OpenAI Unveils CriticGPT to Catch Errors in AI-Generated Code
OpenAI introduces CriticGPT, a new model based on GPT-4 designed to help human reviewers spot bugs and hallucinations in AI-generated code. The tool outperforms average human reviewers and produces fewer false positives during the evaluation process.
OpenAI introduces a new AI model called CriticGPT to help human reviewers identify bugs and errors in code generated by ChatGPT. The tool aims to improve the traditional Reinforcement Learning from Human Feedback process by using generative AI to critique the outputs of other large language models, ensuring these systems behave more accurately.
Developers train CriticGPT by feeding it a dataset of code samples that contain intentionally inserted bugs. Human developers modify ChatGPT's code to introduce various errors and provide sample feedback, which teaches the model how to recognize and flag both common and uncommon coding mistakes.
In testing, CriticGPT outperforms the average human code reviewer, with human trainers preferring its critiques over human-written ones in 63% of cases. The model generates fewer unhelpful nitpicks and false positives, while a new technique called Force Sampling Beam Search allows users to adjust the thoroughness of its bug hunting and control its tendency to hallucinate non-existent errors.