OpenAI introduces CriticGPT, an AI tool to identify errors in AI-generated code

From ABP News: 2024-06-29 11:31:11

OpenAI has developed a new AI model called CriticGPT based on ChatGPT-4 to identify code errors, already improving code review outcomes by 60%. The model is set for integration into OpenAI’s RLHF labeling pipeline, aiming to enhance trainers’ ability to evaluate complex AI outputs effectively.

GPT-4 models powering ChatGPT are designed to improve interactivity and utility through RLHF, allowing trainers to assess and rate responses. As ChatGPT’s reasoning advances, errors become subtler, making it challenging for trainers to detect inaccuracies effectively.

CriticGPT has shown competence in analyzing code and detecting errors that may be overlooked by humans. Trained on a dataset with intentionally inserted bugs, the model effectively flags coding errors, aiding in the detection of hallucinations.

During experiments, teams using CriticGPT provided more holistic critiques and fewer false positives compared to solo trainers. However, concerns regarding the model’s training on brief responses from ChatGPT highlight the need for further development to handle longer, more complex tasks effectively.

There are occasional labelling errors by trainers, with a focus on isolated errors rather than holistic feedback. As AI models become more proficient, there is a concern that human trainers may struggle to provide meaningful feedback within the CriticGPT framework effectively.



Read more at ABP News: OpenAI Unveils New Tool, CriticGPT, To Find Errors In AI-Generated Code: Report