CriticGPT: The AI That’s Changing How We Train AI Models!

OpenAI introduces a new bot, CriticGPT, to help humans help AI!

OpenAI’s New AI Training Technique

OpenAI, the company behind ChatGPT, is exploring a new technique to enhance AI models by incorporating more AI into the training process (Wired). Traditionally, OpenAI used reinforcement learning with human feedback (RLHF), where human trainers rate the AI’s outputs to improve its performance. However, this method has limitations due to inconsistent human feedback and the complexity of rating sophisticated outputs like code. To address these issues, OpenAI developed a new model called CriticGPT, which assists human trainers by catching bugs and providing better critiques.

Enhancing AI Accuracy and Reliability

CriticGPT was fine-tuned using OpenAI’s most powerful model, GPT-4, and showed promising results in assessing code. Human judges preferred CriticGPT’s critiques 63% of the time, suggesting it can enhance the accuracy and reliability of AI models. OpenAI plans to integrate this technique into their RLHF chat stack to reduce errors in human training and improve overall model performance. This approach might also help train AI models that surpass human abilities, as AI continues to evolve.

Future Implications and Competition

The development of CriticGPT is part of OpenAI’s broader efforts to improve large language models and ensure they behave acceptably. This technique could be crucial in training increasingly powerful AI models while maintaining their alignment with human values. Meanwhile, competitors like Anthropic are also advancing their AI training methods to enhance their models. The success of these new techniques will determine their impact on the future of AI development and its alignment with human goals.

(Visit Wired for the full story)

*An AI tool was used to add an extra layer to the editing process for this story.