OpenAI has trained a new model, the CriticGPT based on GPT-4 for find errors in responses generated by ChatGPT .
According to the company, when people rely on CriticGPT to review the codes generated by ChatGPT, they are able to overcome potential issues without additional help 60% of the time.
The series of GPT-4 models that power ChatGPT are corrected through what is called “reinforcement learning from human feedback” (RLHF), which means that the platform improves its performance as users, also called AI trainers, rate the responses given by ChatGPT and flag potential errors.
As ChatGPT improves, its errors will also become more subtle and specialized, which can make it harder for AI trainers to catch inaccuracies when they occur. That’s where CriticGPT can help.
The next step, according to the company, is to incorporate models similar to CriticGPT into the RLHF process, so that human feedback can also be aided by AI feedback.
While CriticGPT’s suggestions aren’t always correct, they can help AI trainers spot issues that might otherwise go unnoticed. And, like the GPT-4 model itself, CriticGPT will also improve as more users use the tool and provide feedback.
According to OpenAI, tests between the two models revealed that CriticGPT’s predictions were considered better than ChatGPT’s own 63% of the time. In addition, the new tool also demonstrated fewer hallucinations — moments in which the AI invents something that is not true and presents it as if it were.
However, CriticGPT still has some limitations and can only help to a certain degree of expertise. If a task or answer is extremely complex, even an expert with the help of a model may not be able to evaluate it correctly.
Source: CNN Brasil
Charles Grill is a tech-savvy writer with over 3 years of experience in the field. He writes on a variety of technology-related topics and has a strong focus on the latest advancements in the industry. He is connected with several online news websites and is currently contributing to a technology-focused platform.
