Why did transformers replace RNNs for most NLP tasks?

Self-attention lets every token weigh every other token in one parallel step, instead of pushing information down a recurrent chain, so long-range dependencies survive and training parallelises. The cost is attention quadratic in sequence length.