OpenAI introduces ‘CriticGPT’ to review code quality

OpenAI recently announced the launch of a new model called CriticGPT based on GPT4 which writes critiques of ChatGPT responses to help human trainers spot errors during “reinforcement learning from human feedback (RLHF)”, unlike other models from the company which are usually consumer-facing.

Based on GPT-4, CriticGPT will assist OpenAI’s human trainers in “catching errors in ChatGPT’s code output.” OpenAI claims that code examined by Critic-GPT can achieve a 60% performance advantage over unreviewed code. The business is also including Critic-GPT-like models in the RLHF labeling process to help AI trainers assess the results of sophisticated AI systems.

According to reports, during the experiments of testing CriticGPT, teams using CriticGPT produced more comprehensive critiques and identified fewer false positives compared to those working alone. “A second trainer preferred the critiques from the Human+Critic-GPT team over those from an unassisted reviewer more than 60 percent of the time, as reported by,” ‘LLM Critics Help Catch LLM Bugs.’

The business is also including CriticGPT-like models in the RLHF labeling process to help AI trainers assess the results of sophisticated AI systems. To teach CriticGPT to quickly identify frequent and uncommon problems, the model was trained by manually altering ChatGPT-generated code, adding new faults, and providing example feedback.

CriticGPT’s recommendations are not always accurate, just like human suggestions are. That being said, the Human+CriticGPT combo is said to perform better than human trainers working alone and to assist trainers in writing “comprehensive critiques” while experiencing fewer hallucinations.

Critics countered that CriticGPT had its limits. CriticGPT appears to have been taught using brief ChatGPT responses. Therefore, additional development is required to tackle longer and more difficult jobs. Another flaw with CriticGPT is “ChatGPT hallucinating,” which occurs when an AI model produces false information but presents it as truth. Additionally, according to OpenAI, CriticGPT is unable to assess a job or response that is exceptionally complicated and may propagate errors from the real world across the solution. As per the business, this new AI model will assist human trainers to “produce better RLHF data for GPT-4,” and OpenAI is also planning to scale this work further.

Future Enhancements and Development Plans for CriticGPT

The latest technology may potentially benefit OpenAI’s next AI models. The business recently disclosed that it is developing a new, more potent AI model. Recently, OpenAI CTO Mira Murati provided details on the intelligence of the next AI model.

According to Murati, GPT-3 was as intelligent as a child, GPT-4 was as intelligent as a high school student, and the upcoming model generation would, for certain jobs, be as intelligent as a PhD holder. Additionally, she disclosed that the upcoming GPT generation should be available in “a year and a half.” She went so far as to say that you could believe that the next-generation chatbot is wiser than you in many situations when you converse with it.

Potential Impact of CriticGPT on Future AI Models

OpenAI and Time Magazine are working together to enhance the intelligence of its new generative models. To improve and train ChatGPT, the artificial intelligence research business has partnered with Time magazine. The firms launched a multi-year content agreement on Thursday, which would provide OpenAI access to both recent and accomplished stories from Time’s archive spanning more than a century. With this new collaboration, Time continues to uphold its mission to increase access to reliable and accurate information worldwide.

You might also be interested in – Musk threatens to ban Apple devices over OpenAI partnership