We Need Efficient and Transparent Language Models