WHAT NEW IN X10.30 FRI

X 上的 AI 合集档案 · 22 个合集 · 60,637 条

jevVIDEO

Jev 三头信号织机:Choice、Score、Yes/No 合一

用 Jev 的 Choice、Score 和是/否三个类型化输出头合并成单一状态,低于 0.85 置信度直接转人工,代码按门禁结果确定性执行。

Flandermaxx
@Flandermaxx9 天前收录原推发布于 2026.10.20
Flandermaxx @Flandermaxx原推发布于 2026.10.20
在 X 上打开原推

WHOEVER LEAKED THIS JEV SIGNAL LOOM SHOULD BE WRITING THE WHITE PAPER I thought asking an agent a question meant prompting the model, waiting for a sentence, and parsing it to figure out what it just said. Apparently not. This loom's setup is much more interesting: → CHOICE head returns one label from an ordered set in 3ms, nothing else allowed, no free-form string on retry → SCORE head ranks the same set by criterion, one score per item, no "explain the ranking" prose anywhere → NOUL head answers yes/no in 4ms with a confidence, below the gate floor drops straight to human review → ONE STATE merges the three heads at 1,257 answers per second, 60 bundles 600 threads, no raw text crosses the gate → CODE acts on the gated state deterministically, flags 5 action drifts above the baseline the author never caught so you get: Choices labeled. Scores ranked. Nouls floored. Actions deterministic. the interesting part isn't making the model smarter. it's giving every question a head whose output type the code can trust. Instead of parsing one freeform answer per decision at Claude prices, you weave three typed heads at $0.42 per 10,000 decisions, 256ms median, and let the gate drop anything under 0.85. Weave the heads. Gate the state. Act in code.

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/2112279612340232483采集于 2026.10.20 10:06