Finding GPT-4’s mistakes with GPT-4

OpenAI has trained CriticGPT, a model based on GPT-4, to catch errors in ChatGPT's code output, and people who get its help outperform unassisted reviewers 60% of the time. The company says it is beginning to integrate CriticGPT-like models into its RLHF labeling pipeline to give trainers explicit AI assistance. CriticGPT's suggestions are not always correct, but the article states that it helps trainers catch many more problems than they would alone, and that Human+CriticGPT critiques were preferred over unassisted critiques by a second trainer more than 60% of the time.