Different concepts are learned at different layers of large language models, with more difficult concepts being fully acquired at deeper layers.