More relevantly, it means that the model parameter count, dataset size, and training procedure are balanced such that they can achieve equal or better performance as GPT-3 while being smaller.