Transformers and the self-attention are powerful architectures to enable large language models, but we need a mechanism for them to understand the order of the different tokens we input into the models.