Chapter 2 — Paths to Superintelligence¶
Core Thesis¶
Multiple independent roads could lead to superintelligence — AI, whole brain emulation, biological cognitive enhancement, brain-computer interfaces, and networked collective intelligence. Multiple paths raise confidence the destination is reachable at all, even though they don't necessarily converge on the same outcome, timeline, or degree of human control over what results.
The Mechanism — Seed AI and Recursive Self-Improvement¶
Turing's 1950 "child machine" — a simple system that learns rather than one pre-loaded with adult competence — is the ancestor of Bostrom's key concept: seed AI, a system sophisticated enough to improve its own architecture. Recursive self-improvement means an early version designs a better version, which (being smarter) designs a still-better version — a process that, under the right conditions, compounds into an intelligence explosion, taking a system from sub-human in most respects to radical superintelligence in a short span. The evolutionary argument for AI's feasibility (Chalmers, Moravec) is weaker than it looks: Box 3's back-of-envelope calculation shows brute-force recapitulation of evolution's search for intelligence would cost 10^31–10^44 FLOPS — decades beyond even continued Moore's Law — though a human-guided search could plausibly cut this by unknown (and unboundable) orders of magnitude.
Whole Brain Emulation — Plagiarism, Not Design¶
Emulation ("uploading") sidesteps understanding cognition altogether: scan a brain, reconstruct its neuronal wiring, run it as software. It needs no theoretical breakthrough, only three enabling technologies — scanning, translation, simulation — pushed further than they've ever gone. C. elegans, a 302-neuron roundworm, has had its full connectivity mapped since the 1980s, yet nobody has emulated even it, because connectivity alone (without synaptic strengths and dynamics) isn't enough. The scale-up ladder (worm → honeybee → mouse → monkey → human) means emulation, unlike AI, should arrive with visible warning — unless the last missing piece turns out to be neurocomputational modeling itself, in which case success could snap from total failure to working emulation with no intermediate warning at all, "like a grand mal seizure" resolving into wakefulness.
Biological Cognition — Slow but Certain¶
Selective embryo selection already delivers measurable gains (Table 5: 1-in-10 selection ≈ 11.5 IQ points; iterated across generations, gains compound past diminishing returns). Iterated embryo selection — deriving new sperm and eggs from embryonic stem cells to compress ten generations of selection into a few years — could in principle push a population toward or past the highest human IQ ever recorded. But generational lag (20+ years for a selected embryo to reach productive adulthood) means this path is inherently slow, and Bostrom is blunt that humans are "probably... the stupidest possible biological species capable of starting a technological civilization" — first through the door, not optimally built for it.
Why Not Cyborgs¶
Brain-computer interfaces look weak as a superintelligence path. The retina already transmits ~10 million bits/second; the bottleneck in human cognition isn't data bandwidth in, it's the neural machinery needed to make sense of data — and upgrading that wholesale is just artificial general intelligence by another name. "Downloading" thoughts directly between brains fails for a deeper reason: brains don't share a standard data format; each brain represents concepts idiosyncratically, so a working brain-to-brain interface would itself need to be an AI-complete translator. Implants help genuine disability (Parkinson's stimulation, locked-in-syndrome cursors) far more plausibly than they'd help a healthy brain get smarter.
Networks and Organizations¶
Collective intelligence — writing, printing, prediction markets, reduced communication overhead — has already raised humanity's problem-solving capacity across history, and could in principle produce a "collective superintelligence" without any single mind becoming smarter. An "awakening" Internet is not spontaneous magic but the endpoint of ordinary incremental engineering (better search, filtering, autonomous agents) that, if it converges on a unified system, collapses back into the AI path anyway.
Key Terms¶
- Seed AI — a system capable of improving its own architecture
- Recursive self-improvement — each smarter version designs the next
- Whole brain emulation — scanning and running a brain's structure as software
- Iterated embryo selection — compressing multiple generations of genetic selection via stem-cell-derived gametes
- Collective superintelligence — superhuman performance from networked, not individually enhanced, minds