New OpenAI GPT-4 service will help spot errors in ChatGPT coding
OpenAI has introduced a new service called CriticGPT, based on GPT-4, to enhance the usefulness of generative AI tools for developers. CriticGPT is designed to identify errors in ChatGPT code outputs, aiming to improve human performance in code review tasks.
Improving AI Systems
OpenAI's initiative focuses on refining the 'Reinforcement Learning from Human Feedback' (RLHF) process to ensure higher quality and reliability in AI systems. The GPT-4 series, including ChatGPT, heavily relies on RLHF to produce reliable and interactive outputs.
Previously, the process involved manual efforts from AI trainers to rate ChatGPT responses for model enhancement. With CriticGPT, OpenAI can now autonomously critique ChatGPT's answers, addressing concerns about AI chatbots surpassing human trainers' capabilities.
Training and Results
CriticGPT was trained by inserting intentional mistakes into ChatGPT-generated code, resulting in critiques preferred by trainers around 63% of the time. The tool effectively reduces nitpicks and hallucinations in code outputs, demonstrating its potential to enhance AI-human collaboration.
However, while CriticGPT shows promise, AI-human collaboration remains crucial for more effective outcomes compared to AI alone.
Future Plans
OpenAI plans to scale its work on CriticGPT and implement it further in practice to enhance the capabilities of AI systems. The company acknowledges that while CriticGPT's suggestions may not always be accurate, they significantly help trainers identify problems in model-written answers.
Despite the complexity of identifying errors across various parts of an answer, OpenAI remains committed to advancing CriticGPT's capabilities for better performance.
