Jablonka & Lamb's (J&L's) book Evolution in Four Dimensions (Jablonka & Lamb Reference Jablonka and Lamb2005) exposes many tangled connections between genome, behaviour, and environment, but it skims over gaps in our knowledge about the information-processing capabilities underlying observed behaviours – ignoring important mechanisms with epigenetic features. Much is said about the physical and chemical mechanisms involved in development, but behavioural competences are described mostly from the outside. Explaining the internal information processing requires the design stance (Dennett Reference Dennett1978).

External behaviours of many animals indicate that they have mechanisms concerned with internal symbolic competences, required for perceiving or acting in structured situations, including planning, predicting, identifying information gaps to be filled, formulating goals, executing plans, learning generalisations, and creatively combining different competences. We need to explain what these competences are, what mechanisms make them possible, how they develop in individuals, and how they evolved. Such competences (in humans and other animals) seem to presuppose something like internal symbolic languages with very specific properties.

When the variety of structurally different combinations of situations and goals rules out preconfigured responses, animals need the ability to represent and make inferences about existing and future configurations and changes; for example, configurations of a partially constructed nest made of interlocking twigs and the affordances (Gibson Reference Gibson1979) for inserting the next twig. This requires internal formalisms for representing structures and possible processes and for constructing, comparing, and planning, including selecting actions from branching collections of possible future sequences. Later, the animal has to produce the actions under the control of the representation. So action sequences linked to complex internal symbolic structures occurred before external linguistic behaviour evolved. Animal behaviours demonstrating such competences include tool-related behaviours (Kacelnik et al. Reference Kacelnik, Chappell, Weir, Kenward, Wasserman and Zentall2006) and the remarkable symbolic competences of the grey parrot Alex (Pepperberg Reference Pepperberg, Balkenius, Zlatev, Kozima, Dautenhahn and Breazeal2001).

Our epigenetic hypothesis about how information-processing develops under the influence of the environment avoids two extreme theories: (1) that all animal competences are somehow encoded separately in the genome, possibly in a large collection of innate modules, and (2) that a small collection of general learning mechanisms (e.g., reinforcement learning) is genetically determined, and everything else is a result of applying those general learning processes. Our “middle way” also synthesizes two apparently opposed views, one expressed by Karmiloff-Smith (Reference Karmiloff-Smith1994, p. 693): “Decades of developmental research were wasted, in my view, because the focus was entirely on lowering the age at which children could perform a task successfully, without concern for how they processed the information”; the other by Neisser (Reference Neisser1976, p. 8): “We may have been lavishing too much effort on hypothetical models of the mind and not enough on analysing the environment that the mind has been shaped to meet.”

What an individual can learn often changes dramatically during its life, indicating a cascaded development of competences partly under the influence of the environment, including competences to acquire new competences (metacompetences), some of which are themselves the result of interaction of earlier metacompetences with the environment. We summarise this relationship in Figure 1, showing multiple routes from the genome to behaviours of various sorts, with competences at different levels of abstraction and different sorts of specificity developed in different ways at different stages. This implies that learning in some parts of the brain is delayed until others have acquired a layer of competences to build on. So if prefrontal lobes are associated with processes further to the right of the diagram, occurring only after many cycles of simpler development, we would expect prefrontal lobes to develop after low-level visual and motor control mechanisms. Evidence consistent with this conjecture has recently been reported in human infants by Gilmore et al. (Reference Gilmore, Lin, Prastawa, Looney, Sampath, Vetsa, Knickmeyer, Evans, Smith, Hamer, Lieberman and Gerig2007).

Figure 1. The environment (including the body and new brain states) can affect all the processes. There are multiple routes from genome to behaviours, some used only after others have produced new competences and metacompetences. (Based on Chappell & Sloman Reference Chappell and Sloman2007)

J&L discuss the evolution of language and, like many others (e.g., Arbib Reference Arbib2005), focus mainly on external language used for communication. This assumes that first there were simple forms of language (e.g., gestures and sounds), and complex forms evolved later. In contrast, we suggest that language first evolved for “internal” use. Because some people restrict the label “language” to symbol systems used for external communication, we use the term g-language (generalised language) to refer to a wider class that includes internal languages. A g-language allows rich structural variability of various kinds as well as compositional semantics for dealing with novel configurations of objects or processes.

Most people assume that language started simple and external and then grew more complex externally before being internalised. We, like Bridgeman (Reference Bridgeman2005), suggest that complex g-languages evolved in many non-human species and also develop in young children, who cannot yet talk. Internal g-languages are needed to provide forms of representation of current and possible future situations and processes that allow wide structural variation in what is represented, with compositional semantics to cope with novelty (Sloman Reference Sloman, MacCafferty and Gray1979). So, rich internal g-languages are precursors to external human languages both in evolution and in child development. After g-languages had evolved for other purposes, including constructing plans that were used to control behaviour, some animals may have started mapping their internal structures onto external behaviours for communication purposes.

Insofar as animals and children can look at different parts of a scene and combine information from most recent saccades with information about parts of the scene that are no longer in view, when planning what to do, they must use representations of spatial organisation of information as well as temporal organisation. In some ways, this requires more complex forms of representation than human spoken languages, combining aspects of verbal language and pictorial languages (analogous to maps, diagrams, and drawings; see also Trehub Reference Trehub1991).

G-languages probably evolved for internal information processing and control of behaviour (through the generation of goals, plans, or instructions), along with generation of questions to specify missing information, and perhaps to formulate hypotheses, explanations, and suppositions. External human language (spoken and gestural) and other symbolically based aspects of human culture (e.g., music, mathematics) also might have built upon these preexisting internal symbolic foundations.

Eventually, instead of a specific g-language, evolution produced competences to acquire a variety of g-languages expressing different kinds of information. This implies that some nonhuman animals' behaviour will be directed and shaped by their internal g-languages, which in turn are shaped by the structure of the external environment, directing evolution down particular paths, and perhaps causing “convergent” evolution of closely related cognitive abilities in birds and mammals with overlapping perceptual and manipulative competences.

If abstract and complex g-language constructs have to be learnt at a late stage of development, but are particularly useful to a species, then some of them could become genetically assimilated or accommodated; in which case they will themselves become heritable and can direct development in particular ways. Environmental cues encountered by these animals will be filtered through their cognitive architecture, thereby tightening the knots between the genome, behaviour, and the environment. Chappell and Sloman (Reference Chappell and Sloman2007) suggest that this employed a separation between parts of the mechanism producing a general class of behaviors and parts that provide parameters that select from that class. The generic competence and the particular parameters might undergo separate trajectories in evolution and development.

If J&L's “assimilate-stretch” principle were extended to cope with the evolution and development of internal g-languages and associated mechanisms, this might be a significant, previously unnoticed, factor in the evolution of cognition. Their examples suggest that assimilate-stretch extends behaviour additively. But qualitatively new capabilities might emerge. For example, if a learned capability becomes genetically assimilated or accommodated, it could form a building block for qualitatively diverse competences. Information that some objects can be deformed by manipulation, can be broken into smaller pieces, can be inserted into spaces, and can, if appropriately assembled, produce fairly rigid structures, might form fundamental parts of a very complex collection of learnable competences, including constructing nests, making or using tools, or extracting objects from containers.

The ideas in this book may turn out to have far-reaching significance for many disciplines. We have tried to show, briefly, how some of that could affect studies of cognition, and internal g-languages, with implications for the evolution of language and many forms of learning. As our cited paper indicates, these forms of development may be required also for intelligent robots that are learning to cope in a wide variety of environments.