“Kaplan et al. (2020) found that larger models are more sample-efficient, reaching a given loss with fewer optimization steps and data points than smaller models.”