The evolution of artificial intelligence has produced surprising capabilities, notably in mimicking human intellect and creativity, instead of the anticipated advancements in robotics. Diffusion models, essential for image generation in tools like DALL·E and Stable Diffusion, generate new images by creatively reassembling digital noise into coherent visuals. Research by physicists suggests that the models’ creativity stems from inherent imperfections in their denoising processes, revealing a deterministic basis for their seemingly innovative outputs. This challenges traditional perceptions of creativity as a higher-order function and aligns it with bottom-up processes seen in natural morphogenesis, where individual components react locally without a central blueprint. The study highlights that technical features, previously considered limitations of the models, are pivotal in fostering creativity. This new understanding could profoundly influence future AI research and reshape our notions of creativity, both artificial and human.
Source link
