Abstract: Artificial intelligence has acquired impressive capabilities over the past few years, mastering tasks that only a few years ago we thought only humans could solve. Yet, compared to the brain, striking limitations as well as profound differences are apparent. In this talk I will highlight two of these: 1) the great amount of prestructure in the early brain prior to the onset of sensory input, and 2) the strongly recurrent nature of cortical networks. I will describe work in a columnar visual cortex (ferret) showing how visual experience transforms a modular prestructure into reliable sensory representations. Network models suggest that this requires properly aligning the feedforward inputs with the recurrent networks, and I will present support for this hypothesis from intracellular recordings and widefield optogenetic manipulations. I will then show that a similar prestructure is found throughout the early cerebral cortex, suggesting a common intracortical mechanism, which we propose is ‘local excitation, lateral inhibition’. Moreover, after the onset of sensory experience, different cortical areas follow a similar path of refinement from modular to cellular-scale organization, pointing to a cortex-wide principle. Models identify an increase in net cross-inhibition, with strong excitatory–inhibitory coupling and weaker excitatory–excitatory connections, as the primary driver of this substantial reorganization. Single-cell optogenetics combined with network modeling confirms this network motif. Our results furthermore demonstrate that this motif acts as an amplifier when input signals are weak and switches to a discriminative regime when they are strong, which may support flexible behavior. These results highlight the role of network interactions in establishing reliable, yet flexible, mature representations, and point to design principles that current AI systems lack.