AI models already have width, the number of nodes in a layer. Each node is a fundamental processing unit that takes inputs and creates outputs within the larger network, like the neurons in your brain working together. There’s also depth, the number of layers in a network. Here’s how Rensselaer Polytechnic Institute’s Ge Wang, Ph.D., a co-author of the study, explains it: “Imagine a city: width is the number of buildings on a street, depth is how many streets you go through, and height is how tall each building is. Any room in any building can communicate with other rooms in the city,” he says in an email. When you add height, as internal wiring or shortcuts within a layer, it creates “richer interactions among neurons without increasing depth or width. It reflects how local neural circuits in the brain … interact to improve information processing.”

Wang and his co-author Feng-Lei Fan, Ph.D., achieve this increased dimensionality within neural networks by introducing intra-layer links and feedback loops. Intra-layer links, Wang says, resemble the lateral connections found in the cortical column of the brain, which is linked with higher-level cognitive functions. These links form connections among neurons in the same layer. Feedback loops, meanwhile, are similar to recurrent signaling, which is where outputs impact inputs. Wang says this could improve a system’s memory, perception, and cognition.

“Together, they help networks evolve over time and settle into stable, meaningful patterns, like how your brain can recognize a face even from a blurry image,” Wang says. “These structures enrich AI’s ability to refine decisions over time, just like the brain’s iterative reasoning.”

AI’ most compelling leap into humanlike behavior may be in the birth of large language models like ChatGPT. On June 12, 2017, eight researchers from Google wrote the paper “All You Need Is Attention,” introducing a new deep-learning architecture known as “transformer,” a technique that focused on attention mechanisms that drastically decreased training times for A.I. models. This paper largely fueled the subsequent AI boom—and is why AI suddenly seemed to be everywhere all at once.

However, it didn’t materialize the “holy grail” of most AI research: artificial general intelligence (AGI). This is the moment when AI systems adapt to truly “think” like a human being. Some futurists consider AGI to be a key pathway to reaching the hypothetical singularity, the point when AI surpasses human cognitive powers.

The reason transformer architecture didn’t fulfill AGI requirements, according to study co-author, Wang, is that it has some inherent limitations. Reuters reported in 2024 that AI companies were no longer seeing the previously observed exponential growth in AI capability. AI’s “scaling law” —the idea that larger data sets and more resources would make models better and better—was no longer working. This prompted some experts to believe that new innovations were needed to push these models even further. This new study suggests that integrating intra-layer links and feedback loops could enable models to relate, reflect, and refine outputs, making them both smarter, as well as more energy efficient.

“One misconception is that our perspective merely proposes ‘more complexity,’ Wang says. “Our goal is structured complexity, adding dimensions and dynamics that reflect how intelligence arises in nature and works logically, not just stacking more layers or parameters.”

One fascinating example of how the authors propose adding this complexity is in feedback loops, with “phase transitions” that would foster intelligent behaviors. The core idea is that these transitions can create new behaviors that were not originally present in the system’s individual components. Just as ice melts into water, so too could AI systems “evolve” to stable states.

“For AI, this could mean a system shifting from vague or uncertain outputs to confident, coherent ones as it gathers more context or feedback,” Wang says. “These transitions can mark a point where the AI truly ‘understands’ a task or pattern, much like human intuition kicking in.”