AI scale became the default because it worked. Add more parameters, more data, more compute — and capability followed, reliably enough that the approach became self-reinforcing. Performance on benchmarks improved. Funding followed performance. The systems got larger. What didn't keep pace was interpretability. Our ability to look inside and say, with any confidence, why a given output occurred. This wasn't an oversight so much as a trade-off that rarely got named as one. Bigger models were harder to examine almost by definition, and for most applications, examination wasn't the point.

CSHL researchers scaled down instead of up and found more. In collaboration with Carnegie Mellon University and Princeton University, CSHL Assistant Professor Benjamin Cowley trained a large AI model to predict how individual neurons in a macaque's visual cortex respond to natural images, then compressed it to roughly 1/1,000 its original size.[²] That compressed model still outperformed larger, state-of-the-art vision models at the same task — by more than 30%.[²] More importantly, it was now small enough to map: each artificial unit matched to a biological one, shifting the question from does the system classify images accurately to the more neuroscientific does each artificial neuron correspond to something real in the brain. That is a different ambition entirely, and one that does move at least in the direction of understanding.

What the model kept, after compression, is the revealing part. Stripped of everything nonessential, the model surfaces neurons that specialize in detecting dots. As Cowley puts it: "In the monkey's brain — and in our brains, too, most likely — there's a group of V4 neurons that love dots."[²] Dots are not trivial. They may be the first step toward recognizing a face — the foundation of social attention, reduced to its mathematical minimum. A model forced to be small converged on this not by design, but because it was what remained. Less like a black box. More like a microscope.