{"version":1,"exportedAt":"2026-11-06T19:09:55.306Z","asOf":"2026-11-07","claims":[{"id":"deep-research-analyst","title":"深度研究能交付专业研究员级别的报告吗？","topic":"agents","organization":"OpenAI","date":"2025-02-02","claim":"Deep research 被介绍为可执行多步骤互联网研究、整合大量资料并生成综合报告的能力。","scope":"“研究员级别”没有在此档案中定义为单一可重复指标。","source":{"title":"Introducing deep research","publisher":"OpenAI","date":"2025-02-02","url":"https://openai.com/index/introducing-deep-research/"},"evidence":[{"id":"deep-research-limits","date":"2025-02-02","relation":"context","title":"发布说明保留幻觉与来源判断方面的局限","summary":"官方同日说明提到，深度研究仍可能产生不正确事实，也可能难以区分可靠资料与传闻。","source":{"title":"Deep research — limitations","publisher":"OpenAI","date":"2025-02-02","url":"https://openai.com/index/introducing-deep-research/"}},{"id":"deep-research-expansion","date":"2025-04-24","relation":"supports","title":"后续扩展产品访问范围","summary":"原发布页面 4 月更新宣布轻量版本和更广的套餐访问范围，说明产品持续交付，但并非准确性复测。","source":{"title":"Deep research — April 24 update","publisher":"OpenAI","date":"2025-04-24","url":"https://openai.com/index/introducing-deep-research/"}}],"assessments":[{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"pending","summary":"多步骤研究产品已发布并扩展；专业研究质量需要限定领域、引用核查和实际任务测试。","caveat":"现有证据不足以给“专业研究员级别”作统一判定，保留为待观察。","basisEvidenceIds":["deep-research-limits","deep-research-expansion"],"author":"编辑人工复核"}],"verifiedAt":"2026-11-07","origin":"seed","assessment":{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"pending","summary":"多步骤研究产品已发布并扩展；专业研究质量需要限定领域、引用核查和实际任务测试。","caveat":"现有证据不足以给“专业研究员级别”作统一判定，保留为待观察。","basisEvidenceIds":["deep-research-limits","deep-research-expansion"],"author":"编辑人工复核"},"status":"pending","asOf":"2026-11-07"},{"id":"operator-chatgpt-integration","title":"Operator 会从独立预览并入 ChatGPT 吗？","topic":"agents","organization":"OpenAI","date":"2025-01-23","claim":"Operator 的研究预览提出，浏览器任务能力未来会整合进 ChatGPT。","scope":"核对产品整合这一具体路线，不把发布演示解释为任意任务都能完成。","source":{"title":"Introducing Operator","publisher":"OpenAI","date":"2025-01-23","url":"https://openai.com/index/introducing-operator/"},"evidence":[{"id":"operator-integration","date":"2025-07-17","relation":"supports","title":"官方更新宣布并入 ChatGPT agent","summary":"原发布页面 7 月更新宣布 Operator 已整合进 ChatGPT agent，并说明独立站点将逐步停止使用。","source":{"title":"Operator — July 17 update","publisher":"OpenAI","date":"2025-07-17","url":"https://openai.com/index/introducing-operator/"}}],"assessments":[{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"supported","summary":"官方后续公告明确支持产品整合路线已经交付。","caveat":"整合是产品状态证据，不构成浏览器 Agent 成功率的独立评估。","basisEvidenceIds":["operator-integration"],"author":"编辑人工复核"}],"verifiedAt":"2026-11-07","origin":"seed","assessment":{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"supported","summary":"官方后续公告明确支持产品整合路线已经交付。","caveat":"整合是产品状态证据，不构成浏览器 Agent 成功率的独立评估。","basisEvidenceIds":["operator-integration"],"author":"编辑人工复核"},"status":"supported","asOf":"2026-11-07"},{"id":"chatgpt-search-rollout","title":"ChatGPT Search 会向免费用户开放吗？","topic":"search","organization":"OpenAI","date":"2024-10-31","claim":"ChatGPT Search 发布时面向部分付费用户开放，并计划在之后几个月扩展到免费用户。","scope":"本档案核对开放范围的承诺，不对搜索答案质量作总体判断。","source":{"title":"Introducing ChatGPT search","publisher":"OpenAI","date":"2024-10-31","url":"https://openai.com/index/introducing-chatgpt-search/"},"evidence":[{"id":"search-logged-in","date":"2024-12-16","relation":"supports","title":"扩大到所有登录用户","summary":"原始发布页面的日期更新称，在可用地区向所有登录用户开放搜索。","source":{"title":"ChatGPT search — December 16 update","publisher":"OpenAI","date":"2024-12-16","url":"https://openai.com/index/introducing-chatgpt-search/"}},{"id":"search-no-signup","date":"2025-02-05","relation":"supports","title":"后续公告称无需注册即可使用","summary":"同一页面 2025 年 2 月更新进一步宣布，在 ChatGPT 可用地区，搜索向所有人开放且无需注册。","source":{"title":"ChatGPT search — February 5 update","publisher":"OpenAI","date":"2025-02-05","url":"https://openai.com/index/introducing-chatgpt-search/"}}],"assessments":[{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"supported","summary":"后续官方公告支持开放范围逐步扩大的承诺已经落实。","caveat":"这里依据可核验的发布公告，不代表今天各地区、各账户的实际访问状态。","basisEvidenceIds":["search-logged-in","search-no-signup"],"author":"编辑人工复核"}],"verifiedAt":"2026-11-07","origin":"seed","assessment":{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"supported","summary":"后续官方公告支持开放范围逐步扩大的承诺已经落实。","caveat":"这里依据可核验的发布公告，不代表今天各地区、各账户的实际访问状态。","basisEvidenceIds":["search-logged-in","search-no-signup"],"author":"编辑人工复核"},"status":"supported","asOf":"2026-11-07"},{"id":"claude-computer-use","title":"模型能操作鼠标和键盘，就能稳定完成工作吗？","topic":"agents","organization":"Anthropic","date":"2024-10-22","claim":"Claude 3.5 Sonnet 的 computer use 公测让模型通过截图、光标和虚拟键盘操作计算机。","scope":"公开测试阶段的能力演示，不等于无人监管的完整任务自动化。","source":{"title":"Developing a computer use model","publisher":"Anthropic","date":"2024-10-22","url":"https://www.anthropic.com/news/developing-computer-use"},"evidence":[{"id":"osworld-2024","date":"2024-10-22","relation":"context","title":"同日公布的基准暴露完成率边界","summary":"官方发布说明报告 screenshot-only OSWorld 成绩为 14.9%，并强调滚动、拖拽等动作仍可能出错。","source":{"title":"Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku","publisher":"Anthropic","date":"2024-10-22","url":"https://www.anthropic.com/news/3-5-models-and-computer-use"}}],"assessments":[{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"partial","summary":"计算机操作接口确实被开放测试；同一时期的基准说明其任务完成能力仍有明显边界。","caveat":"该分数仅对应当时模型与设置，不能当作当前版本的成绩。","basisEvidenceIds":["osworld-2024"],"author":"编辑人工复核"}],"verifiedAt":"2026-11-07","origin":"seed","assessment":{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"partial","summary":"计算机操作接口确实被开放测试；同一时期的基准说明其任务完成能力仍有明显边界。","caveat":"该分数仅对应当时模型与设置，不能当作当前版本的成绩。","basisEvidenceIds":["osworld-2024"],"author":"编辑人工复核"},"status":"partial","asOf":"2026-11-07"},{"id":"google-ai-overviews","title":"AI 搜索摘要，能把复杂搜索交给模型吗？","topic":"search","organization":"Google","date":"2024-05-14","claim":"Google 推出扩大范围的 AI Overviews，提出由定制 Gemini 模型帮助用户处理更复杂的搜索与研究问题。","scope":"目标是降低搜索整理成本；摘要的事实性仍需结合来源判断。","source":{"title":"Generative AI in Search: Let Google do the searching for you","publisher":"Google","date":"2024-05-14","url":"https://blog.google/products-and-platforms/products/search/generative-ai-google-search-may-2024/"},"evidence":[{"id":"aio-errors","date":"2024-05-30","relation":"challenges","title":"Google 复盘错误结果并调整触发条件","summary":"Google 承认出现少量不准确或无帮助的摘要，说明讽刺内容、信息稀缺与查询理解问题，并公布技术调整。","source":{"title":"AI Overviews: About last week","publisher":"Google","date":"2024-05-30","url":"https://blog.google/products-and-platforms/products/search/ai-overviews-update-may-2024/"}}],"assessments":[{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"partial","summary":"AI 摘要已进入搜索产品；后续官方复盘表明，复杂问题的答案仍受来源和查询理解影响。","caveat":"这是发布者对问题的说明，不是覆盖所有查询的独立准确率审计。","basisEvidenceIds":["aio-errors"],"author":"编辑人工复核"}],"verifiedAt":"2026-11-07","origin":"seed","assessment":{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"partial","summary":"AI 摘要已进入搜索产品；后续官方复盘表明，复杂问题的答案仍受来源和查询理解影响。","caveat":"这是发布者对问题的说明，不是覆盖所有查询的独立准确率审计。","basisEvidenceIds":["aio-errors"],"author":"编辑人工复核"},"status":"partial","asOf":"2026-11-07"},{"id":"gpt4o-omni","title":"实时多模态模型，会一次开放所有能力吗？","topic":"models","organization":"OpenAI","date":"2024-05-13","claim":"GPT-4o 被介绍为跨文本、视觉和音频的端到端多模态模型，展示低延迟语音交互。","scope":"初始发布同时说明，不同模态能力分阶段开放。","source":{"title":"Hello GPT-4o","publisher":"OpenAI","date":"2024-05-13","url":"https://openai.com/index/hello-gpt-4o/"},"evidence":[{"id":"gpt4o-card","date":"2024-08-08","relation":"context","title":"语音能力有专门评估与缓解措施","summary":"GPT-4o 系统卡补充音频、视觉和文本的能力评估，讨论声音生成等风险与限制，并描述外部评估工作。","source":{"title":"GPT-4o System Card","publisher":"OpenAI","date":"2024-08-08","url":"https://openai.com/index/gpt-4o-system-card/"}}],"assessments":[{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"partial","summary":"官方后续系统卡支持多模态架构和评估工作的存在，开放范围与实际工作流仍需分开考察。","caveat":"这不是对当前版本所有模态可用性或任意任务可靠性的测试。","basisEvidenceIds":["gpt4o-card"],"author":"编辑人工复核"}],"verifiedAt":"2026-11-07","origin":"seed","assessment":{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"partial","summary":"官方后续系统卡支持多模态架构和评估工作的存在，开放范围与实际工作流仍需分开考察。","caveat":"这不是对当前版本所有模态可用性或任意任务可靠性的测试。","basisEvidenceIds":["gpt4o-card"],"author":"编辑人工复核"},"status":"partial","asOf":"2026-11-07"},{"id":"sora-world-simulation","title":"生成视频，离“理解物理世界”还有多远？","topic":"models","organization":"OpenAI","date":"2024-02-15","claim":"Sora 的研究介绍把视频生成视为通往理解、模拟物理世界的模型的一条路径。","scope":"这是研究方向与目标，不是已经实现通用世界模拟的可检验承诺。","source":{"title":"Video generation models as world simulators","publisher":"OpenAI","date":"2024-02-15","url":"https://openai.com/index/video-generation-models-as-world-simulators/"},"evidence":[{"id":"sora-turbo","date":"2024-12-09","relation":"supports","title":"Sora 从研究演示走向产品","summary":"OpenAI 发布 Sora Turbo，向 Plus 和 Pro 用户提供独立视频产品，支持生成最长 20 秒的视频；发布说明仍列出物理与复杂动作方面的局限。","source":{"title":"Sora is here","publisher":"OpenAI","date":"2024-12-09","url":"https://openai.com/index/sora-is-here/"}}],"assessments":[{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"pending","summary":"视频生成功能已形成产品，但“理解物理世界”缺少统一完成标准，继续观察。","caveat":"产品上线不能独立证明实现通用世界模型，也不能因某个失败片段就否定研究方向。","basisEvidenceIds":["sora-turbo"],"author":"编辑人工复核"}],"verifiedAt":"2026-11-07","origin":"seed","assessment":{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"pending","summary":"视频生成功能已形成产品，但“理解物理世界”缺少统一完成标准，继续观察。","caveat":"产品上线不能独立证明实现通用世界模型，也不能因某个失败片段就否定研究方向。","basisEvidenceIds":["sora-turbo"],"author":"编辑人工复核"},"status":"pending","asOf":"2026-11-07"},{"id":"gpt4-benchmarks","title":"考试表现接近人类，意味着可靠通用能力吗？","topic":"models","organization":"OpenAI","date":"2023-03-14","claim":"GPT-4 发布说明称，它在多种专业与学术测试上表现接近人类水平，并报告了模拟律师资格考试等结果。","scope":"发布说明中的测试表现与日常任务可靠性是两个不同问题。","source":{"title":"GPT-4 research","publisher":"OpenAI","date":"2023-03-14","url":"https://openai.com/index/gpt-4-research/"},"evidence":[{"id":"gpt4-report","date":"2023-03-15","relation":"context","title":"技术报告同时记录可靠性局限","summary":"GPT-4 技术报告指出模型仍可能生成不真实内容、犯推理错误；基准分数不能单独证明可靠完成真实工作。","source":{"title":"GPT-4 Technical Report","publisher":"OpenAI / arXiv","date":"2023-03-15","url":"https://arxiv.org/abs/2303.08774"}}],"assessments":[{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"partial","summary":"有官方测试结果支撑特定能力，也有同一技术报告明确列出的局限。","caveat":"当前档案没有独立重跑原始试题；不把发布者报告的分数当作对所有现实任务的保证。","basisEvidenceIds":["gpt4-report"],"author":"编辑人工复核"}],"verifiedAt":"2026-11-07","origin":"seed","assessment":{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"partial","summary":"有官方测试结果支撑特定能力，也有同一技术报告明确列出的局限。","caveat":"当前档案没有独立重跑原始试题；不把发布者报告的分数当作对所有现实任务的保证。","basisEvidenceIds":["gpt4-report"],"author":"编辑人工复核"},"status":"partial","asOf":"2026-11-07"},{"id":"copilot-productivity","title":"AI 编码助手能让开发者快 55% 吗？","topic":"apps","organization":"GitHub / METR","date":"2022-09-07","claim":"GitHub 的受控编程实验报告，使用 Copilot 的参与者更快完成指定任务，平均用时减少约 55%。","scope":"这是特定实验与参与者的结果，不是所有开发任务的通用提升率。","source":{"title":"Quantifying GitHub Copilot’s impact on developer productivity and happiness","publisher":"GitHub","date":"2022-09-07","url":"https://github.blog/news-insights/research/research-quantifying-github-copilots-impact-on-developer-productivity-and-happiness-/"},"evidence":[{"id":"metr-2025","date":"2025-07-10","relation":"context","title":"真实成熟项目中的结果并不相同","summary":"METR 对熟悉成熟开源项目的 16 位开发者、246 个任务进行随机试验，允许使用当时 AI 工具的一组平均耗时增加 19%。","source":{"title":"Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity","publisher":"METR","date":"2025-07-10","url":"https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/"}}],"assessments":[{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"partial","summary":"原实验提供了特定任务上的支持；后续研究提醒我们不能把同一提升率外推到成熟项目。","caveat":"两项研究的人群、任务、工具和时间不同。后一个研究不构成对原始受控实验的直接复现或证伪。","basisEvidenceIds":["metr-2025"],"author":"编辑人工复核"}],"verifiedAt":"2026-11-07","origin":"seed","assessment":{"id":"editor-review-2026-11-07","date":"2026-11-07","status":"partial","summary":"原实验提供了特定任务上的支持；后续研究提醒我们不能把同一提升率外推到成熟项目。","caveat":"两项研究的人群、任务、工具和时间不同。后一个研究不构成对原始受控实验的直接复现或证伪。","basisEvidenceIds":["metr-2025"],"author":"编辑人工复核"},"status":"partial","asOf":"2026-11-07"}]}