[
  {
    "id": "code-metr-rct",
    "sector": "code",
    "title": "熟悉代码库的开发者：AI 并非总能提速",
    "organization": "METR",
    "published": "2025-07-10",
    "type": "随机对照研究",
    "url": "https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/",
    "takeaway": "16 位资深开源开发者处理 246 项真实任务；允许使用当时 AI 工具时，完成时间增加 19%。",
    "limitation": "仅代表 2025 年初工具及熟悉大型仓库的样本；原页已链接 2026 年后续研究，不能作为当前全部工具结论。",
    "tags": [
      "真实任务",
      "生产率",
      "研究边界"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "code-claude-code",
    "sector": "code",
    "title": "Claude Code：终端中的编程代理",
    "organization": "Anthropic",
    "published": "2025-02-24",
    "type": "产品发布",
    "url": "https://www.anthropic.com/news/claude-3-7-sonnet",
    "takeaway": "Claude Code 以有限研究预览发布，允许开发者从终端委派工程任务；与 Claude 3.7 Sonnet 同时推出。",
    "limitation": "这是历史发布档案；企业自评与演示不等于独立验证，也不证明可以免除测试和代码审查。",
    "tags": [
      "编程代理",
      "终端",
      "工作流"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "daily-uk-playbook",
    "sector": "daily",
    "title": "英国政府 AI Playbook：从选型走向安全部署",
    "organization": "英国 Government Digital Service",
    "published": "2025-02-10",
    "type": "政府实践指南",
    "url": "https://www.gov.uk/government/publications/ai-playbook-for-the-uk-government",
    "takeaway": "新版手册提出公务人员使用 AI 的十项原则，并说明 AI 的能力、风险及政府选择、采购和部署方法；扩展了 2024 年生成式 AI 框架。",
    "limitation": "指南服务于英国政府情境，不是其他司法辖区的合规结论，也不保证具体系统部署有效。",
    "tags": [
      "政务",
      "采购",
      "部署治理"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "science-praim",
    "sector": "science",
    "title": "PRAIM：德国乳腺筛查的真实世界研究",
    "organization": "吕贝克大学等研究团队",
    "published": "2025-01-07",
    "type": "同行评审观察研究",
    "url": "https://www.nature.com/articles/s41591-024-03408-6",
    "takeaway": "研究纳入 463,094 名接受筛查的女性；AI 支持组癌症检出率为每千人 6.7，比较组为 5.7，调整后的相对提高为 17.6%。",
    "limitation": "观察性实施研究不能等同随机试验；检出率提高不直接证明死亡率下降，适用人群和工作流程也有限定。",
    "tags": [
      "医学影像",
      "真实世界",
      "乳腺筛查"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "media-veo2",
    "sector": "media",
    "title": "Veo 2：电影语言成为视频生成接口",
    "organization": "Google",
    "published": "2024-12-16",
    "type": "模型发布",
    "url": "https://blog.google/innovation-and-ai/models-and-research/google-labs/video-image-generation-update-december-2024/",
    "takeaway": "官方介绍 Veo 2 可响应镜头、类型及电影效果描述，并披露最高 4K、分钟级生成能力。",
    "limitation": "模型能力上限不等于各产品入口的输出上限；复杂动作和长视频一致性仍需逐镜检查。",
    "tags": [
      "视频生成",
      "电影语言",
      "Veo"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "robotics-pi0",
    "sector": "robotics",
    "title": "π0：跨机器人本体的通用策略原型",
    "organization": "Physical Intelligence",
    "published": "2024-10-31",
    "type": "研究发布",
    "url": "https://www.pi.website/blog/pi0",
    "takeaway": "π0 结合图像、文本和动作数据，可直接输出底层运动指令，并面向多种机器人进行训练和任务适配。",
    "limitation": "作者明确称其为早期原型；跨机器人训练不等于无需额外数据即可执行任意家庭任务。",
    "tags": [
      "机器人基础模型",
      "动作生成",
      "多任务"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "sports-table-tennis",
    "sector": "sports",
    "title": "机器人乒乓球：29 场人机比赛的实测",
    "organization": "Google DeepMind",
    "published": "2024-08-07",
    "type": "研究论文",
    "url": "https://arxiv.org/abs/2408.03906",
    "takeaway": "研究以 29 场机器人对人类比赛评估策略，机器人赢得 13 场，胜率约 45%，目标为业余人类水平。",
    "limitation": "样本量及对手水平限定了结论，不能表述为击败职业选手或达到世界冠军水平。",
    "tags": [
      "乒乓球",
      "运动机器人",
      "实测"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "robotics-bmw",
    "sector": "robotics",
    "title": "Figure 02 在 BMW 工厂完成钣金放置试验",
    "organization": "BMW Group",
    "published": "2024-08-06",
    "type": "企业现场试验",
    "url": "https://www.press.bmwgroup.com/global/article/detail/T0444265EN/successful-test-of-humanoid-robots-at-bmw-group-plant",
    "takeaway": "BMW 披露在 Spartanburg 工厂开展数周试验，Figure 02 将钣金件放入指定夹具，随后用于底盘装配。",
    "limitation": "该来源证实的是有限时段和特定工序的试验，不是当时已实现全厂自主生产或经济回报。",
    "tags": [
      "人形机器人",
      "制造",
      "现场试验"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "design-flux1",
    "sector": "design",
    "title": "FLUX.1：三种发布方式与不同许可",
    "organization": "Black Forest Labs",
    "published": "2024-08-01",
    "type": "模型发布",
    "url": "https://bfl.ai/blog/24-08-01-bfl",
    "takeaway": "FLUX.1 发布 pro、dev、schnell 三个版本；schnell 使用 Apache 2.0，dev 的初始定位为非商业应用的开放权重模型。",
    "limitation": "开放权重不等于统一商业许可；商业使用必须核对具体版本条款，不能混用三种版本的授权结论。",
    "tags": [
      "开放权重",
      "图像生成",
      "许可"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "media-gen3",
    "sector": "media",
    "title": "Gen-3 Alpha：以时间描述强化镜头控制",
    "organization": "Runway",
    "published": "2024-06-17",
    "type": "模型发布",
    "url": "https://runway.com/research/introducing-gen-3-alpha",
    "takeaway": "模型联合训练视频和图像，利用密集时间描述改善场景过渡和关键元素控制；发布计划包含内容审核和 C2PA 来源标准。",
    "limitation": "页面样片由厂商选择，不是独立成功率评测；发布计划与当时实际开放功能应区分。",
    "tags": [
      "视频生成",
      "镜头控制",
      "内容溯源"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "education-learnlm",
    "sector": "education",
    "title": "LearnLM：面向学习过程的模型微调",
    "organization": "Google",
    "published": "2024-05-14",
    "type": "研究与产品发布",
    "url": "https://blog.google/products-and-platforms/products/education/google-learnlm-gemini-generative-ai/",
    "takeaway": "Google 推出以教育研究为依据、面向学习微调的 LearnLM 模型家族，强调主动学习、认知负荷和适应学习需求。",
    "limitation": "教学法导向的模型设计不自动等于长期学习增益；仍需要真实课堂与不同学习者群体的效果验证。",
    "tags": [
      "学习科学",
      "个性化",
      "模型评估"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "science-alphafold3",
    "sector": "science",
    "title": "AlphaFold 3：从蛋白质拓展到分子相互作用",
    "organization": "Google DeepMind / Isomorphic Labs",
    "published": "2024-05-08",
    "type": "同行评审论文",
    "url": "https://www.nature.com/articles/s41586-024-07487-w",
    "takeaway": "统一模型预测蛋白质、核酸、小分子等复合物结构；论文报告在多个分子相互作用任务上优于所比较的专用方法。",
    "limitation": "结构预测不等于药物有效、安全或能够成药；仍需实验验证，不能据此推导临床疗效。",
    "tags": [
      "生命科学",
      "结构预测",
      "药物研发"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "design-firefly3",
    "sector": "design",
    "title": "Firefly Image 3 进入 Photoshop 工作流",
    "organization": "Adobe",
    "published": "2024-04-23",
    "type": "产品发布",
    "url": "https://blog.adobe.com/en/publish/2024/04/23/adobe-advances-creative-ideation-with-new-firefly-image-3-model",
    "takeaway": "Firefly Image 3 发布测试版本，可在 Photoshop 测试版及 Firefly 网页应用中使用，强调写实质量、风格和细节控制。",
    "limitation": "质量提升是厂商发布主张；具体品牌一致性和可用率仍需用实际设计任务验证。",
    "tags": [
      "视觉设计",
      "Photoshop",
      "工作流"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "sports-olympic-agenda",
    "sector": "sports",
    "title": "国际奥委会发布 Olympic AI Agenda",
    "organization": "国际奥委会",
    "published": "2024-04-19",
    "type": "战略文件",
    "url": "https://www.olympics.com/ioc/news/ioc-takes-the-lead-for-the-olympic-movement-and-launches-olympic-ai-agenda",
    "takeaway": "议程明确 AI 体育应用的愿景、原则、重点领域和治理机制，并提出国际奥委会的四项承诺，强调以运动员为中心。",
    "limitation": "这是战略与治理安排，不能直接当作各项目已部署或已经产生训练成效的证据。",
    "tags": [
      "奥运",
      "体育治理",
      "运动员"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "sports-tacticai",
    "sector": "sports",
    "title": "TacticAI：为角球提供战术建议",
    "organization": "Google DeepMind / Liverpool FC",
    "published": "2024-03-19",
    "type": "同行评审研究",
    "url": "https://deepmind.google/blog/tacticai-ai-assistant-for-football-tactics/",
    "takeaway": "系统结合预测和生成模型分析角球；在研究的人类专家评估中，相较原有战术布局，专家 90% 的情况下偏好系统建议。",
    "limitation": "90% 是专家偏好率，不是进球率或比赛胜率；研究重点为角球，不代表完整比赛指挥能力。",
    "tags": [
      "足球",
      "战术分析",
      "辅助决策"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "media-sora-research",
    "sector": "media",
    "title": "Sora 研究：最长一分钟的生成视频",
    "organization": "OpenAI",
    "published": "2024-02-15",
    "type": "技术报告",
    "url": "https://openai.com/index/video-generation-models-as-world-simulators/",
    "takeaway": "报告介绍扩散 Transformer 对视频和图像的统一训练，可生成最长一分钟的高清视频，支持不同分辨率和宽高比。",
    "limitation": "报告明确列出物理模拟、因果关系及长期一致性等局限；研究展示不等同于真实世界模拟器已可靠。",
    "tags": [
      "视频生成",
      "扩散模型",
      "一致性"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "science-who-lmm",
    "sector": "science",
    "title": "WHO：医疗多模态模型的四十余项建议",
    "organization": "世界卫生组织",
    "published": "2024-01-18",
    "type": "国际组织指南",
    "url": "https://www.who.int/news/item/18-01-2024-who-releases-ai-ethics-and-governance-guidance-for-large-multi-modal-models",
    "takeaway": "WHO 发布四十余项建议，覆盖诊疗、患者使用、行政事务、医学教育与科研；明确指出错误输出、偏差和自动化偏见的风险。",
    "limitation": "指南不是产品批准，也不意味着模型可替代临床判断；患者信息和输出准确性须有专门保障。",
    "tags": [
      "医疗治理",
      "多模态",
      "患者安全"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "daily-graphcast",
    "sector": "daily",
    "title": "GraphCast：机器学习生成十天天气预报",
    "organization": "Google DeepMind",
    "published": "2023-11-14",
    "type": "研究发布",
    "url": "https://deepmind.google/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/",
    "takeaway": "GraphCast 可在单台 TPU v4 上不到一分钟生成十天预报；在评估的 2,760 个变量与预报时效组合中，89.3% 优于 HRES。",
    "limitation": "这是特定评估设置下的比较，不代表所有天气事件或地区均更准；公共安全决定应依据官方预警。",
    "tags": [
      "天气",
      "公共服务",
      "预测"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "education-unesco",
    "sector": "education",
    "title": "UNESCO：生成式 AI 教育与研究指南",
    "organization": "UNESCO",
    "published": "2023-09-07",
    "type": "国际组织指南",
    "url": "https://www.unesco.org/en/articles/guidance-generative-ai-education-and-research",
    "takeaway": "指南以人为本，建议政府建立政策与能力建设安排，并关注数据隐私和教育机构对工具的适用性验证。",
    "limitation": "这是治理建议，不是某个教学产品提高成绩的实验证据；具体落地需结合当地制度。",
    "tags": [
      "教育治理",
      "隐私",
      "教师"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "robotics-rt2",
    "sector": "robotics",
    "title": "RT-2：把网络知识转成机器人动作",
    "organization": "Google DeepMind",
    "published": "2023-07-28",
    "type": "研究发布",
    "url": "https://deepmind.google/blog/rt-2-new-model-translates-vision-and-language-into-action/",
    "takeaway": "RT-2 将视觉语言模型与机器人控制数据结合，以视觉—语言—动作模型实现指令到动作的转换；团队进行了超过 6,000 次机器人试验。",
    "limitation": "测试中的泛化不代表任意环境下都可靠；物理操作仍受训练动作、设备与安全约束限制。",
    "tags": [
      "具身智能",
      "VLA",
      "泛化"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "design-firefly",
    "sector": "design",
    "title": "Firefly 首发：图像与文字效果生成",
    "organization": "Adobe",
    "published": "2023-03-21",
    "type": "产品发布",
    "url": "https://news.adobe.com/news/news-details/2023/adobe-unveils-firefly-a-family-of-new-creative-generative-ai",
    "takeaway": "首个 Firefly 模型进入测试，聚焦图像和文字效果；Adobe 披露训练来源包括 Adobe Stock、开放许可及版权到期的公有领域内容。",
    "limitation": "商业安全是厂商的设计目标，不构成所有生成物无侵权风险的保证；发布时仍为测试版。",
    "tags": [
      "生成设计",
      "版权",
      "图像"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "education-khanmigo",
    "sector": "education",
    "title": "Khanmigo：GPT-4 教学助手的有限试点",
    "organization": "Khan Academy",
    "published": "2023-03-14",
    "type": "试点发布",
    "url": "https://support.khanacademy.org/hc/en-us/community/posts/13992414612877-Introducing-Khanmigo",
    "takeaway": "Khan Academy 宣布面向有限用户开放实验性 AI 学习向导 Khanmigo，由 GPT-4 支持。",
    "limitation": "试点发布不能证明学习成绩改善；应分别评价解题准确性、引导方式和学生独立完成任务的能力。",
    "tags": [
      "AI辅导",
      "课堂助手",
      "试点"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "daily-nist",
    "sector": "daily",
    "title": "NIST AI RMF 1.0：把风险管理带入部署流程",
    "organization": "美国国家标准与技术研究院",
    "published": "2023-01-26",
    "type": "风险管理框架",
    "url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10",
    "takeaway": "NIST 发布供组织自愿使用的 AI 风险管理框架，帮助设计、开发、部署和使用 AI 的组织管理风险与可信性。",
    "limitation": "自愿框架不是认证或法律合规保证；采用框架仍需形成可检查的控制措施与证据。",
    "tags": [
      "风险管理",
      "公共治理",
      "可信AI"
    ],
    "verified": "2026-11-12"
  },
  {
    "id": "code-copilot-study",
    "sector": "code",
    "title": "Copilot 对照实验：单一编程任务更快",
    "organization": "GitHub",
    "published": "2022-09-07",
    "type": "企业研究",
    "url": "https://github.blog/news-insights/research/research-quantifying-github-copilots-impact-on-developer-productivity-and-happiness-/",
    "takeaway": "在实现 HTTP 服务的受控任务中，使用 Copilot 的参与者完成任务约快 55%。",
    "limitation": "这是特定实验任务，不能直接解释为所有开发项目效率提高 55%；研究来自产品提供方。",
    "tags": [
      "代码补全",
      "生产率",
      "对照实验"
    ],
    "verified": "2026-11-12"
  }
]