The compute used to train the final model in the 2017 Transformer paper was roughly 1/10,000th of the training compute required for GPT-3