BERT is ~100m parameters. GPT3.5 is 1000x bigger. A Bert sized decoder model would likely generate nonsense For most applications BERTs output is used to fine

BERT is ~100m parameters. GPT3.5 is 1000x bigger. A Bert sized decoder model would likely generate nonsense For most applications BERTs output is used to fine tune additional NN layers, eg for text classification