-
值得看的一些编程博客
This is one of those “N technical things every programmer must read” lis…
-
「Norm convergence of multiple ergodic averages fo」:数学直觉如何落到可检验的步骤
I’ve just uploaded to the ArXiV my paper “Norm convergence of multiple e…
-
智能体栈里的KV Cache:延迟、成本和翻车点
自回归推理里,历史 token 的 K/V 因果掩码下不变、可以缓存,Q 不行——本文给出可自证的归纳证明,推导显存公式与长上下文并发上限,拆解…
-
我们只招最潮的人:招聘信号失真
An acquaintance of mine, let’s call him Mike, is looking for work after …
-
「Some new pages」笔记:问题从哪来、证明卡在哪
I’ve added some new pages to this blog. Firstly, I’ve moved my old page …
-
智能体栈里的从 logits 到文本:延迟、成本和翻车点
语言模型每一步只输出一组 logits,真正决定输出哪个 token 的是解码策略。本文从 softmax 与 temperature 的几何直…
-
为什么要认真做生产力和速度
A common topic of discussion among my close friends is where the bottlen…
-
「Darwin’s finches and introgressive hybridisation」:数学直觉如何落到可检验的步骤
Last week, as mentioned previously, I attended a very inspiring interdis…
-
别只背名词:Diffusion + Transformer怎么用到生产
DiT 把扩散模型的去噪网络从 U-Net 换成 Transformer,靠的是把图像压成 latent patch token、用 adaLN…
-
读故障复盘能学到什么
I love reading postmortems. They’re educational, but unlike most educati…