OpenAI, a company known for the ChatGPT chatbot has unveiled CriticGPT- an advanced AI tool which is based on the GPT-4 architecture, which has been designed to help developers identify and correct errors in code generated by ChatGPT. This innovation further aims to significantly improve the accuracy of AI systems by providing reliable outputs.

Improving code review

CriticGPT has been designed to enhance code review processes by 60 per cent when compared to traditional methods. This new tool will soon integrate OpenAI’s Reinforcement Learning from Human Feedback (RLHF) pipeline, which will offer AI trainers for more effective means to assess and refine AI-generated content, regardless of complexity.

Training process and performance

To develop CriticGPT, human developers will be able to edit codes which were written by ChatGPT- by introducing various errors and providing feedback. This training will enable the CriticGPT to recognize common and uncommon coding mistakes. It was further stated that the results were impressive and further demonstrated CriticGPT’s ability to pinpoint errors which might be overlooked by human reviewers.

However, CriticGPT will be facing challenges in identifying issues which were spread across multiple code strings, a common source of AI ‘hallucinations’ or errors. Despite this, CriticGPT’s performance marks a significant step forward in AI-assisted code review.

Future developments in AI

CriticGPT is one of the several new tools that aims at enhancing large language models and their generative capabilities. OpenAI is said to be working on the next-generation AI model, which will be significantly more intelligent. 

As per the OpenAI CTO Mira Murati, while GPT-3 has the intelligence of a toddler and GPT-4 that of a high-schooler, the upcoming model will possess the intelligence equivalent to a PhD holder in specific tasks. This next-gen model is further expected to be released in about 1.5 years, and it may even surpass human intelligence in some areas (reports state).

Collaboration with Time Magazine

As per the further advancement of the AI models, OpenAI has reportedly partnered with Time Magazine in a multi-year agreement. This will grant OpenAI access to over 100 years of Time’s archives, which could be used to train and enhance ChatGPT. 

This partnership highlights Time’s commitment to making accurate and reliable information which is widely accessible, while also supporting the development of smarter AI systems.