wassname comments on How useful is mechanistic interpretability?

[wassname]

I also think current methods are mostly focused on linear interp. But what if it just ain’t linear.