Bostrom was asking the right questions in 2014. His Chapter 4 remains one of the best conceptual introductions to takeoff dynamics. But the empirical evidence has shifted the probability distribution toward slower, more continuous takeoff. The scaling laws (Kaplan 2020, Chinchilla 2022) show capability improves as a power law of compute – smooth and predictable, not discontinuous. AI Impacts found large technological discontinuities are historically rare across 37 trends. The actual trajectory (GPT-3 to frontier models) looks like Paul Christiano’s slow takeoff, not Yudkowsky/Bostrom’s FOOM. The hardware overhang has a real precedent (AlexNet 2012) but current trends favor ever-more-compute approaches. Fair verdict: these are possibilities worth considering, not predictions – which, to Bostrom’s credit, is largely how he framed them.
So what
Chapter 4 is a well-constructed philosophical analysis of possible takeoff dynamics. Its taxonomy (slow/moderate/fast takeoff) and vocabulary (optimization power, recalcitrance, hardware overhang) remain standard in the field. However, 12 years of evidence – scaling laws, AI Impacts’ discontinuity research, and the actual trajectory of AI development – have shifted probability toward slow, continuous takeoff rather than FOOM. The hardware overhang argument has a genuine historical precedent (deep learning 2012) but faces headwinds from current compute-hungry trends.
Verdict
MixedBag