Preeti
OpenAI has developed a new AI model called CriticGPT to critique the work of its existing AI models, specifically ChatGPT. CriticGPT, based on GPT-4, aims to catch errors in ChatGPT’s code output and provide AI assistance to AI trainers. The model has been integrated into OpenAI’s Reinforcement Learning from Human Feedback (RHLF) labelling pipeline for efficient self-learning.
In experiments, CriticGPT was found to produce fewer nitpicks and hallucinations compared to ChatGPT critiques, making it preferred by trainers in 63% of cases. However, there are limitations to CriticGPT, such as being trained on short ChatGPT answers, potential hallucinations, and the inability to detect errors spread across multiple parts of an answer. Additionally, in complex tasks, even with model assistance, correct evaluation may be challenging.
OpenAI acknowledges that CriticGPT’s suggestions are not always correct, and further development is needed to address these limitations. The company aims to improve the model’s ability to critique long and complex tasks and reduce hallucinations. Despite these challenges, CriticGPT represents a significant advancement in evaluating outputs from advanced AI systems, providing valuable feedback for AI trainers.
Source link
Source link: https://www.medianama.com/2024/06/223-criticgpt-openai-new-tool-for-better-ai-code-reviews/
GIPHY App Key not set. Please check settings