Hidden material on the back of creativity AI

Original version From This story It appeared Quanta Magazine.

We were once promised to the cars of the robot and the servant of the robot. Instead, we see the emergence of artificial intelligence systems that can beat us in chess, analyze huge texts of composing texts and songs. This has been one of the great wonders of the modern era: Physical work that is easy for humans is very difficult for robots, while algorithms are increasingly capable of imitating our minds.

Another surprise that has attracted long -term researchers is that this kind of strange and strange creativity is.

Release models, the spine of image -makers such as Dall · E, Imagen and Patable Diffusion are designed to produce carbon versions of the images they are trained. However, in practice, they seem to improvise the elements in the images to create something new – not only colorful color spots, but coherent images of semantic meaning. The Paradox is behind the release models, saying: “Giulio Biroli, a scholar and physicist at Supérieure école Normale in Paris:” If they were fully, they just have to remember. ” “But they don’t do it – they are really capable of producing new samples.”

To produce images, the emission models use a process known as densoising. They convert an image to digital noise (a set of pixels), then collects it again. It is like putting a painting through the crushing over and over again so that all that remains is the fine dust candle, then attach the pieces together. The researchers have been surprised for many years: If the models are just rebuilding, then how does it just get into the picture? It is like put your chopped painting again in a completely new work of art.

Two physicists have now made an astonishing claim: these technical deficiencies in the process of removal that lead to the creativity of publishing models. In an article presented at the International Conference on Machine Learning 2025, this dual has developed a mathematical model of trained publishing models to show that their so-called creativity is actually a definite process-a direct and inevitable result of their architecture.

By turning on the black box of publishing models, new research can have great consequences for future research of artificial intelligence – and perhaps even for our understanding of human creativity. “The real power of the article is that it creates very accurate predictions of very illegal something,” said Luka Ambrogoni, a computer scientist at Radod University in the Netherlands.

Down

Mason Cambodian, a graduate student studying applied physics at Stanford University and the main author of the new article, has long been fascinated by morphogenesis: a process that has their own life systems.

One way to understand the growth of embryos in humans and other animals is named after what is known as a Turing pattern, which is the twentieth -century mathematician Alan Turing. Turing patterns explain how cell groups can organize themselves in separate organs and organs. Importantly, this coordination is all locally. There is no CEO of cell monitoring to ensure that they all match a final body program. In other words, solitary cells do not have several designs of the body to establish their work. They only act in response to the signals of their neighbors and make reforms. The system is usually running low, but for example, it occasionally becomes clumsy.