“Muennighoff et al. (2023) found that in data-constrained training, the value of excess model parameters decays faster than the value of repeated data tokens.”