Biological Computing Underhang

A provocative post by Elliot Callender titled 'Biological Computing Underhang' has gone viral on LessWrong, presenting a stark technical comparison between biological and artificial intelligence. The core argument is that the human neocortex is fundamentally limited by its serial processing depth—each cortical area can only perform about 14 sequential computational steps (equivalent to ReLU transforms in a neural network) per 4-millisecond corticothalamic cycle. In contrast, models like GPT-3 operate with 192 such layers and complete a forward pass in under 2ms. This creates a computational 'underhang' where, given sufficient data, artificial neural networks (ANNs) can be trained to strictly dominate human cortical capabilities, even with unlimited developmental time.

Callender extends this to the concept of 'primitives'—fundamental representations embedded during childhood critical learning periods, like a toddler learning linear algebra concepts that become wired into early visual cortex. Adults cannot establish new primitives with the same neural efficiency, making certain complex abstractions and their compositions biologically unattainable. The post raises a pivotal evolutionary question: if adding more cortical layers seems genetically simple, why are all mammalian brains stuck at six? The suggested answer points to metabolic constraints, error propagation challenges in biological networks, and the Pareto-optimal trade-off between depth and the heavy energy cost of neural interconnect. The implication is that AI's architectural freedom from these biological constraints may enable forms of reasoning and concept mastery permanently out of human reach.