WHAT NEW IN X11.05 THU

X 上的 AI 合集档案 · 23 个合集 · 69,408 条

grokVIDEO

一页论文:Grok Bot 如何决定自主行动还是先问

用「能否在午饭前撤销」这一个分数把任务分流到 act/draft/ask/stop 四道门,并限制每天三次询问,四周后询问率从 0.61 降到 0.09。

the0xbt
@the0xbt22 天前收录原推发布于 2026.10.13
the0xbt @the0xbt原推发布于 2026.10.13
在 X 上打开原推

This is f**king insane... A one-page paper on how a single Grok Bot earns the right to act without you, and the whole method fits above the fold. The bot never asks how important a task is. It asks one thing, can this be undone by lunch, and that number picks one of four doors: act, draft, ask, stop. A rename scores 0.98 and simply happens. A $2,400 wire scores 0.04 and waits behind ask. Anything under 0.05 never even gets asked about. The part almost nobody builds: three asks a day, no rollover. When they run out the bot does not fall silent, it writes the work up as a draft and leaves it on the desk. And every yes gets appended to the rules file as one line. Six lines on day 1, thirty-one on day 28, twenty-five of them born from a yes. The same question never returns. Ask rate across four weeks on the same budget: 0.61 -> 0.34 -> 0.18 -> 0.09. Sixty tasks, four doors, zero wrong. Paper is below. The routing config is inside it.

这条是引用着另一条写的:x.com/rich_odinn/status/2109563714068169203

x.com/the0xbt/status/2109958524520337705采集于 2026.10.14 10:35