WHAT NEW IN X10.28 WED

X 上的 AI 合集档案 · 22 个合集 · 58,925 条

jevVIDEO

Jev 决策网格:一次调用批量回答 5 个问题

用 Jev 把入站状态打包成类型化记录,一次调用并行返回 5 个 Choice/Score 答案,再按 0.85 置信度逐条门控,10,000 状态 $0.42、中位 256ms。

Flandermaxx
@Flandermaxx6 天前收录原推发布于 2026.10.22
Flandermaxx @Flandermaxx原推发布于 2026.10.22
在 X 上打开原推

WHOEVER LEAKED THIS JEV DECISION GRID SHOULD BE SETTING THE STACK I thought asking an agent five questions meant five prompts, five responses, five chances to drift. Apparently not. This grid's setup is much more interesting: → STATE packs the inbound case into one typed record, 15 walls and 600 cards indexed, no raw text crosses the boundary → QUESTIONS bundles 5 typed asks against the same state (Choice A, Choice B, Score A, Score B, Noul), no five round-trips, one call → JEV returns all 5 answers in parallel at 256ms median, no waiting on the first to compare against the second, no re-reading the state five times → GATE compares each answer's confidence against 0.85 independently, drops the ones below to review, forwards the rest to code → CODE acts on the batch per-answer deterministically, flags 6 cases where confidence tanked on question B while question A still passed clean so you get: States packed. Questions batched. Answers parallel. Gates independent. the interesting part isn't asking fewer questions. it's giving every state a grid where five answers come back in one shot. Instead of running 5 separate model calls at 5 round-trips each, you ask one batched decision at $0.42 per 10,000 states, 256ms median, output free, and the gate scores each answer on its own confidence. Pack the state. Batch the questions. Gate each answer.

这条是引用着另一条写的:「长文:如何搭好 Jev 驱动的 agent 技术栈」

x.com/flandermaxx/status/2112989044573434101采集于 2026.10.22 10:07