Jev 与文本生成模型的取舍
Jev 不生成自由文本。Diogo 用一段对比视频说明并行决策模型与逐字生成模型在任务形态上的差别。

作者将批量问答扩大至每批约300个问题,Jev成功响应,而Gemini在生成前拒绝了全部请求。输入仅约5.5万token,上下文窗口约1M token,问题在于Gemini执行结构化输出的方式。
5/7 Then I widened the batch to roughly 300 questions per call. ✅ Jev answered. ❌ Gemini rejected all 300 test requests before generating anything. The input was only ~55k tokens in a ~1M-token context window. The problem was how Gemini enforces structured output using a