To say that language is shaped by characteristics of the brain implies that to understand language evolution, we need to know how the brain's capabilities and limits apply to language. However, Christiansen & Chater (C&C) say little more than that the brain had developed the needed perceptual, motor, and general cognitive abilities before language emerged. In their own research and theorizing, C&C have steadfastly objected to the arguments of Chomsky, Pinker, and other researchers that seem to suggest that calling something we cannot explain “innate” excuses us from trying to understand how it works (grammar is an instinct – duh!). Yet C&C seem to have done something close to that here, saying that language evolution is shaped by the brain without telling us how. This could be a strategic divide-and-conquer step, but its use by the authors without explanation is disappointing. I take a stab here at rectifying the omission in a way that leaves the main idea intact: hypothesizing an important quantitative brain evolution that would have been a necessary enabler of cultural language evolution.

There has been dramatic progress in the last few decades, some of it attributable to Christiansen and Chater in various works, in creating and evaluating computational mechanisms that do much of what human brains do in acquiring and using language based solely on the kinds of language experience that all humans have. These models offer empirical examples of fundamentally simple mechanisms that learn to do things that have often been declared impossible to learn. I focus here on Latent Semantic Analysis (LSA; Landauer & Dumais Reference Landauer and Dumais1997), and mention two others, the Topics model of Steyvers and Griffiths (Reference Steyvers, Griffiths, Landauer, McNamara, Dennis and Kintsch2007) and the BEAGLE model of Jones and Mewhort (Reference Jones and Mewhort2007), which employ different means to achieve similar results in mimicking human use of many tens of thousands of words to convey meaning. Then I add some extrapolative conjectures on how these successes suggest ways in which culturally steered (Baldwinian) genetic evolution could have played a crucial role in language evolution by expanding something already important.

These models all learn to simulate human language phenomena solely by mathematical analysis of large-text corpora that closely resemble what average literate people have read. They all eschew any direct aid from language experts – such as grammatical annotations or semantic hierarchies – that must themselves be results of the causal mechanisms sought. Each in its own way achieves a representation of words and passages as numerical vectors that obey the compositional constraint that every meaningful passage be a combination of components chosen from the same set; for example, sets of English words or Chinese characters. All three models closely simulate human performance in language tasks, such as answering synonym tests and mimicking free association. For more examples, LSA learns word meanings at the same daily rate as children; and has come close to equaling humans in judging coherence of book chapters, the topical similarity of conversational utterances, the semantic similarity of test essays, and in identifying translated documents in widely differing languages when given only partially overlapping experience with them. (For more on these examples and many others, see Landauer et al. Reference Landauer, McNamara, Dennis and Kintsch2007.) Lest a reader be tempted to discount these achievements because LSA almost completely ignores grammar (reflecting an apparent assumption of grammar's near equivalence with language in the target article and elsewhere in the innateness debate), consider the following two observations: (a) in a study in which two independent expert readers and LSA each estimated the quality of substantive content of essays, the mutual information between LSA and humans was 90% of that between two humans; and (b) the possible information in a typical sentence attributable to word choice is about nine times that of word order (Landauer Reference Landauer and Ross2002).

Now for the conjecture on how Baldwinian evolution may have been essential to language. The fundamentally simple mathematics by which LSA learns and uses its word and passage vector representations requires a representational space of about 10¹³ bits.Footnote ¹ (The other models make similar demands.) This is some 10,000 times the amount of learned information accumulated in a normal human lifetime (Landauer Reference Landauer1975). Of this, more than one-tenth is knowledge of word meanings. By this analogy, the space needed to support language is tauntingly similar to the increases in primate and human cortex size over the probable period of development of gesture, proto-, and fully Homo sapiens sapiens languages (Deacon Reference Deacon1997). Acknowledging some unresolved questions in this arena, it appears that language may have simply needed much more relevant brain space. Importantly, the new space was most likely the same kind of space – analogous to increasing general-purpose memory capacities in a computer – not something wonderfully novel like Universal Grammar (UG).

So, suppose that the explosive human cortical growth started with a mutational decrease in the usual amount of death of neonatal neural cells, and this let some early male get a tiny bit better at composing multi-utterance mating calls, which let him sire more people like him … The resulting genetically adapted brain size growth would be no more mysterious, or take longer to happen, than the differentiation of Great Danes and Chihuahuas; and it could have eventuated in great vocabularies that made possible great minds that supported intelligent cultures that in time invented many grammars whose possibly equally simple underlying mechanisms we have not yet discovered. Learning how every one of more than 100,000 words relates to any other word, and how their combinations yield meaning, appears as complex and difficult to me as learning a grammar, yet succumbs to mathematical simulations, given enough space.