the machines are always trained with the same dataset for each task. the biggest difference right now is small technical modifications on models that are also pre trained on gigantic unlabelled datasets. this doesn't feel like we're teaching them to do the test specifically at all