yes, since you can learn to represent the same problem with less amount of params, however most of the architectures are optimized for the linear product, so we gotta figure out a new architecture for it