Large language models are astonishingly good at coding and getting better. What will this mean for software and software developers?