← 返回列表

@teortaxesTex: They've recreated DeepSeek-Optical Contexts Compression from first principles… to be clear, these "tokens" are, rather, KV slots. In disk sp...

@teortaxesTex 3 信息等级 3 1 噪音/剔除；2 较弱；3 普通事实；4 重要行业动态；5 极重大事件。该分数是信息显著性，不是投资建议。 发布：2026-06-12T08:21 抓取：2026-06-12 11:19
🔗 原文链接
AI 算力
摘要

推文称有人从第一性原理复现了DeepSeek的光学上下文压缩技术，指出其令牌实际上是KV槽，磁盘空间仍大于文本和渲染图像，故该技术有效是因为多数缓存设计臃肿。

客观事实
  • DeepSeek光学上下文压缩技术被从第一性原理复现
  • 该技术的令牌实际上是KV槽
  • 磁盘空间仍大于文本和渲染图像，有效源于缓存设计臃肿
DeepSeek

原文

They've recreated DeepSeek-Optical Contexts Compression from first principles…
to be clear, these "tokens" are, rather, KV slots. In disk space it's still larger than both the text and even the rendered image. So the only reason this works is that most cache designs are BLOATED https://t.co/clpsLC2SXW

likes: 21 | retweets: 1 | replies: 0 | views: 9819