Table 1: Perplexity of 355 million parameter models trained for 10 billion tokens on the Pile. Yet, some subquadratic gated-convolutions match attention on the non AR slice! Can we capture the strengths of both gated convolutions and attention in one purely sub-quadratic architecture? We find the AR gap is because gated convolution models (e.g. Hyena, H3, RWKV, RetNet) need model dimension that sc