The notion that the secret to more effective artificial intelligence was always hidden inside the skull of a macaque monkey is subtly unnerving. Not figuratively, but literally, biologically, awaiting mapping and replication. In doing so, a group of researchers at Cold Spring Harbor Laboratory may have unintentionally opened a door that the field was unaware existed. They have now shrunk an AI vision model to the size of an email attachment.
The figures are almost humorous. There were 60 million variables in the initial AI model used to comprehend how primates perceive their surroundings. That figure decreased to about 10,000 after the team used compression methods modeled after the brain. One of the study’s lead authors, Ben Cowley, an assistant professor at Cold Spring Harbor, put it simply: “That is incredibly small.” He wasn’t exaggerating. To put things in perspective, contemporary large language models require data centers the size of warehouses in order to run on billions of parameters. This could be contained in a tweet.
The study, which was published in Nature in late February 2026, concentrated on a particular area of the visual cortex known as V4, which is a group of neurons in charge of processing curves, colors, textures, and what Cowley refers to as “complicated proto-objects.” It turns out that these neurons are remarkably well-organized. Each one specializes. Some react to edges, some to particular hues, and—possibly most oddly—some seem to be totally focused on dots.A subset of V4 neurons in the monkey’s brain—and probably our brains as well—love dots, according to Cowley. When you realize that the human eye is essentially a dot-detection machine that has been trained over millions of years of evolution, it almost seems humorous.
The research team collaborated with colleagues at Princeton University and Carnegie Mellon University. Together, they presented carefully selected natural images to macaque monkeys, such as birds, everyday scenes, and household items, and monitored which neurons fired in reaction to each image. Large AI models were then trained to predict those neural responses, and the models were methodically compressed without significantly compromising accuracy. Before compression, the resulting compact version outperformed rival systems by over 30%, and it continued to perform admirably even after being reduced to a small portion of its original footprint.
It’s possible that what the compression reveals rather than the compression itself is the greater story here. Cowley’s team discovered something that resembled biology more than traditional AI when they looked into the shrunken model’s internal operations. The compact neurons first deconstructed images into low-level features, such as edges, colors, and simple shapes. Each unit then combined that raw data in a unique way to create unique preferences. The majority of AI systems are not built with that architecture. It’s more in line with how a living brain functions, which begs the unsettling question: have we been creating AI the hard way all along?
The implications go far beyond visual processing, according to Mitya Chklovskii, a group leader at the Simons Foundation’s Flatiron Institute who was not involved in the study. He claimed that compact models inspired by biology could advance AI toward something more truly human-like. Despite their immense power, some researchers believe that current neural networks are still poorly understood—vast black boxes that generate results without providing much insight into the process. According to Cowley, “We’re very impoverished in our understanding of how these AI systems work, much like our own brain.” The new model begins to bridge that gap because it is small enough to dissect.
It is difficult to overlook the medical possibilities. Cowley has already started to consider the implications of creating AI models that mimic the brains of Alzheimer’s patients. Researchers may eventually be able to target and possibly rebuild those lost connections if they are able to model how synaptic connections deteriorate over time and determine exactly which visual stimuli cause neurons to communicate. For now, it’s speculative. However, as this develops, it’s hard not to feel that the goals of this research are far greater than the model itself.
The power required for a human brain is about 20 watts, which is less than that of a typical lightbulb. In the past, similar tasks have required thousands of times more from AI systems. The gap has always appeared to be an engineering ceiling or a basic restriction. Cowley’s team quietly and subtly suggests that there may not be a ceiling at all. Perhaps we were focusing on the wrong type of brain.
