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。

该帖称赞 TypeSafeAI 的 Jev 通过评分过滤历史记录而非摘要来压缩上下文,可将 156,000 个 token 一次性减少到 62,000 个,并希望 Claude Code 和 Codex 也能采用这一思路。
This usage is really great. I hope it's implemented in Claude Code and Codex. > Claude Code's context compression, without summarization, reduces 156,000 to 62,000 tokens in one shot. TypeSafeAI's "Jev" The idea of filtering history by scores without summarizing is great