OpenAI’s CrtiticGPT is a new AI tool the company is developing but the beta version is expected to be out sometime later. This upcoming tool is meant to improve AI-generated code. 

Back up by a published paper, OpenAI affirms that users who utilise CriticGPT to analyse ChatGPT code perform 60% better than users who do not use the AI tool. 

The AI giant stated that it is currently integrating CriticGPT models into its RLHF (reinforcement learning from human feedback) machines which assists the company’s AI trainers. RLHF collects feedback which the AI trainers compare and rate the various responses provided by ChatGPT. 

OpenAI's CriticGPT and Why it is Needed

Over the years since its inception, ChatGPT has improved by leaps and bounds regarding its accuracy with answers and suggestions. Though this is a benefit, it also poses a challenge for RLHF. This is due to the lack of inaccuracies or mistakes available for AI trainers to draw their comparisons from. Alignment of models becomes more of a hurdle as responses become more accurate to which trainers are unable to provide feedback. CriticGPT is OpenAI’s answer to solving this problem. 

GPT-4 powers ChatGPT and CriticGPT is also based on this very model. Additionally, RLHF trained and integrated with CriticGPT. However, the new AI tool perceived more inaccuracies to be critiqued. AI trainers were then instructed to enter the inaccuracies into ChatGPT’s code and compose sample feedback under the pretence that the trainers found the bug.

The trainers are then asked to compare numerous critiques of the reviewed code. This is done to make it easier to detect a critique that catches the bug. CriticGPT also creates lengthier and more extensive critiques.

Suggested:

From OpenAI, ChatGPT Enterprise is Now Accessible.

ChatGPT for macOS: The AI Chatbot’s Integration into Apple’s OS.