"A self-attention model is a linear MLP with dynamic weights" sounds about right

→ View original post on X — @rasbt,