“According to Pearce & Song (2024), the local power-law exponent for optimal model size scaling is close to Kaplan et al.'s 0.73 in the small-model regime but converges to Chinchilla's 0.5 as compute gets larger.”