Kloss(@kloss_xyz)

this is pretty insane. anyone who's scrubbed through hours of footage looking for the one frame where something broke knows AI still can't really watch video for you imagine doing a jigsaw puzzle with 500 pieces after glancing at the box for two seconds and never being allowed to look at it again… that's roughly how most AI models handle an image. one look, convert it to words, reason off that this visual model keeps looking… zooms into the part that matters, pulls the exact frame out of the video, rewa

6.6内容质量
this is pretty insane. anyone who's scrubbed through hours of footage looking for the one frame where something broke knows AI still can't really watch video for you imagine doing a jigsaw puzzle with 500 pieces after glancing at the box for two seconds and never being allowed to look at it again… that's roughly how most AI models handle an image. one look, convert it to words, reason off that this visual model keeps looking… zooms into the part that matters, pulls the exact frame out of the video, rewa

TL;DR · AI 摘要

klöss on X: "this is pretty insane. anyone who's scrubbed through hours of footage looking for the one frame where somet...

核心要点

  • 主题聚焦：this is pretty insane. anyone who's scrubbed thr
  • 来源：Kloss(@kloss_xyz)，建议结合原文判断细节。
  • AI 分析暂不可用，本条为保底评分与摘要。
#AI
打开原文

klöss on X: "this is pretty insane. anyone who's scrubbed through hours of footage looking for the one frame where something broke knows AI still can't really watch video for you imagine doing a jigsaw puzzle with 500 pieces after glancing at the box for two seconds and never being allowed to look… / X

klöss

@kloss_xyz

这简直令人难以置信。任何花数小时浏览视频寻找某个关键帧的人，都清楚AI目前还无法真正帮人观看视频。想象一下，你只瞥了眼装有500块拼图的盒子两秒钟，然后被要求在完全不能再次查看盒子的情况下完成拼图……这基本上就是大多数AI模型处理图像的方式。它们匆匆一瞥，将图像转换为文字，然后基于这些文字进行推理。而这个视觉模型却不同，它会持续关注视频中关键的部分，从视频中精准提取出相关帧，重新观看不确定的片段，然后再继续思考。从机器人、工地、工厂流水线、游戏回放，甚至到在家清点宠物……这些场景都蕴含着大量AI尚未涉足的视觉应用场景，而现在，AI正在学习如何真正"观看"。

$

00:00

/$

Andrew M. Dai

@AndrewDai

6h

多年来目睹模型在静态图像上表现优异，却对任何运动或角度变化的物体完全失效。今天，我们正在解决这个问题。现在尝试Glimpse，我们的首个视觉推理模型，在关键公共视觉推理任务上已达到与Astra相当的前沿性能水平。

显示更多

付费合作

5:30 PM · Oct 16, 2026

·

1,361

浏览量

5

20

4