OpenAI introduces CriticGPT, a GPT 4-based model that can detect errors in ChatGPT’s code output

OpenAI CriticGPT

OpenAI has recently unveiled CriticGPT, which is a new artificial intelligence (AI) tool based on GPT 4 and helps to find errors produced by the popular chatbot, ChatGPT. Unlike the other large language models (LLMs) from the company, which are more approached towards the users, CriticGPT is designed to “write critiques of ChatGPT responses to help human trainers or coders spot mistakes during reinforcement learning from human feedback.”

Additionally, OpenAI added that CriticGPT improves code review outcomes by more than 60 percent as compared to previous models. The AI company plans to integrate CriticGPT into its Reinforcement Learning from Human Feedback (RLHF) soon. CriticGPT can also write critiques that highlight the inaccuracies in the ChatGPT’s answers.

Further, OpenAI plans to develop CriticGPT’s capabilities by aiming to integrate advanced methods to improve human-generated feedback for GPT-4. This initiative underlines the ongoing importance of human oversight in refining AI technologies despite their increasing dependence on other automation capabilities.

However, there are certain limitations with CriticGPT. For starters, currently, it cannot generate longer and more comprehensive critiques, and its AI is trained on ChatGPT answers that are short. For future possibilities, it may be trained to deal with longer and more complex tasks.

OpenAI has also cautioned that the model still produces hallucinations, which may lead trainers to make mistakes after seeing these errors. The firm also added that real-world errors can be distributed across various parts of an answer, and they plan to pinpoint these mistakes, which can help multiple coders look out for all errors.

Also read: OpenAI rolls out ChatGPT app for Mac users: All the details