The team observed an abstract “pull” that lures neural networks into an optimal functional state, so they never stray far from their dedicated “set points.”