OpenAI has announced a new model named CriticGPT, aiming to identify errors in the code generated by the ChatGPT robot. CriticGPT is based on the GPT-4 model, and OpenAI says this new model has demonstrated high effectiveness, contributing to improving human performance in code review rather than replacing them.
The OpenAI initiative aims to develop the “Reinforcement Learning through Human Feedback” (RLHF) process to achieve a high level of quality and reliability in artificial intelligence systems. The GPT-4 models, which power the publicly available versions of ChatGPT, heavily rely on this process to ensure the reliability and interactivity of their outputs. So far, this process has been manual, relying on the efforts of human trainers who evaluate ChatGPT responses to enhance the model’s performance.
With the launch of CriticGPT, OpenAI can now evaluate ChatGPT responses autonomously, addressing concerns about artificial intelligence surpassing many human trainers.
OpenAI collaborated in developing the CriticGPT model by training it with specific feedback from trainers who deliberately introduced errors in the code generated by ChatGPT. The results were encouraging, as trainers showed a preference for CriticGPT criticism by up to 63%, thanks to the tool’s ability to reduce minor criticisms and hallucinations.
CriticGPT represents a significant advancement in artificial intelligence technologies, providing a mechanized way to assess the quality of these models’ results. With further development, CriticGPT can significantly accelerate the advancement of artificial intelligence models, increasing their reliability and efficiency.
