Also not every domain need deep learning at all. With tabular data gbm is still king (I am happy for every example where DL outperformed gbm on tabular data, as I am also would be happy to use more dl, but I can't just use more complex architectures for the sake of complexity.) Also not every company has infinite data which is needed for the data hungry DL models. So there are situations where knowing transformer architecture would be feasable, but with the available data it is gonna be gbm still. There are still data scientict positions out there where you can still have a big impact on the business as a 'sklearn kiddie' (okay maybe xgboost or lightgbm kiddie) and it would not help any more if you would know all the DL architectures.
In the end it should be all about business impact (if you are working in the competitive sector) and not who is using the latest, freshest architectures.
Also not every domain need deep learning at all. With tabular data gbm is still king (I am happy for every example where DL outperformed gbm on tabular data, as I am also would be happy to use more dl, but I can't just use more complex architectures for the sake of complexity.) Also not every company has infinite data which is needed for the data hungry DL models. So there are situations where knowing transformer architecture would be feasable, but with the available data it is gonna be gbm still. There are still data scientict positions out there where you can still have a big impact on the business as a 'sklearn kiddie' (okay maybe xgboost or lightgbm kiddie) and it would not help any more if you would know all the DL architectures.
In the end it should be all about business impact (if you are working in the competitive sector) and not who is using the latest, freshest architectures.