An alternate title for this post might be, “Twitter has a kernel team!?”…

On Thursday, Charlie Fefferman continued his lecture series on interpola…

单头 attention 只有一组 softmax 权重，只能在一种相似度度量下做一次聚合。Multi-Head Attention 通过多套独…

机器人商业化量产落地，核心不在于极致的单机技术优化，而在于供应链降本、无改造部署、场景复用、数据迭代四大核心维度的平衡。摒弃定制化研发、前置基建…

On large platforms, it’s impossible to have policies on things like mode…

Today, Charlie wrapped up several loose ends in his lectures, including …

> 本文从零推导注意力机制点积方差的来源，解释缩放因子如何防范梯度弥散，并作为大模型 Scaling Laws 数值稳定的基石。

HN comments are terrible . On any topic I’m informed about , the vast ma…

I’m continuing my series of articles for the Princeton Companion to Math…

从 cross-attention 到 self-attention 的退化路径 → 为什么 self-attention 是 O(1) 跳数 …