You mentioned using a specific transformer architecture in your last project; why was that better than a standard RNN for that use case?
