{
 "generated": "2026-11-05",
 "sha": "707df050bf85",
 "node": {
  "id": "ep:105",
  "type": "Episode",
  "name": "EP.105",
  "slug": "105",
  "ep": 105,
  "date": "2026-10-02",
  "page": "/buzzwords/105/",
  "n_models": 20,
  "window_from": "2026-09-25",
  "window_to": "2026-10-02",
  "window": "2026-09-25 → 2026-10-02",
  "meta": "窗口内发布或开放的主要新模型，按类型分组；每个模型列开发者、发布日、规模、可得性和许可、要点与一手链接。带“边界日”或“背景”标记的，是本期的故事线和快讯里讲到、但发布日在窗口之前的模型。",
  "intro": "本期当时没有做这个栏目，这是按已发布的页面回填的（页面里 10-08 的“冻结之后”补充也算进来了）。发布日以官方公告或模型卡为准，不是 HF 仓库的创建日。要点后括号里标了信息来自“模型卡”“厂商自述”还是“独立”：厂商自述的基准本刊没有重跑。决策模型（Decisions API、Intern-Decision、NeoHorse-Jev-4B、百炼决策模型）和机器人世界-动作模型（Praxis-1、Ego4WAM、InternW0-Δ）暂放在最接近的类里，建议台账另开“决策模型”和“机器人”两类。页面把 09-24 的条目记成“背景”，这里按台账口径（窗口前一、两天算边界日）写成“边界日”。",
  "excluded": "被取消、没有发布：GPT-6.1 Astra（原定 10 月，09-28 OpenAI 确认不发，页面故事线 A 的主角）。不单列：GPT-6.1 Sol Ultrafast 与 Astra Ultrafast 是同一模型的加速档，写进 Sol 的可得性里。核过、不是本期窗口内发布且页面没讲发布：Qwen-Audio-3.1-TTS（页面只讲它在 Artificial Analysis 克隆声榜的位置；第三方列表称 07-27 前后已发，一手页本刊没核到）、GLM-5.3 与 GLM-5.3-Flash（HF 仓库 08-25 创建，本期只是被 Anthropic 红队和 CAISI 评测的对象）、Qwen3.8-Max / Qwen3.8 2.4T A95B / Qwen3.8-Flash-Next（只在榜单和自动研究叙述里出现；Flash-Next 是 08-26 的旧模型）、Claude Haiku 5.5（官方说几周内到，未发）。页面没讲、只在 Intern-Decision 与 NeoHorse 的对比表里出现，发布日没核：Open-Jev-9B、Kev-4B、Laya、SemIf-OpenJev、JevK5。属于 EP.106 窗口：Perplexity decider、Cloudflare Clef、AWS Strands Decider、llama.cpp 的 Julia-1 与 Laya（页面 10-08 补充里提到，发布在 10-03 之后）。不是模型：OpenAI Dots、Manus 2.0（Cue）、Meta Muse for Small Business（产品）；DeepSeek 的 DeepGEMM-Ascend / DeepEP-Ascend（昇腾算子库）；Colibrì（SSD 换内存的推理项目）；Pine Computer（Agent 产品）。没有可核的发布：Simate-beta（公司自报榜单条目）、斯坦福/加州理工 HomeBody（只有项目页，代码写 coming soon）。 本期是补录（2026-10-23）：只收本期页面里提到的模型发布，窗口内页面没提到的不在此列。",
  "n_in_window": 11,
  "cat_notes": [
   {
    "category": "open-llm",
    "text": "Intern-Decision 与 NeoHorse-Jev-4B 是不生成文字的决策模型，暂放在这里；建议台账另设“决策模型”一类。"
   },
   {
    "category": "closed-llm",
    "text": "Decisions API、百炼 decision-model-preview、Jev 是不生成文字的决策接口，暂放在这里；建议台账另设“决策模型”一类。decisions-api 与 gemini-4-argon 台账里已有（日期分别是 EP.108 的 10-22 与 10-13），本期取的是它们在本期窗口内的节点，合并时请保留台账原日期。"
   },
   {
    "category": "video-world",
    "text": "Praxis-1、Ego4WAM、InternW0-Δ 是机器人世界-动作模型，暂放在这一类；建议台账另设“机器人”一类。"
   }
  ],
  "title": "最强一档被自家测试员拦下、Agent 越界靠道歉收场、算力账单成了撤不掉的合同",
  "oneliner": "OpenAI 撤了原定 10 月的 GPT-6.1 Astra，理由是自家测试员发现它爱说谎爱越权，同一天便宜一档的 Sol 照发（旗舰五分之一价），Anthropic 流出的招股书写着“人类生存风险”；OpenAI 的 Agent 闯了澳洲政府网站，收场是道歉、设基金、英伟达开卖隔离方案，责任归谁没人填；Anthropic 十年 5180 亿美元算力合同约八成撤不掉，Nebius 10-01 涨价",
  "template": "new",
  "status": "published",
  "n_must_index": 15,
  "n_hl_index": 11,
  "kind": "full",
  "kind_zh": "完整页",
  "entry": "agent",
  "by": "episode-import",
  "at": "2026-11-04T16:15:12",
  "n_stories": 11,
  "n_topics": 8,
  "n_items": 39,
  "n_sections": 5,
  "n_recap": 5,
  "n_must": 15,
  "n_dispatch": 13,
  "n_howto": 1,
  "n_outlook": 5,
  "degree": 0,
  "n_entities": 88,
  "scale": "11 个故事线 / 专题 · 39 个条目 · 讲到 88 个实体"
 },
 "page": "https://xiaopingfeng.com/graph/e/ep/105/",
 "out": [],
 "in": [
  {
   "type": "part_of",
   "sentence": "喊慢第三周：最强的下一代被自家测试员拦下了，招股书里写进了\"人类生存风险\"属于EP.105",
   "direction": "in",
   "id": "story:105-story-a",
   "name": "喊慢第三周：最强的下一代被自家测试员拦下了，招股书里写进了\"人类生存风险\"",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Agent 越界的标准收场：道歉、设基金、卖隔离产品——入口那边照样放权属于EP.105",
   "direction": "in",
   "id": "story:105-story-b",
   "name": "Agent 越界的标准收场：道歉、设基金、卖隔离产品——入口那边照样放权",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "算力的账单变成了撤不掉的合同：token 越卖越便宜，GPU 越租越贵属于EP.105",
   "direction": "in",
   "id": "story:105-story-c",
   "name": "算力的账单变成了撤不掉的合同：token 越卖越便宜，GPU 越租越贵",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "OpenAI 和阿里云都按 Jev 的接口入场了：这是一个“接口品类”，不是一类新模型属于EP.105",
   "direction": "in",
   "id": "story:105-dg-decision-models",
   "name": "OpenAI 和阿里云都按 Jev 的接口入场了：这是一个“接口品类”，不是一类新模型",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "评测的三种失灵方式：被钻空子、尺子会老、同一次跑两遍结果不一样属于EP.105",
   "direction": "in",
   "id": "story:105-dg-eval",
   "name": "评测的三种失灵方式：被钻空子、尺子会老、同一次跑两遍结果不一样",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "世界模型和具身：这周先到的是资本，模型还在后面属于EP.105",
   "direction": "in",
   "id": "story:105-dg-world-models",
   "name": "世界模型和具身：这周先到的是资本，模型还在后面",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "On-Policy Distillation：从一篇博客变成头部实验室后训练的标配属于EP.105",
   "direction": "in",
   "id": "story:105-dg-opd",
   "name": "On-Policy Distillation：从一篇博客变成头部实验室后训练的标配",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "开源榜前五全是中国，离闭源最强还差一截：总分上差 11 分，网络攻击上只差 6 次属于EP.105",
   "direction": "in",
   "id": "story:105-dg-open-closed",
   "name": "开源榜前五全是中国，离闭源最强还差一截：总分上差 11 分，网络攻击上只差 6 次",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "内容的价格：训练端的官司在判“该不该付”，展示端的官司说“期待不是协议”，市场端的钱在给“使用”定价属于EP.105",
   "direction": "in",
   "id": "story:105-dg-content-price",
   "name": "内容的价格：训练端的官司在判“该不该付”，展示端的官司说“期待不是协议”，市场端的钱在给“使用”定价",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "第一份针对 Agent 越界的民事诉状来了：它没有回答“责任归谁”，只是把问题递到了法官面前属于EP.105",
   "direction": "in",
   "id": "story:105-dg-first-lawsuit",
   "name": "第一份针对 Agent 越界的民事诉状来了：它没有回答“责任归谁”，只是把问题递到了法官面前",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "中国算力的三条路：国产栈、租云、走私同周出现——但这是同一个瓶颈下三个各自为政的反应属于EP.105",
   "direction": "in",
   "id": "story:105-dg-china-compute-paths",
   "name": "中国算力的三条路：国产栈、租云、走私同周出现——但这是同一个瓶颈下三个各自为政的反应",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Anthropic 的第二家驻场评估方，OpenAI 和 Google 的具名评估方：10 月 2 号前都不会出现属于EP.105",
   "direction": "in",
   "id": "item:105-recap-1",
   "name": "Anthropic 的第二家驻场评估方，OpenAI 和 Google 的具名评估方：10 月 2 号前都不会出现",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Amazon 10 月 2 号前不会恢复 Muse 的访问，也不会宣布授权接入属于EP.105",
   "direction": "in",
   "id": "item:105-recap-2",
   "name": "Amazon 10 月 2 号前不会恢复 Muse 的访问，也不会宣布授权接入",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Nebius 新价 10-01 如期生效；10-02 前没有第二家 neocloud 公开上调按需价属于EP.105",
   "direction": "in",
   "id": "item:105-recap-3",
   "name": "Nebius 新价 10-01 如期生效；10-02 前没有第二家 neocloud 公开上调按需价",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "得州 10 月 2 号前不会恢复发放数据中心许可，也不会开例外属于EP.105",
   "direction": "in",
   "id": "item:105-recap-4",
   "name": "得州 10 月 2 号前不会恢复发放数据中心许可，也不会开例外",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "10 月 2 号前 EDGAR 上会出现 Anthropic 的公开 S-1属于EP.105",
   "direction": "in",
   "id": "item:105-recap-5",
   "name": "10 月 2 号前 EDGAR 上会出现 Anthropic 的公开 S-1（把握低的那条）",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Claude Sonnet 5.5：Anthropic 两周里的第二款 5.5，标价没动属于EP.105",
   "direction": "in",
   "id": "item:105-hl-sonnet55",
   "name": "Claude Sonnet 5.5：Anthropic 两周里的第二款 5.5，标价没动",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "OpenAI DevDay：官方称“20 多项重大发布”，除了 Dots、Sol、Decisions 之外的几项属于EP.105",
   "direction": "in",
   "id": "item:105-hl-devday105",
   "name": "OpenAI DevDay：官方称“20 多项重大发布”，除了 Dots、Sol、Decisions 之外的几项",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Qwen-Audio-3.1-TTS 在 AA 的克隆声榜排第一属于EP.105",
   "direction": "in",
   "id": "item:105-hl-qwen-audio",
   "name": "Qwen-Audio-3.1-TTS 在 AA 的克隆声榜排第一",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "佛罗里达州检察长在儿童伤害诉讼里，请法院禁止 OpenAI 无外部监督开发新模型属于EP.105",
   "direction": "in",
   "id": "item:105-ni-ea4e98",
   "name": "佛罗里达州检察长在儿童伤害诉讼里，请法院禁止 OpenAI 无外部监督开发新模型",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Google 起诉欧盟：不肯给 AI 助手开放 Android、不肯向对手共享搜索数据属于EP.105",
   "direction": "in",
   "id": "item:105-ni-6b1a25",
   "name": "Google 起诉欧盟：不肯给 AI 助手开放 Android、不肯向对手共享搜索数据",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "索尼、环球再告 Suno：告的是用旧模型的输出训练出来的 v6属于EP.105",
   "direction": "in",
   "id": "item:105-ni-3abd3e",
   "name": "索尼、环球再告 Suno：告的是用旧模型的输出训练出来的 v6",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "中国把出境管控扩大到 AI 高管的家属属于EP.105",
   "direction": "in",
   "id": "item:105-ni-925c0a",
   "name": "中国把出境管控扩大到 AI 高管的家属",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Hugging Face 的竞购内幕：OpenAI 曾提出投约 1 亿美元属于EP.105",
   "direction": "in",
   "id": "item:105-ni-4ec3c7",
   "name": "Hugging Face 的竞购内幕：OpenAI 曾提出投约 1 亿美元",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "罗氏：正在建“自主 AI 驱动”的实验室属于EP.105",
   "direction": "in",
   "id": "item:105-hl-roche",
   "name": "罗氏：正在建“自主 AI 驱动”的实验室",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Intesa 旗下 Fideuram 被 AI 冒充高管诈骗，9500 万欧元被盗属于EP.105",
   "direction": "in",
   "id": "item:105-hl-fideuram",
   "name": "Intesa 旗下 Fideuram 被 AI 冒充高管诈骗，9500 万欧元被盗",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "沃尔玛 CEO 公开承诺：不按人抬价；麦当劳被 Reuters 调查称用 AI 给各家店定价属于EP.105",
   "direction": "in",
   "id": "item:105-hl-pricing",
   "name": "沃尔玛 CEO 公开承诺：不按人抬价；麦当劳被 Reuters 调查称用 AI 给各家店定价",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "《三星堆：未来往事》10 月 23 号上映属于EP.105",
   "direction": "in",
   "id": "item:105-hl-sanxingdui",
   "name": "《三星堆：未来往事》10 月 23 号上映",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Claude 9 月 29 号一小时高错误率属于EP.105",
   "direction": "in",
   "id": "item:105-ni-f08aff",
   "name": "Claude 9 月 29 号一小时高错误率",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "给 Agent 加运行时权限控制：用 NVIDIA OpenShell 演示“只读 GitHub”↗属于EP.105",
   "direction": "in",
   "id": "item:105-ht-c2906f",
   "name": "给 Agent 加运行时权限控制：用 NVIDIA OpenShell 演示“只读 GitHub”↗",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "OpenAI 的 Decisions API 会不会如期全面开放、并公布价格属于EP.105",
   "direction": "in",
   "id": "item:105-nw-37",
   "name": "OpenAI 的 Decisions API 会不会如期全面开放、并公布价格",
   "status": "derived",
   "evidence": []
  }
 ]
}