I have just uploaded to the arXiv my paper “A quantitative formulation o…

把 Transformer block 里那个看起来最不起眼的两层 MLP 真正讲清楚——4 倍扩张比的来历、逐位置而不是跨位置的设计、Geva…

It’s common to see people advocate for learning skills that they have or…

In one of my recent posts, I used the Jordan normal form for a matrix in…

从公式到工程把 LayerNorm 讲清楚：它在每个 token 内部如何做归一化，为什么比 BatchNorm 更适合变长序列和自回归训练，p…

2025年火爆全球的生成式AI与人形机器人两大赛道，看似是AI技术迭代的两大终极方向，共享深度学习、强化学习的技术基底与千亿级资本红利，但二者的…

This is a pseudo-transcript (actual words modified to be more readable t…

As you may have noticed, I had turned off the “Snap preview” feature on …

从 ResNet 的核心思想出发，讲清 Transformer 里残差连接真正解决的不是“信息保留”这种空泛说法，而是优化路径、梯度传播和迭代修…

Which is faster, keyboard or mouse? A large number of programmers believ…