WHAT NEW IN X10.29 THU

X 上的 AI 合集档案 · 22 个合集 · 60,306 条

jevVIDEO

Jev 1.13 承接 agent 里的 yes/no 分支:百万 token 输入 $0.042

把「下一个谁干、这条来源相关吗、草稿能发吗」这类小判断从全模型调用挪到 Jev 的 Choice/Score/Noul 层,并说明决策与执行代码的边界。

Dark Factor
@DarkFactorr5 天前收录原推发布于 2026.10.24
Dark Factor @DarkFactorr原推发布于 2026.10.24
在 X 上打开原推

this is f*cking insane, ur agent pays frontier prices to answer yes or no. Jev 1.13 answers the same fork for $0.042 per million tokens and charges nothing for output every agent hits the same small questions all day, which worker goes next, is this source relevant, is the draft ready, and each one becomes a full model call before any real work starts Jev by TypeSafe pulls those questions into their own layer: > state: the task, what's finished and which workers are free right now > question: Choice to route, Score to rate, Noul for a straight yes or no > threshold: anything below ur confidence line drops to review > code: executes the answer and checks it actually happened • what one fork costs > 10,000 decisions at 1,000 input tokens each land at $0.42 > Browser Use found flights in 7 seconds for $0.0039 > Vercel's team reported safety classification 5 to 18x faster than their previous model • where Jev stops and code takes over > the menu gets rebuilt after every action, never yesterday's options > questions that share the same state get sent together, not one by one > a DONE with high confidence still needs code to confirm the file exists and the part that broke my brain: that $0.0039 flight run burned 90,558 Jev input tokens it found the flights, it never booked a single ticket the cheap part was never the hard part the hard part is knowing exactly where the decision ends and the work begins ↓

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/darkfactorr/status/2113714777226690905采集于 2026.10.24 09:57