Compute-optimal training

Choosing the model size and the amount of training data that give the lowest loss for a fixed training budget. Chinchilla found that both should grow at about the same rate.