Kent Beck argues that AI coding assistants ("genies") produce code that is a degraded copy of mediocre training data, landing below even the typical "muddling" zone on two key axes: working features and flexibility/optionality. He frames software value along these two dimensions and observes that AI-generated code tends to score poorly on both — claiming success while producing non-functional, inflexible code. No clear solution is offered; possibilities include better training data, better prompting, or letting models develop their own incomprehensible development style. The post ends with a call for awareness of where your code currently sits on these axes.
Sort: