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。

通过将一个大问题拆成五个小问题,准确率从 62.6% 提升到 95.0%;Jev 本身不生成文本,而是在给定候选项中挑选,拆分提问可超越 Haiku 4.5。
ONE BIG QUESTION: 62.6% ACCURATE. FIVE SMALL ONES: 95.0%. SAME MODEL. the whole lesson of this article is in that jump. jev doesn't write. it only picks from answers you allow. ask it one vague question per email and it guesses. split it into 5 small ones and it beats haiku 4.5