OpenAI's Superalignment Team has recently announced a groundbreaking development in AI governance and alignment research. This groundbreaking study investigates the utilization of deep learning's generalization properties to empower 'weak supervisors', such as humans, to effectively manage more advanced AI models.

In a remarkable achievement, the research team has successfully utilized the GPT-2 model to achieve supervisory performance levels comparable to those of GPT-4. This achievement demonstrates that smaller AI models can effectively generalize and solve complex problems.

This research aims to address the limitations of current AI alignment methods, including Human Feedback-based Reinforcement Learning (RLHF). With future AI systems capable of complex and creative actions beyond human understanding, the safe supervision and control of these systems becomes increasingly crucial.

The team expressed confidence that this new approach represents a significant advancement in aligning superhuman AI models. The research focuses on the supervisory relationship between humans and AI, operating under the premise that human intelligence may be inferior to AI in the future.

To support this research, OpenAI has launched a $10 million grant program, highlighting the significance of safely adapting superintelligent AI systems. The research team anticipates that this approach will offer new solutions to future challenges in AI alignment.

The study focuses on finding methods for humans to effectively control AI, even if human performance is inferior. The Superalignment Team is committed to developing alignment technologies to prevent AI from surpassing human control and causing societal harm.

The team used the GPT-2 model to oversee the more advanced GPT-4, a technique expected to aid in addressing critical issues in AI alignment. The research revealed that although the GPT-4 model, trained on GPT-2's outputs, demonstrated better overall performance, it exhibited lower performance in specific areas. This suggests that lower-performing AI can still have a significant impact on more advanced AI systems.

이영우 강남포스트 학생기자  webmaster@ignnews.kr

<저작권자 © 강남포스트, 무단 전재 및 재배포 금지>