How so? I see lots of similarities both in the training and inference/prompting stages.