{"version":"2026-10-19.observations.2","reading":{"id":"reading-measures","title":"曲线突然跃升，意味着工作能力突然出现吗？","deck":"先检查分母、评分量尺与模型配置，再理解增长和预测。","editorialAt":"2026-10-15","sourceIds":["D-RUAN","D-MIRAGE","D-WEI","C-EPOCH"],"sections":[{"title":"同一表现会被两种评分画成不同曲线","paragraphs":["人工构造例：五个字符中答对四个，完全匹配规则给0，逐字符评分给80%。这是说明量尺的假设示例，不是AI实跑。Schaeffer等展示离散评分可形成表观跃迁；Wei等的涌现定义讨论小模型未显现而大模型显现的任务表现。两者提醒检查度量，并不意味着所有突破都不存在。"],"sourceIds":["D-MIRAGE","D-WEI"]},{"title":"相关分数可以帮助预测什么","paragraphs":["Ruan等用跨模型评测成绩的低维表示预测下游测试，并在留出模型上检验。它是观察性成绩预测，没有随机改变模型训练或证明某个用户工作必然可完成。论文>10%指较弱训练模型与留出较强模型的成绩跨度，不是“预测误差超过10%”。","人工构造的阅读例：用若干代码、常识和数学测试成绩作为输入，预测另一项测试的成绩，再让新模型实际做该测试来核对。它估的是测试表现，不是某个未来日期或完整岗位能力。"],"sourceIds":["D-RUAN"]},{"title":"模型日横截面不是历史实测","paragraphs":["ECI当前快照综合多种测试；X轴按模型发布日展开，只表示该模型在当前文件里的位置。运行日、成绩首次公开日与所用预算不齐全，所以没有把点连成每日能力历史，也不推算行业增长率或倍增时间。不同配置和区间重叠仍要核对。"],"sourceIds":["C-EPOCH"]},{"title":"怎样用于一年、三年的判断","paragraphs":["可以明确写出条件：某个固定任务和协议持续改善且独立验收通过，才扩大委派；审查/返工代价增加则收缩。用相关分数或消息份额替代实际通过和采用数据，会把预测做得过于精确。具体工作比例继续保持未知，条件变化进入场景复核。"],"sourceIds":[]},{"title":"把未来条件变成学习检查点（编辑建议）","paragraphs":["以周报为人工构造学习例：现在先练每个数字都有出处、期间与单位，交一份能用程序复算的结果和来源对账表。","一年复核点：用未见过的跨文件材料练口径统一、独立验收和失败接管；只有留出材料在相同质量与预算下持续通过，才扩大草稿范围。","三年复核点：练版本追溯、已撤销事实与异常处理，比较跨季度的完整人时；返工增加或接收失败就收回委派。时间是重新检查的节点，不是到期必然获得能力的预测。"],"sourceIds":[]}],"actions":["先看评分是否要求整题/局部","区分模型日、观察期、成绩公开日","把未来判断写成可被证据推翻的条件"],"seriesIds":["epoch-eci","aime-answers"],"taskIds":["C03","M01"],"frontierIds":["F07","F08","F24"],"evidenceStatus":"active","actualCoverage":null},"sources":[{"id":"D-RUAN","title":"观察性缩放与成绩预测","url":"https://proceedings.neurips.cc/paper_files/paper/2024/file/1cded4f97cf5f01a284c574110b7e3b9-Paper-Conference.pdf","version":"NeurIPS2024正式论文","locator":"§3–4；§4.2留出模型预测","scope":"跨模型测试分数的低维表示及下游成绩预测","license":"仅解释和引用；不复制论文或完整模型表","limit":"观察性预测不是训练干预或具体任务通过证据；>10%指训练与留出模型的成绩跨度，不是预测误差","publishedAt":null,"readAt":"2026-10-15","collectedAt":null,"status":"active","nextReview":"2027-01-13","methods":[],"family":"D-RUAN"},{"id":"D-MIRAGE","title":"离散评分如何制造表观跃迁","url":"https://arxiv.org/html/2304.15004v2","version":"arXiv v2（2023-05-22引用版本，非首稿日）","locator":"Figures1/5/6；离散与连续评分比较","scope":"评测量尺如何改变可见曲线","license":"原论文CC BY-NC-ND 4.0；本站仅独立概念/事实说明与原文链接，不再发布论文全文","limit":"部分量尺的反例，不证明全部能力突破都是假象","publishedAt":"2023-05-22","readAt":"2026-10-15","collectedAt":null,"status":"active","nextReview":"2027-01-13","methods":[],"family":"D-MIRAGE"},{"id":"D-WEI","title":"涌现能力的操作性定义","url":"https://arxiv.org/html/2206.07682v2","version":"arXiv v2（2022-10-26引用版本，非首稿日）","locator":"§2；Table1/Figures1–3","scope":"小模型未显现、大模型显现的任务表现及度量条件","license":"原论文CC BY 4.0；本站仅有限事实摘要与原文链接","limit":"操作性定义与指定评测，不能直接预测工作自动化临界点","publishedAt":"2022-10-26","readAt":"2026-10-15","collectedAt":null,"status":"active","nextReview":"2027-01-13","methods":[],"family":"D-WEI"},{"id":"C-EPOCH","title":"Epoch ECI固定快照","url":"https://epoch.ai/data/eci-documentation","version":"2026-10-14取得的CSV；ECI与FrontierCode独立指标","locator":"eci_scores.csv；processed_data_for_eci.csv指定benchmark行","scope":"跨测试拟合的综合指数；组件保留模型版本，可能取同名同日不同配置的最高成绩","license":"Epoch数据CC BY 4.0（当前官方链接核对），署名Epoch AI","limit":"模型日不等于测量日；本次快照不是逐日历史；组件0–1分不称工作通过率","publishedAt":null,"readAt":"2026-10-15","collectedAt":"2026-10-14","status":"active","nextReview":"2027-01-13","methods":["https://epoch.ai/data/eci-documentation/methodology"],"family":"C-EPOCH"}],"actualCoverage":null}