OpenAI's GPT-3.5 and GPT-4 both fell down when challenged with greater and greater levels of this basic computer task, pointing to a fundamental omission in their design.