Generative AI can’t find its own errors. Do we need better prompts?

If a large language model can ‘self-correct’ why doesn’t it just give the right answer to begin with? Google’s DeepMind scientists report on this paradox.