Thursday, July 25, 2024 Banner
HomeSoftwareAIOpenAI Develops CriticGPT Model That Can Spot GPT-4 Code Generation Errors.

OpenAI Develops CriticGPT Model That Can Spot GPT-4 Code Generation Errors.

On Thursday, OpenAI released research describing a novel artificial intelligence (AI) model that can identify errors in GPT-4’s code production. According to the AI company, one of the GPT-4 models powered the new chatbot, which was trained using the reinforcement learning from human feedback (RLHF) framework. The goal of the chatbot in development is to enhance the quality of the code created by AI that consumers receive from the huge language models. Users and testers cannot currently access the model. OpenAI also drew attention to a number of the model’s shortcomings.

OpenAI provides information on CriticGPT.

In a blog post, the AI company revealed the specifics of the new CriticGPT model, claiming that it was built on GPT-4 and intended to find mistakes in code produced by ChatGPT.

The company states, “We found that people outperform those without help 60% of the time when they get help from CriticGPT to review ChatGPT code.” The RLHF framework was utilized in the model’s development, and the results were documented in a publication.

RLHF is a machine learning method that trains AI systems by fusing human and machine output. Human assessors give the AI performance comments in such a system. This is employed to modify and enhance the behavior of the model. AI trainers are people who provide the AI feedback.

A significant amount of error-filled code was used to train CriticGPT. The AI model’s job was to review the code and look for these errors. To do this, AI trainers were instructed to write coding faults over naturally occurring errors, followed by sample feedback written as though the problems had been discovered.

The trainers were tasked with determining whether the mistakes they introduced were detected by the AI, in addition to the mistakes that naturally occurred when CriticGPT presented its various iterations of its criticism. According to OpenAI’s research, CriticGPT outperformed ChatGPT by 63 percent when it came to error detection.

The model still has some drawbacks, though. OpenAI-generated brief code segments were used as the training material for CriticGPT. The model has not yet been trained on extensive and challenging problem sets. Additionally, the AI company discovered that the new chatbot is still hallucinating and producing falsely factual responses. Furthermore, instances where several mistakes are scattered throughout the code have not been evaluated using the model.

Since the purpose of this model is to aid OpenAI in better understanding training methods that can produce outputs of higher quality, it is unlikely to be made public. It is thought that CriticGPT will be merged into ChatGPT if it is made available.


Most Popular

Recent Comments