“The sequence-to-sequence team at Google Brain initially used 8 GPUs, which was considered a large amount of compute at the time.”