{"version":"2026-10-19.observations.2","reading":{"id":"reading-jobbench","title":"AI会做，不等于你已经省下工作","deck":"JobBench把委派偏好带入专业任务选题；能力、人的意愿、实际省时和岗位变化仍需分别核对。","editorialAt":"2026-10-19","sourceIds":["R-JOBBENCH","R-WORKBANK","R-WORKPATTERNS","R-TAOBAO","R-JB-EVAL","R-JB-SCORES"],"sections":[{"title":"JobBench测到了什么","paragraphs":["JobBench把专业工作打包为任务与多种参考文件，要求处理冲突材料并提交可检查产物。评分链中的关键条件要一起满足才计相应权重；流畅的文字不能替代证据推理。这里衡量的是指定环境中的交付能力。"],"sourceIds":["R-JOBBENCH"]},{"title":"想委派什么，要由本人和情境决定","paragraphs":["WORKBank研究美国职业任务的委派偏好及希望保留的人类参与程度。JobBench借用这类偏好来选题，没有在每次基准运行中测量真实员工的满意度或采用效果。","对自己的工作，先说清愿意交出的步骤、想保留的判断和原因；未填写表示未知。这是编辑建议，不把美国职业平均偏好自动填成中国用户或个人意愿。"],"sourceIds":["R-WORKBANK","R-JOBBENCH"]},{"title":"从能力到就业，四层证据各回答一件事","paragraphs":["任务能力：产物能否在明确模型、工具、材料与评分条件下合格？个人效果：加入准备、监督、审核、返工后，人是否真的省时？组织效果：合格交付、客户结果和协调流程是否改变？劳动市场效果：招聘、工资和岗位结构为何变化？后一层不能由前一层的分数替代。","跨企业随机实验的公开摘要报告，使用工具者的邮件活动耗时下降，却没有检测到任务数量或构成变化。淘宝客服随机实验摘要报告，聊天更快与AI可处理聊天的客户评分下降并存。它们说明应同时检查工作量、质量和接管；不据此给本团队承诺收益。"],"sourceIds":["R-WORKPATTERNS","R-TAOBAO"]},{"title":"先查协议，再读分数","paragraphs":["论文v1列main/easy各65项任务，固定执行器README列65/63；本站本轮未独立清点HF文件，按文档各自版本保留。更关键的是：执行器要求逐题归一化后等权平均，而官网固定分数说明采用全部得分除以全部满分。这是尚未解决的聚合公式冲突，不能假定只是版本更新。","两种公式在每题满分不同时会产生不同结果。因此本页只解释方法，不发布可比增长率，也不把外部加权成绩映射成工作自动化或岗位替代比例。"],"sourceIds":["R-JOBBENCH","R-JB-EVAL","R-JB-SCORES"]},{"title":"把方法带回多文件分析","paragraphs":["编辑构造例子，未执行AI：经营报告同时含年度与季度材料。先要求来源定位、期间与单位表、冲突清单，再让人核验口径与结论。与F01的财报分析只有部分步骤相交，不是JobBench对F01的直接验证。","净节省人时＝人工基线H0−新流程全部人时H1；净省时率＝1−H1/H0。工作量与质量应一致，缺数据不估计，H0为零时比例不可计算，负收益保留。C03/F01已有完整人时字段可继续使用，不把单次生成时间换算为编制。"],"sourceIds":[]}],"actions":["求职者：准备带来源核对、复算过程和本人贡献说明的作品；有岗位材料后再核查要求。","团队负责人：先列愿意委派与保留的步骤，用已有记录检查全部人时和质量；暂不改变分工也合理。","普通读者：分别复述能力、个人效果、组织效果、劳动市场四层；不知道的保持未知。"],"seriesIds":[],"taskIds":["F01","C03"],"frontierIds":["F20","F08","F24"],"evidenceStatus":"active","actualCoverage":null},"sources":[{"id":"R-JOBBENCH","title":"JobBench：专业证据推理与委派偏好","url":"https://arxiv.org/html/2605.26329v1","version":"v1","locator":"§2.1–2.3；附录E；§3.1","scope":"选题与链式评分方法","license":"仅链接与有限摘要，不再分发论文全文或任务包","limit":"偏好用于选题；基准没有测量员工采用、满意度或净省时。","publishedAt":"2026-05-25","readAt":"2026-10-19","collectedAt":"2026-10-19","status":"active","nextReview":"2026-11-19","methods":[],"family":"JobBench"},{"id":"R-WORKBANK","title":"WORKBank：员工希望保留多少参与","url":"https://arxiv.org/abs/2506.06576v3","version":"v3","locator":"摘要与版本历史","scope":"美国职业任务偏好与人类参与程度","license":"仅链接与有限摘要，不再分发论文全文或任务包","limit":"本轮阅读摘要；不是中国用户或当前用户的意愿调查，不是实际采用率。","publishedAt":"2025-06-06","readAt":"2026-10-19","collectedAt":"2026-10-19","status":"active","nextReview":"2026-11-19","methods":[],"family":"R-WORKBANK","revisedAt":"2026-02-01"},{"id":"R-WORKPATTERNS","title":"跨企业实验：个人活动省时与任务组合","url":"https://www.aeaweb.org/articles?id=10.1257/aeri.20250275","version":"AEA forthcoming摘要；2026-10-19读取","locator":"Abstract","scope":"随机提供生成式AI访问的知识工作者实验","license":"仅链接与有限摘要，不再分发论文全文或任务包","limit":"本轮仅核对公开摘要，不声称已阅读会员全文；个人活动省时不等于组织总产出增加。","publishedAt":null,"readAt":"2026-10-19","collectedAt":"2026-10-19","status":"active","nextReview":"2026-11-19","methods":[],"family":"R-WORKPATTERNS"},{"id":"R-TAOBAO","title":"淘宝客服实验：速度、质量与接管","url":"https://arxiv.org/abs/2605.14830v2","version":"v2","locator":"Abstract；版本历史","scope":"客服Agent与人工监督的随机现场实验","license":"仅链接与有限摘要，不再分发论文全文或任务包","limit":"本轮阅读摘要，未独立复现；结论限该客服流程，不推算软件或金融岗位收益。","publishedAt":"2026-05-14","readAt":"2026-10-19","collectedAt":"2026-10-19","status":"active","nextReview":"2026-11-19","methods":[],"family":"R-TAOBAO","revisedAt":"2026-06-01"},{"id":"R-JB-EVAL","title":"JobBench执行器：逐题等权汇总","url":"https://github.com/Job-Bench/job-bench-eval/blob/660b33a86ed032559df0b2fdc1ad38cf7ceb83f3/README.md","version":"660b33a86ed032559df0b2fdc1ad38cf7ceb83f3","locator":"Scoring；dataset说明","scope":"固定代码文档中的评分公式与任务数","license":"仅链接与有限摘要，不再分发论文全文或任务包","limit":"与官网固定分数说明存在未解决的聚合口径冲突；未复跑或独立清点HF数据文件。","publishedAt":null,"readAt":"2026-10-19","collectedAt":"2026-10-19","status":"active","nextReview":"2026-11-19","methods":[],"family":"JobBench"},{"id":"R-JB-SCORES","title":"JobBench官网：总点数汇总说明","url":"https://github.com/Job-Bench/Job-Bench.github.io/blob/828b1607c6fdb389ee4762f50d1fd675a40912e9/docs/score-updates/2026-09-15.md","version":"828b1607c6fdb389ee4762f50d1fd675a40912e9","locator":"Aggregation；2026-09-15分数更新","scope":"特定运行的按总点数聚合记录","license":"仅链接与有限摘要，不再分发论文全文或任务包","limit":"不代表所有历史版本均使用此公式；与执行器README冲突未澄清，不连接成趋势。","publishedAt":"2026-09-15","readAt":"2026-10-19","collectedAt":"2026-10-19","status":"active","nextReview":"2026-11-19","methods":[],"family":"JobBench"}],"actualCoverage":null}