用 Jev 把 2000 条评论变成 67 列结构化特征喂给 CatBoost
Jev 把非结构化文本压成 38 个有界问题、29 个评分信号和 67 个数值列，模型 RMSE 从 3.09 降到 1.77。
JEV JUST DID THE F*CKING JOB YOU’VE BEEN PRETENDING YOUR CHATBOT CAN DO FOR TWO YEARS 2,000 reviews → 67 numeric columns → error cut from 3.09 to 1.77 you fed your agent a document it returned another fucking paragraph and you called that automation congrats you automated reading and created nothing Jev took the same ugly text and forced it through 38 bounded questions → 29 graded signals → 9 probability questions → expected values → uncertainty → 67 columns ordinary software could actually use then CatBoost trained on those columns round one reached 1.87 RMSE round five reached 1.77 the mean baseline was 3.09 this is the part everyone keeps missing Jev is not interesting because it writes prettier answers it is interesting because it can turn language into infrastructure text → decisions → columns → thresholds → models support tickets risk reports CRM notes research papers fraud cases all the shit trapped inside prose becomes something your system can filter, join, cache, score and audit stop building agents that talk build systems that decide
这条是引用着另一条写的：x.com/fluixoo/status/2103123493567025321