The hidden ingredients behind the creativity of artificial intelligence

The original version to This story Appear Quanta magazine.

We have once promised self -driving cars and robot maids. Instead, we saw up artificial intelligence Systems that can strike us in chess, analyze huge sets of text, and composition This was one of the great surprises of the modern era: the physical tasks that facilitate people very difficult for robots, while algorithms are increasingly able to simulate our thought.

Another surprise that researchers have long received is the talent of these algorithms for their strange type of creativity.

The proliferation models, the backbone of images generation tools such as Dall · E and Imagen and stable proliferation, are designed to create carbon copies of the pictures that have been trained on. In practice, they seem to be improvised, mixed inside the images to create something new – not just illogical points of colors, but the coherent images with a semantic meaning. This is the “paradox” behind the proliferation models, he said Julio PerollyHe said that a researcher at Amnesty International and Physicists in Superior in Paris: “If they work perfectly, they should just save it.” “But they don’t do it – they are already able to produce new samples.”

To create pictures, Proliferation models use a process known as a reduction. They convert a picture into digital noise (an unsheatted group of pixels), then reassemble it. It is similar to placing the painting over and over to tear until all that I left is a pile of soft dust, then the cutting together. For years, the researchers asked: If the models are re -assembled, how can the grandmother enter the image? It is like re -assembling your tearing painting in a completely new artwork.

Now two physicists have made an amazing claim: they are the artistic defects in the process of reducing themselves that lead to the creativity of prevalence. in paper The duo at the International Conference on ID 2025 presented a sporting model for trained spread models to show that what is called their creativity is in fact an inevitable process-which is an inevitable direct result of its architecture.

By shedding light on the black square of prevalence, new research can have great impacts on future artificial intelligence research – and perhaps even for our understanding of human creativity. “The real strength of the paper is that it makes very accurate predictions of something very not trivial,” he said, he said, “He said, he said, he said,” Luca AmbrogenComputer scientist at Radbod University in the Netherlands.

The bottoms are up

Mason CampA highly studied student studying applied physics at Stanford University and the main author of the new paper, has long been fascinated by the formation: the operations that live systems bring together.

One way to understand the development of embryos in humans and other animals is through what is known as a Torring styleIt was named after the mathematics scientist in the twentieth century Alan Torring. Torring patterns explain how cell groups can organize themselves in distinctive organs and limbs. Decally, this coordination occurs at the local level. There is no CEO overseeing cell trillion to ensure that they are all compatible with the final body plan. Individual cells, in other words, have no final plans for the body that builds on their work. They take measures and make corrections in response to signs from their neighbors. This system is usually turned on from the bottom up smoothly, but from time to time it is going well, for example, hands are produced with additional fingers, for example.