OpenAI Develops CriticGPT Model Capable of Spotting GPT-4 Code Generation Errors

OpenAI published a study on Thursday about a new artificial intelligence (AI) model that can detect GPT-4 errors in code generation. The AI ​​firm stated that the new chatbot was trained using the Human Feedback Reinforcement Learning Framework (RLHF) and was powered by one of the GPT-4 models. The underdevelopment chatbot was designed to improve the quality of AI-generated code that users get from large language models. The model is currently not available for users or testers. OpenAI also highlighted several limitations of the model.

OpenAI shares details about CriticGPT

The AI ​​firm shared details of the new CriticGPT model in a blog post, saying it was based on GPT-4 and designed to identify errors in the code generated by ChatGPT. “We found that when people get help from CriticGPT to review ChatGPT code, they outperform those without help 60 percent of the time,” the company says. The model was developed using the RLHF framework and the findings have been published in a paper.

RLHF is a machine learning technique that combines machine output with humans to train AI systems. In this system, human evaluators provide feedback on the AI's performance. It is used to adjust and improve the behavior of the model. Humans who provide feedback to the AI ​​are called AI trainers.

CriticGPT was trained on a large volume of code data that contained errors. The AI ​​model was responsible for finding these errors and critiquing the code. Because of this, the AI ​​trainers were asked to write the errors in the code on top of naturally occurring errors and then write example comments as if they had caught those errors.

After the CriticGPT shared their multiple variations of their critique, the coaches were asked to detect whether the errors they inserted were detected by the AI ​​along with the natural errors. OpenAI, in its research, found that CriticGPT was 63 percent better than ChatGPT at detecting errors.

However, the model still has certain limitations. CriticGPT was trained on short strings of code generated by OpenAI. The model has not yet been trained on long and complex tasks. The AI ​​firm also found that the new chatbot continues to mislead (generating factually incorrect answers). Also, the model has not been tested in scenarios where there are multiple scattered errors in the code.

This model is unlikely to be made public, as it is designed to help OpenAI better understand training techniques that can produce higher quality results. If CriticGPT makes it public, it is believed to be integrated into ChatGPT.