The 20-Watt Constraint: Why the Brain Forced Computation onto Discrete Attractor Topologies, Not Gradient Descent

The human brain maintains flexible, generalizable intelligence on a continuous power budget of approximately 20 Watts. Modern deep learning architectures, by contrast, require megawatt-scale facilities to approach comparable inference capabilities. This energetic disparity is not merely a matter of efficiency—it represents an architectural incompatibility. If the brain operated on end-to-end continuous gradient backpropagation, it would have been metabolically eliminated by evolution. However, a central finding in neuroscience over the past fifteen years—that neural population activity is constrained to low-dimensional intrinsic manifolds—remains fundamentally descriptive. It tells us where the brain computes, but not how the search is executed. We propose that the brain's computational paradigm is algorithmically isomorphic to Discrete Structural Annealing (DSA): the state space is a low-dimensional discrete topological manifold, the search is executed via discrete operators (Swap, Reverse, Mutation), and the search dynamics are driven by a "temperature" schedule regulated by neuromodulators. We present five falsifiable neuroscientific predictions, each detailing a specific causal experimental design. If validated, these predictions will shift neuroscience from describing manifolds to causally manipulating them. This is a roadmap for high-impact discoveries, charting a search space that can be empirically terminated or confirmed.