-
把「The cubic nonlinear Schrödinger equation in two」写成可执行的阅读路径
I’ve just uploaded to the arXiv the paper “The cubic nonlinear Schröding…
-
别只背名词:多模态融合怎么用到生产
CLIP 靠 batch 负样本和温度做对比学习,不是拼 caption;projector、Flamingo cross-attention、…
-
把 Verilog 写安全有多难
Troll? That’s how people write Verilog 1 . At my old company, we had a t…
-
「Dyadic models」:数学直觉如何落到可检验的步骤
One of the oldest and most fundamental concepts in mathematics is the li…
-
智能体栈里的ViT:延迟、成本和翻车点
ViT 把图像切成 patch 再当 token,但 patch size 和分辨率一旦定下来,序列长度、attention 的显存和算力也同时…
-
编程语言怎么影响 token 效率和正确性
This somewhat widely cited post (I keep seeing it cited, anyway) suggest…
-
「2006 ICM: Étienne Ghys, “Knots and dynamics”」笔记:问题从哪来、证明卡在哪
Almost a year ago today, I was in Madrid attending the 2006 Internationa…
-
智能体栈里的MoE:延迟、成本和翻车点
MoE 把参数量和每 token 计算量解耦,但代价从模型结构转移到了路由:过载丢 token、专家塌缩、路由抖动是训练现场的三种症状,All-…
-
采样还是追踪:性能数据两套方法
Perf is probably the most widely used general purpose performance debugg…
-
「(Emmanuel Kowalski) The large sieve inequalities」笔记:问题从哪来、证明卡在哪
[This post is authored by Emmanuel Kowalski.] This post may be seen as c…