The effectiveness of MAE pre-pretraining for billion-scale pretraining