Original version With This story appeared in How much warehouse.
We used to promise cars and maids of robots. Instead, we saw the creation of artificial intelligence systems that can beat us in chess, analyze the huge rice of text and create sonnets. It was one of the great surprises of the newfangled era: physical tasks that are simple for people turn out to be very hard for robots, while the algorithms are increasingly able to imitate our intellect.
Another surprise that has long been embarrassed by researchers is the talent of algorithms for their own strange creativity.
Diffusion models, the spine of the tools generating image, such as Dall · E, Imagen and Stable Diffusion, have been designed to generate a copy of carbon images in which they were trained. In practice, however, it seems to be improvised, mix elements in paintings to create something recent – not only nonsense drops of colors, but consistent images of semantic meaning. This is a “paradox” behind the diffusion models, he said Giulio BiroliAI researcher and physicist at École Normale Supérieure in Paris: “If they acted perfectly, they should simply remember,” he said. “But this is not the case – in fact they can produce new samples.”
To generate images, Dyfusion models use the process known as denoising. They convert the image to the digital noise (inconsistent pixel collection), and then it again. It is as if putting the image repeatedly through the shredder, as long as everything remains, it is a pile of miniature dust, and then patching the pieces back. Over the years, scientists have wondered: does the models just bear how recent appears in the picture? It’s like submitting a crushed image to a completely recent work of art.
Now two physicists have a surprising claim: this technical imperfections in the denoising process itself leads to the creativity of diffusion models. IN paper Presented at an international conference on machine learning 2025, the duo has developed a mathematical model of trained diffusion models to show that their so-called creativity is in fact a deterministic process-indirect, inevitable consequence of their architecture.
By highlighting the black box of diffusion models, recent research can have great implications for future AI research – and maybe even for our understanding of human creativity. “The real power of paper is that he makes very accurate forecasts of something very non -trial,” he said Luca AmbrogioniIT specialist at Radboud University in the Netherlands.
Bottom
CambrbA graduate studying physics used at Stanford University and the main author of The Fresh Article, has long been fascinated by morphogenesis: processes in which live systems self -organize.
One way to understand the development of embryos in humans and other animals is something that is known as A Turing patternNamed in honor of the 20th-century mathematics Alan Turing. Turing patterns explain how cell groups can organize on separate organs and limbs. Most importantly, this coordination takes place at the local level. There is no general director supervising cell trillion to make sure that they all are in line with the final body plan. In other words, individual cells do not have a ready body plan on which you can base your work. They simply take action and make corrections in response to signals from their neighbors. This bottom-up system usually works smoothly, but from time to time it goes not so-an example, producing hands with additional fingers.