CriticGPT is an AI model created by OpenAI that critiques code generated by ChatGPT and identifies bugs in the code. The issue of AI hallucinations, where models assert falsehoods under the guise of authority, has led some scholars to believe that generative AI cannot detect or correct its errors. However, OpenAI disagrees and introduced CriticGPT as a version of GPT-4 to help find and correct mistakes in code.
CriticGPT focuses on programming code writing, aiming to improve the accuracy of AI-generated code by catching mistakes. The researchers at OpenAI trained CriticGPT by having human programmers review its critiques of programming code and provide feedback. The model was also trained to identify deliberate bugs inserted by human contractors, with the goal of learning to associate bugs with explanations.
The results showed that CriticGPT was able to find more bugs in the code than human reviewers, leading to more comprehensive critiques. However, the model also had a tendency to hallucinate bugs that were not actually present, raising concerns about the tradeoff between bug detection and hallucinations.
One potential issue with OpenAI’s approach is the lack of transparency, as the CriticGPT model and its training data are not publicly available for examination. This raises questions about oversight and accountability in AI development. Overall, CriticGPT shows promise in improving code quality with the assistance of AI, but challenges remain in balancing accuracy and potential errors.
Source link
Source link: https://www.zdnet.com/article/openai-proposes-a-second-neural-net-to-catch-chatgpts-code-mistakes/