If you run a model on slower hardware are you getting more experience? Surely its a factor of model output reviewed and not human time.