source": "arXiv 2610.08791 · 10/8 22:30",
          "url": "https://arxiv.org/abs/2610.08791",
          "related": [
            "GitHub",
            "机器之心"
          ],
          "related_urls": [
            "https://github.com/Einsia/phys-last-exam",
            "https://www.jiqizhixin.com/articles/2026-10-08-6"
          ]
        }
      ]
    },
    {
      "heading": "💼 行业 · 投融资 · 政策",
      "items": [
        {
          "title": "OpenAI 年化收入约 500 亿美元，低于传闻",
          "tag": "市场",
          "detail": "据 FT 首发、CNBC 证实（据报道），OpenAI 向投资人披露 9 月底年化收入约 500 亿美元，低于上月广泛流传的 680 亿美元。知情人士称 680 亿含合作伙伴的总收入。消息传出后，Nvidia、Oracle、CoreWeave 等 AI 股下跌。",
          "why": "AI 估值叙事第一次被收入口径「打折」",
          "source": "CNBC（FT 首发）· 10/9 00:52",
          "url": "https://www.cnbc.com/2026/10/08/open-ai-revenue-nvidia-oracle-coreweave.html",
          "related": [
            "HN 349 分"
          ],
          "related_urls": [
            "https://news.ycombinator.com/item?id=50008187"
          ]
        },
        {
          "title": "Anthropic 启动 Cyber Mission",
          "tag": "政策",
          "detail": "包含两项新举措：一是关键基础设施防御计划（CIDP），为电网、水务、交通等 OT 防御方提供前沿模型、驻场工程师和威胁研究，首批合作方包括 CrowdStrike（据 Axios）；二是 OSS Scanner，用最强模型免费为开源项目定期扫描漏洞。Project Glasswing 已并入扩大后的网安验证计划。",
          "why": "前沿模型厂商把攻防能力主动投向防守方",
          "source": "Anthropic 官方 · 10/9 05:04",
          "url": "https://www.anthropic.com/news/anthropic-cyber-mission",
          "related": [
            "The Verge：OSS Scanner"
          ],
          "related_urls": [
            "https://www.theverge.com/ai-artificial-intelligence/1008521/anthropic-open-source-oss-scanner"
          ]
        },
        {
          "title": "Anthropic 新版使用政策：禁止虐待 Claude",
          "tag": "政策",
          "detail": "年度政策更新于 11/12 生效。整合了针对影响力行动的「欺骗性活动」专章；新增 Claude 自主执行物理动作时的管控，明确医疗、金融等高风险用途的要求；首次把「持续且无谓地虐待模型」列为违规，可能导致封号。",
          "why": "Agent 进入物理世界，规则先行",
          "source": "Anthropic 官方 · 10/8",
          "url": "https://www.anthropic.com/news/2026-usage-policy-update",
          "related": [
            "The Decoder",
            "HN 142 评论"
          ],
          "related_urls": [
            "https://the-decoder.com/being-mean-to-claude-can-now-get-your-account-suspended-under-anthropics-new-tos/",
            "https://news.ycombinator.com/item?id=50008565"
          ]
        }
      ]
    },
    {
      "heading": "🇨🇳 国内动态",
      "note": "DeepSeek / Qwen / Kimi / GLM / MiniMax 本窗口无新模型发布（已查 X 官号与 HF 组织页）；阶跃 Step 5 限免、Meta 封禁字节广告（付费墙）等未收录。",
      "items": [
        {
          "title": "生数 Vidu Q4 Preview：视频生成 0.09 元/秒起",
          "tag": "模型发布",
          "detail": "Vidu Q4 正式版的创作者预览版，重点提升人物表演、镜头调度和复杂特效。支持最多 15 张参考图、参考音色输入，以及 2K/4K 输出；首发价 0.09 元/秒起，官方称同样预算最多可生成 5 倍内容。",
          "why": "国产视频模型继续打价格+可控性",
          "source": "机器之心（据报道）· 10/8 21:30",
          "url": "https://www.jiqizhixin.com/articles/2026-10-08-11"
        }
      ]
    },
    {
      "heading": "💬 讨论热点",
      "style": "bullets",
      "items": [
        {
          "title": "被解雇的 OpenAI 安全研究员发公开信",
          "detail": "Jasmine Wang、Tomek Korbak、Mikita Balesni 否认违规处理敏感信息，警告这会形成寒蝉效应；同日 Bengio 表示，重视安全的人或许该离开前沿实验室。",
          "url": "https://techcrunch.com/2026/10/08/fired-openai-safety-researchers-dispute-misconduct-claims-warn-of-chilling-effect/",
          "related_urls": [
            "https://x.com/ShakeelHashim/status/2108294430046486806"
          ]
        },
        {
          "title": "特朗普：用「人工智能」一词的人是「敌人」",
          "detail": "特朗普要求改称「超级智能」（Super Intelligence），HN 52 评论；Simon Willison 直言「太蠢了」。",
          "url": "https://x.com/Techmeme/status/2108243700753313944",
          "related_urls": [
            "https://x.com/simonw/status/2108346075874734127"
          ]
        },
        {
          "title": "AI 数学进展引发密码学担忧",
          "detail": "密码学家 Matthew Green 发帖「我们可能会失去公钥密码学」；以太坊研究者讨论钱包安全的「地堡模式」（HN 55/58 分）。",
          "url": "https://news.ycombinator.com/item?id=50012043",
          "related_urls": [
            "https://the-decoder.com/ai-math-breakthroughs-have-ethereum-researchers-debating-how-fast-wallet-security-could-collapse/"
          ]
        }
      ]
    }
  ],
  "brand": "AI 前沿日报"
}