Amid the rise of billion-parameter models, I argue that toy models, with just a few neurons, remain essential—and may be all neuroscience needs.