Sublime
An inspiration engine for ideas
The State
Chatrine Lanot • 1 card
"Buying books would be a good thing if one could also buy the time to read them." –Arthur Schopenhauer https://t.co/08NYlltVow
很多人担心DeepSeek的低成本训练会冲击显卡市场，但我认为其实是利好
首先一个误区是其他厂商模仿DeepSeek就不需要那么多卡了。
其实DeepSeek-R1的低成本训练方法是可以scaling的。也就是说用更多卡，理论上效果只会更好。他本质上是一种improvement of scaling law，可以参考我下面画的不太严谨的示意图。在deepseek出来之前，其他大模型用PRM (process reward model)的时候，已经观察到scaling law失效，边际效应递减了，因为需要额外的卡训练PRM模型来监督推理过程，但是deepseek的出现重新验证了scaling... See more
Earlier in my career when someone was being longwinded and convoluted I'd think "I just don't understand it yet" and now I think "that person just doesn't understand it yet."
Jack Altmanx.comcomputer networks
Rishabh Rawat • 2 cards