The Futile Quest for Artificial Intelligence (2006) (archive.skeptic.com)

Peter Kassan’s 2006 critique argues that repeated claims of “human-level” AI are premature. Using Jeff Hawkins’ high-profile brain-modeling announcement as a springboard, Kassan traces AI’s history and fault lines (connectionism, computationalism, robotics) to show that modeling the brain is far harder than enthusiasts assume. Neuroscience lacks a unifying theory: estimates of brain anatomy keep changing (cortex ~30 billion neurons; “about a thousand trillion” synapses), and previously ignored components like glial cells have active roles. Simple examples underscore the gap: C. elegans’ ~300 neurons and ~7,000 synapses remain poorly modeled, while typical artificial neural networks use dozens to a few hundred artificial “neurons” and at most millions of weighted connections—orders of magnitude smaller and far simpler than biological reality. Kassan highlights concrete technical roadblocks: realistic synapses may require thousands–millions of parameters and millisecond-to-submillisecond dynamics, implying astronomical data and compute; if each synapse were represented compactly, the code/data would dwarf today’s largest software projects (he estimates simulation scale millions of times larger than Windows). Testing is essentially impossible due to combinatorial explosion of inputs and emergent, un-specifiable behavior in connectionist systems. Moore’s Law won’t magically close these gaps (software scales differently), so Kassan concludes that brain-based routes to AGI are currently infeasible and that AI research should temper grand claims and confront these fundamental engineering and scientific limits.