[phowon]

The BERT paper also introduced BERT Base, with is 12 layers with approximately the same number of parameters as GPT, but still outperforms GPT on GLUE.