Systems neuroscience—from Lashley's distributed engrams through Pribram's field-based processing to Freeman's oscillatory dynamics—has long argued that intelligence is a whole-brain property requiring feedback-driven computation. We formalize this tradition using Reinforcement Learning and Approximate Dynamic Programming (RLADP) and propose that vertebrate intelligence falls into four qualitatively distinct levels—rodent, primate, human, cetacean—each defined by a different architecture for generating and propagating backpropagated feedback signals. The transition between levels is not parametric but architectural, and each architecture demands a different energy strategy. A conserved allometric rule for cortical ion channels holds across nine of 10 mammalian species, fixing the biophysical cost of computation per unit volume; human neurons uniquely violate this rule, reducing channel density to redirect energy toward long-range white matter connectivity. We show that white matter is an active communication system whose superlinear scaling creates a geometric cost trap, that the corticothalamic loop provides master timing for forward-backward cortical processing cycles, and that timing degradation causes qualitative intelligence failure. The biological strategies cataloged here—from selective connectivity reduction to cellular energy reallocation to cortical reorganization—have parallels with the communication-energy wall now constraining artificial intelligence.