How strong is the NLP technology of Recurrent.ai?

Forcibly answer a wave, taken from Recurrent.ai Zhihu:

20 19, 1 June, 2006, Yang, co-founder of Recurrent.ai, and Google Brain developed the "TransformerXL Model" with international significance. Compared with traditional RNN and standard transformer, its effective modeling length is increased by 80% and 450% respectively, and the speed of transformer model in testing stage is increased by more than 1800 times.

20 19 June, Yang, co-founder of Recurrent.ai, and Google's collaborators proposed a new pre-training language model XLNet. This model has surpassed the performance of BERT model, which once kept the best performance record in 20 tasks, and reached the most advanced level in 18 tasks.

This is a process of constantly refreshing your own world record. ...