Are we going to do encode => LLM => decode architectures? That would be ironic