Demystifying language models