Study Guide: Chapter 4 — The Kinetics of an Intelligence Explosion

Core Idea

Rate of intelligence increase = optimization power / recalcitrance. Recalcitrance plausibly falls right at human-parity (content overhang, hardware overhang, threshold effects), and optimization power plausibly rises (attention, funding, competition) — so if a takeoff happens, fast/moderate is more likely than slow, despite zero historical precedent for change this rapid.

Key Terms

Optimization power · recalcitrance · slow/moderate/fast takeoff · crossover · hardware overhang · content overhang · algorithm overhang

Case Summary

A system with human-level reading comprehension at machine reading speed could digest the Library of Congress in weeks — instantly "weakly superintelligent" through content alone, no new algorithm needed. Box 4's math: constant effort + hyperbolically falling recalcitrance → thousandfold capability growth within 18 months of crossover.

Application Checklist

• [ ] Separate "when will X reach parity" from "how fast does X exceed parity once reached" — different questions, different answers
• [ ] Check for overhangs (compute, pre-made content, dormant algorithms) that let a system jump capability cheaply
• [ ] Watch for anthropomorphized intuitions that compress "dumb-to-human" into a wide gap and "human-to-superhuman" into a narrow one
• [ ] When a system crosses from external-driven to self-driven improvement, expect the growth curve to bend sharply

Self-Test

1. Why does Bostrom think recalcitrance might fall, not rise, right around human-level machine intelligence?
2. What is the "crossover," and why does it matter for whether growth becomes exponential?
3. Why is the "fast takeoff has zero historical precedent" objection insufficient to rule it out?