... so training? Even "regular" pretraining is "inference plus training at the end" if you want to frame it that way