The dominant recipe for building better language models has not changed much since the Chinchilla era: spend more FLOPs, add more parameters, train on ...