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。

作者认为专为可监控性设计的 Jev 模型在大规模场景下可能表现良好,但对 CoT 方法存有顾虑,因为 GPT-6 可能规避监控。
Overall it's super fast! A specialized Jev model designed specifically for monitorability could work really well at scale! I have my reservations of CoT as a method due to GPT-6 evading monitors, but it could prevent more obvious harm.