WHAT NEW IN X10.28 WED

X 上的 AI 合集档案 · 22 个合集 · 58,924 条

jevPOST

Puffle 团队实测后把 Jev 从 harness 里拆掉

作者认可决策模型能当 agent 的侦察兵省成本、也能提升 RAG 精度,但实测 Jev 判断经常出错把 agent 带偏,最后换回 qwen 3.7 flash。

Kush
@kushbhuwalka2 天前收录原推发布于 2026.10.26
Kush @kushbhuwalka原推发布于 2026.10.26
在 X 上打开原推

beware of jev - we had to rip it out of the @puffleai harness i do believe decision models will be big for harnesses 1) they can act like scouts for your agent, checking if a file is worth reading, massively saving cost and latency. 2) they solve precision in RAG so when jev came out, I stuck it in immediately. but jev constantly makes wrong assessments which leads the agent astray. its just not smart enough. we're early on decision models. we need decision models that can reason like a small LLM can. back to qwen 3.7 flash

x.com/kushbhuwalka/status/2114464959631614055采集于 2026.10.26 09:59