The Grand Theory of Intelligence – Gwern

Gwern Branwen argues intelligence is search over Turing machines, with no master algorithm or fluid intelligence — just compute and learned special cases.

• Intelligence is search over Turing machines of varying lengths; no master algorithm exists, only a vast ensemble of learned special cases.
• Human variation in intelligence reduces to more compute for searching more Turing machines longer — no IQ gland or fluid-intelligence region in the brain.
• Large neural networks are just giant ensembles of small specialized models; a small model can always be distilled for any specific task.
• General-purpose learning is expensive, slow, and unreliable compared to hardwired gene-encoded solutions — which is why most organisms don’t evolve it.
• Human-level intelligence is rare to evolve because static niches favor hardcoded neural nets over costly general search.
• Short-lived organisms face an especially high barrier to evolving general intelligence since learning payoff requires enough lifespan to amortize it.

2024-11-21 · Watch on YouTube