How good are LLMs at fixing their mistakes? A chatbot arena experiment with Keras and TPUs