Ubiquitous, not universal: the limits of AI scaling
Valentin Radu · Zenodo (CERN European Organization for Nuclear Research) · 2026
Nature spent 10^25 to 10^40 learning instances to produce general intelligence. Current AI performs roughly 10^7 gradient updates. This book measures the eighteen to thirty-three order of magnitude gap, identifies the constraints that prevent closing it, and projects where efficiency improvements lead when capability scaling stalls.