Ubiquitous, not universal: the limits of AI scaling

Preface

The dominant narrative in 2024-2026 was exponential progress toward artificial general intelligence within a decade. This book measures the distance to that goal, identifies the constraints that bind, and projects what we will actually achieve.

The approach is simple: measure how much computation nature spent to produce general intelligence, measure how much computation we are spending on AI, identify the constraints that prevent closing the gap, and trace where efficiency improvements lead when capability scaling stalls.

The evidence suggests we will not reach AGI on the current path. Frontier capabilities stagnate at “impressively competent within distribution, fragile outside.” But while capability scaling fails, efficiency explodes: GPT-4 quality moves from datacenters to laptops, then smartphones, then all devices.