-
P95 并没有你想的那么强
Reaching 95%-ile isn’t very impressive because it’s not that hard to do.…
-
把「“Math Doesn’t Suck”, and the Chayes-McKellar-Win」写成可执行的阅读路径
As you may already know, Danica McKellar, the actress and UCLA mathemati…
-
别只背名词:T5怎么用到生产
T5 把翻译、分类、摘要、问答都改写成统一的文本到文本任务,但这个接口不是免费的。本文拆开 span corruption 的 corrupti…
-
Overwatch 性别随机试验里真正测到了什么
A recurring discussion in Overwatch (as well as other online games) is w…
-
读懂「Random matrices: the circular law」:先把定义、反例和适用范围钉死
Van Vu and I have recently uploaded our joint paper, “Random matrices: t…
-
别只背名词:GPT 系列怎么用到生产
GPT-1 到 GPT-4 四次跃迁里,next-token prediction 这个目标函数从未换过,真正变化的是任务接口、参数规模和对齐机…
-
为什么很难买到真正好用的东西
There’s a cocktail party version of the efficient markets hypothesis I f…
-
「Printer-friendly CSS, and nonfirstorderisability」:数学直觉如何落到可检验的步骤
I recently discovered a CSS hack which automatically makes wordpress pag…
-
智能体栈里的BERT:延迟、成本和翻车点
BERT 的每个设计选择都有可核对的实测代价。本文钉住 BERT 这一条路线本身:80/10/10 掩码策略真正缓解了多少 pretrain-f…
-
80 年代以来 CPU 真正多了什么
This is a response to the following question from David Albert: