There is no single agreed definition of intelligence. Psychologists, neuroscientists, and machine learning researchers each carved their own, and the disagreements are real. But the threads overlap in interesting ways — and the overlap is where AGI conversations should start.
The psychometric tradition
Modern psychology grew up measuring intelligence with tests. Charles Spearman noticed in 1904 that performance on very different cognitive tests correlated, and named the shared factor g. A century of follow-up work — Cattell, Horn, Carroll, the CHC (Cattell–Horn–Carroll) synthesis — broke g into fluid reasoning, crystallised knowledge, working memory, processing speed, and several narrower abilities.
The lesson is empirical: human cognitive performance is not infinitely many independent skills. A small number of latent factors explain a surprising fraction of the variance. Whether that generalises to machine minds is an open question.
The neuroscience view
Neuroscience tends to define intelligence functionally: the brain's ability to model its environment well enough to act effectively in it. Efficient coding, predictive processing, and the free energy principle all describe the brain as a prediction machine that minimises surprise.
On this view, intelligence is less about scoring on tests and more about how compactly and accurately a system represents the world.
The machine learning view
ML researchers, especially in the AGI tradition, define intelligence in terms of generalisation across tasks. Legg and Hutter (2007) proposed an influential formal definition: an agent's intelligence is its expected performance across a wide distribution of environments, weighted by simplicity.
Closely related ideas — compression, Solomonoff induction, MDL — frame intelligence as the ability to find short descriptions of complex data. ARC-AGI and similar benchmarks try to operationalise this as few-shot abstraction.
Where the definitions converge
All three traditions agree on a small core: intelligence is about turning experience into useful internal models, and using those models flexibly in novel situations. Sample efficiency, transfer, abstraction, and goal-directed action recur everywhere.
Where they diverge is on whether subjective experience matters, whether language is necessary, and whether intelligence is one thing or many. None of those disagreements have been settled.
Key terms
- g factor
- The shared general factor across psychometric tests; named by Spearman (1904).
- Fluid vs crystallised
- Cattell's split between reasoning on novel problems and accumulated knowledge.
- Predictive processing
- The brain treated as a hierarchical prediction system that minimises prediction error.
- Legg–Hutter intelligence
- Expected reward across all computable environments, simplicity-weighted.
- Few-shot abstraction
- Inducing a rule from a handful of examples; central to ARC-AGI.
Connects to AGI
When people argue about whether a system is 'really' intelligent, they are almost always using different definitions. AGI specifically targets the ML view — wide generalisation across tasks — but cannot ignore the others if it wants to be useful to humans.