“A fundamental mathematical equivalence exists between linear attention (attention without the softmax function), linear Recurrent Neural Networks, and state-space models (SSMs).”