Transcendence: Generative Models Can Outperform The Experts That Train Them

This paper investigates the concept of "transcendence" in generative models, where AI surpasses human performance, using the example of an autoregressive transformer trained on chess games. Low-temperature sampling is found to be crucial for achieving transcendence, as it enables the model to outperform individual human experts by combining and denoising expert knowledge. The study presents two theorems that establish conditions for transcendence, both in the case of a single noisy expert and multiple experts. Experiments with ChessFormer models demonstrate the model's ability to transcend human-level performance, particularly when trained on diverse datasets. The research also connects transcendence to ensemble learning and offline reinforcement learning, highlighting the importance of dataset diversity for optimal performance. The study suggests future research in various domains and addresses the implications of AI models surpassing human abilities, while emphasizing the role of denoising in decision-making.