Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
https://rikiyat.github.io/blog/weight-space-programming/
好多人说这个思路overhyped了,好巧不巧我刚好有一个大概0.05b的模型有些边边角角的问,结果就真的能通过原地修改某个head的key projection把问给解决了,解决了。。。
@aigc1024