I remember seeing some theoretical analysis that compared computing differences between transformers, LSTMs and RNNs and I think that RNNs are theoretically bet

I remember seeing some theoretical analysis that compared computing differences between transformers, LSTMs and RNNs and I think that RNNs are theoretically better (can learn more complex functions). Can't find it now.