“A Transformer model can achieve the same quality as a much larger LSTM model while using 10 to 100 times less compute and being 10 times smaller.”