The original version of This story appeared in How many magazine.
We had once promised independent cars and robot chambermaids. Instead, we have seen the rise of artificial intelligence systems that can beat us in failures, analyze huge oars of text and compound sonnets. This was one of the major surprises of the modern was: the physical tasks that are easy for humans are very difficult for robots, while algorithms are able to imitate our intellect.
Another surprise that has long been in perplexed research is the talent of these algorithms for their own type of strange creativity.
The diffusion models, the backbone of image generation tools such as dall · e, image and stable diffusion, are designed to generate carbon copies of images on Whhhhhhohho have been formed. In practice, however, they seem to improvise, mixing elements in images to create something new – not just absurd stains, but images consistent with semantic meaning. It is the “paradox” behind the diffusion models, says Giulio BiroliA researcher and physicist for AI AI at the normal school of superior to Paris: “If they worldwide, they simply show memorization,” he said. “But they don’t die – they are actually able to produce new samples.”
To generate images, Diffusion models use a process called Denising. They convert an image to digital noise (an incoherent collection of pixels), then re -gather it. It is like putting a painting several times through a shredder until you have a pile of fine dust, then bring the pieces together. For years, researchers have Wondard: if the models only go up, then how is novelty in the image? It is like enjoying your shredded painting in a brand new work of art.
Now, two physicists have made a surprising affirmation: it is the technical imperfections of the deenizing process itself which leads to the creativity of the diffusion models. In a paper Presented at the international conference on 2025 automatic learning, the duo has developed in the mathematical model of diffusion models formed to show that their creativity So-He-Cland is in fact a deterministic process-a direct and inevitable consequence of their architecture.
By illuminating the black box of diffusion models, the new research could have great implications for future research on AI-and perhaps even for our sub-demand of human creativity. “The real force of the paper is that it makes very precise predictions of something very not embraced,” said Luca Ambrogionito the computer scientist at Radboud University in the Netherlands.
Up to the top
Mason KambA graduate student studying physics applied at the University of Stanford and the main author of the new article, has long been fascinated by morphogenesis: the processes by which living systems self-assemble.
A way to highlight the development of embryos in humans and other animals is what is called a Turing patternThe name of the 20th century mathematician Alan Turing. Turing models explain how cell groups can be organized in district organs and members. Crucialyy, this coordination takes place at the local level. There is no CEO supervising billions of cells to ensure that an ally complies with a final body plan. Individual cells, in other words, do not have a finished plan of a body on Whhh to base their work. They simply act and make corrections in response to the signals of their neighbors. This upward system is generally smooth, but from time to time, it goes from time to time with additional fingers, for example.
