In Agnostic's post below about the recent results with starlings bearing on language evolution, I left a comment that ended up veering somewhat off topic. But the result was the coalescence of some thoughts of mine concerning the nature of g that I felt were worthy of reproducing on the front page. Please note that what follows is my own possibly idiosyncratic perspective and does not reflect any widely held consensus. Indeed, it is probably in tension with a significant ongoing research effort devoted to explaining g in terms of a purely psychological contruct known as "working memory." But in any case I hope that it is useful to you in your own thinking about this most mysterious of constructs in the human sciences.


g as a construct, strictly speaking, probably does not correspond to some kind of all-purpose mental workspace where problem elements from any domain can be placed, maintained, and manipulated in the service of adaptive behavior. It is true that many people do construe g in this way, and it may turn out that g is in fact something like that, but I seriously doubt it. Rather, g should be thought of as a source of variance in all mental ability tests, its nature being indifferent to the specific contents of the test and the cognitive processes required by the solution of the test items. Thus, I think, g will prove to be not software but hardware.

On this view it not meaningful to apply a g-like construct to whatever general-purpose cognitive processes that these starlings may be using to distinguish these sequences. You would only be able to do that if you put these starlings through a battery of diverse cognitive tasks (preferably calling upon different a priori posited components in the box-and-arrow diagrams that cognitive psychologists like to draw up) and found that a starling that did well on one task tended to do well on all of them and a starling that did poorly on one tended to do poorly on all. And what if this sequence distinction showed a large loading on the g extracted from this hypothetical positive-manifold correlation matrix? It would not exactly be a mistake to say that sequence distinction captures the essence of g, but there would be a much better way to put it given what we currently know. I would posit that sequence distinction samples so many brain regions or cognitive processes that the pervasive hardware features of the brain giving rise to g (number of neurons? quality of myelination? efficiency of chemical transmission at the synapses? we know that IQ is positively correlated with both gray and white matter volume) gets aggregated at the expense of specific sources of variance particular to the circuit that executes this task.

Consider reaction time. We know that reaction time is negatively correlated with IQ but not that highly; clearly, there is a lot of specificity in a RT task. We can think of the circuit that a given person uses in a RT task as being extremely simple (I have seen MEG results showing that choice RT in the Hick paradigm doesn't even lead to activation in the frontal lobes) and varying greatly from person to person because of idiosyncracies in sensory induction, motor response, and the stochastic wiring of the circuit itself. Thus, the ratio of variance attributable to consistent hardware components that affect efficiency of processing to variance attributable to idiosyncratic properties of the circuit is relatively small. But what about highly g-loaded tasks such as reasoning through matrix patterns, learning the meanings of words, or decomposing visual 2D gestalts into tangible 3D components? The circuits that carry out these tasks probably still vary from person to person but are surely much, much more complex; under MEG or fMRI the whole brain probably lights up. Thus, the circuit samples g at so many points that specificity gets largely drowned out.

Well, what would we see if we put starlings through a whole battery of cognitive tasks and looked at the correlation matrix? Would we get positive manifold? A high first:second eigenvalue ratio? You know, I actually bet we would. But would sequence discrimination show a large loading on the extracted g? An interesting question.