AI vs AI: OpenAI Creates CriticGPT to Identify Flaws in GPT-4 Output
Dubai’s Roads and Transport Authority (RTA) has received a five-star rating from the British Safety Council for 2024. This award recognizes RTA’s high standards in health and safety management at work sites, construction areas, and during operations.
CriticGPT assists human trainers who use “reinforcement learning from human feedback” (RLHF) to train AI systems. It focuses on analyzing code generated by ChatGPT, spotting potential mistakes, and helping humans catch errors more easily.
To train CriticGPT, researchers provided it with code containing intentional errors. This helped CriticGPT learn to recognize and flag different coding mistakes.
Tests showed that CriticGPT found errors 63% more effectively than human reviewers when dealing with real-world mistakes from large language models (LLMs). Teams using CriticGPT alongside human reviewers produced more detailed error reports and reduced false errors compared to reviews done by AI alone.
However, CriticGPT has limitations. It was trained on short code snippets, which may not work as well for longer, more complex tasks. It also reduces, but doesn’t completely eliminate, false errors, and human reviewers can still be misled.
The researchers note that CriticGPT is most effective when errors are isolated. Real-world errors can be spread throughout an output, which may be challenging for future versions of CriticGPT.
OpenAI plans to incorporate AI tools like CriticGPT into the review process for training large language models. These tools would help identify errors in code, making it easier for humans to spot mistakes and improve the evaluation of complex AI outputs.
