The Evolved Apprentice

Two Framing Ideas About Human Evolution.

Human evolutionary change has been rapid and extensive; so much so that the genetic similarity and recent divergence between the human and the chimp lineages came as a profound surprise. Three million years ago humans were relatively minor elements of a rich East African mammalian fauna. Since then, our lineage has expanded  geographically, demographically and ecologically. Over roughly the same period, our lineage has experienced an explosive increase in co-operation. We are the only large mammal that depends for essential resources on co-operation with non-relatives. Likewise, tool-use. Beginning about 2.5 million years ago, we became obligate technovores, with the pace of innovation picking up over the last 200,000 years. These changes have been accompanied by others in morphology, life history and family organization. We are not what we used to be. Tellingly, this pattern has not been mirrored in other lineages, as it would be if this trajectory had an external cause. So a first framing idea is that human evolutionary change has been self-generated through positive feedback. Specifically: a feedback loop driven by the increasing complexity of human social environments, and by the problems this complexity causes for co-operation management.

The second idea is that human cognitive competences have a high information load. Ordinary adaptive decision-making depends on an agent being sensitive to complex, subtle features of their environment. The information-hungry nature of human action first became apparent in thinking about language, but other competences turn out to be information-hungry as well. For instance, in much social interaction we are effortlessly tuned to the moods, intentions and beliefs of others. If you have made some horrendous social blunder, it is almost impossible not to be aware of it. Yet since others obey norms of politeness in such situations, this awareness often depends on sensitivity to small cues of posture, expression, tone of voice; shifts in conversational focus and so forth. As with language, so with the skills of social navigation, bargaining, co-ordinating effective plans, acquiring and using technology. Almost all of us acquire these skills, but they depend on acquiring and deploying significant informational resources.

Both these framing ideas are on the money. Positive feedback loops are indeed critical in driving human evolution, and much routine human action has a high information load. But I am sceptical about the way these framing ideas have been developed, and in the rest of this essay I explain that scepticism and gesture towards an alternative.

Co-operation, Defection, Co-ordination.

Increased levels of co-operation in early humans was probably triggered by some external change. Our environments became more seasonal and open, and perhaps that exposed our ancestors to more predators, so they had to act together to survive. But once co-operation became a key feature of human lifeways, it induced an arms race. Agents acting together can secure resources, construct technologies and defend territories that would be utterly beyond any lone wolf. But often the profit of co-operation is not contingent on everyone paying its full costs, and hence there is a temptation to secure a share of co-operation’s profits while avoiding its costs. Co-operation is thus risky as well as profitable; and so co-operators must be vigilant. Vigilant co-operation selects for intelligence. But as others become more intelligent, defection threatens to become increasingly well disguised. As co-operating groups became built from increasingly intelligent agents; as co-operating groups become increasing differentiated through division of labour; and as they become larger; effective vigilance became ever more difficult. Hence selection for vigilance selects for still higher intelligence, establishing a positive feedback loop between individual intelligence and social complexity.

This picture identifies co-operation management as the key to human cognitive evolution. Each agent has an interest in maximising return from co-operation while minimising investment in co-operative activities. This sets up a vigilance requirement that becomes steadily more challenging. However, while stable co-operation depends on policing defection, in small scale social worlds, this picture overstates the problem of identifying cheats, and understates the informational challenge of co-ordination; of making co-operation work. In a village, everyone knows who the bastards are. The ancient policing problem is not one of identification but motivation; of risking confrontation when they are bad bastards. Once something like language evolved, bastards were fingered by effective gossip networks. But they may have been outed even earlier, by an ancient feature of human economic life: our evolution as social foragers. There has been a revolution in human foraging. Our lineage has evolved from being omnivorous marginal scroungers to being dominant predators; from being predator targets to taking the food from other predators.

There is much controversy about this process: about its timing; about the relative importance of scavenging and hunting; about female co-operation and the role of plant foods and small game in human diets; about the invention of cooking and its role in human evolutionary history. But by 400,000 years ago, humans had evolved into social foragers exploiting high value, heavily defended resources. Large and medium size game animals had become a key resource. Large herbivores (buffalos, for example) are neither easy nor safe to kill, especially with short-range technology. So this foraging mode depended hunting peoples having intimate knowledge of their targets and the local terrain. Social foraging requires information: and the more ambitious, profitable, or varied the targets, the more information it requires. It depended on technology, and the skills to use it expertly. And it depended on co-ordinated co-operation; hunting large animals with short-range weapons poses a formidable co-ordination challenge. Lone hunters or small parties can kill large animals once lethal standoff technology has been invented: spear-throwers; bows; poison darts. But 400,000 years ago, these innovations lay in the distant future.

Thus generating the profit from collective foraging was informationally demanding. But identifying the shirkers who threatened that profit was not. Killing safely with stone-tipped spears required groups to act together in co-ordinated yet flexible ways, in conditions of stress, danger and time pressure. Such conditions generate information: agents leak information about their character, judgement and capacities when engaged in intimate, high-stress collective activity, especially when the activity persists for days. On a three-day hike, you find out a lot about your companions, especially if you get lost or the weather turns foul. That is true even when you have modern equipment and face no real danger. The greater the stress, discomfort or danger, the more you and they learn. Collective foraging would frequently involve stress, discomfort and danger, and shirkers and cheats would unambiguously identify themselves. Our ancestors did not need the social sensitivity of Jane Austen to know with whom they were dealing. But they did need to be informed and intelligent to act together to extract resources from a recalcitrant and dangerous world.

Modularity, Novelty, Expertise.

Foraging illustrates the information-hungry character of human skills. Famously, in theories of human evolution, information-hunger has been linked to a modularity hypothesis. Adaptive response to our complex environment depends on innate, evolved, special purpose cognitive mechanisms, for it is only such mechanisms that enable us gather and deploy the information on which action depends. In turn, the modularity hypothesis presupposes that information demands are discrete and predictable. So while ordinary human action involves solving problems with a high information load, both the problems, and the information  needed to solve them, are relatively constant over evolutionary time. Hence we get the famous “massive modularity” hypothesis, and the idea that our minds are ensembles of innately equipped special purpose devices; devices which adapt us to the challenges posed to our foraging Pleistocene ancestors; challenges which largely persist today. Our minds are integrated arrays of devices each of which solves a particular problem with remarkable efficiency. These devices are efficient because they come pre-equipped with much of the information they need. We are born knowing what human languages are like, what human minds are like, what human social worlds are like. Children need to learn the specific moral norms of their community. But they do not need to learn what a moral norm is.

This nativist explanation of information-hungry competence presupposes that the information structure of human selective environments is stable. If the information a child needs about her world is stable over evolutionarily significant time frames, selection can build that information into human minds. Not otherwise. Some domains are stable. Material technology is a plausible example of a discrete, stable, and fitness-critical domain, for the causal properties of sticks, stones and bones do not change. However, many central aspects of the human world have changed fundamentally. Think, for example, of the Holocene Revolution. Over the last 10,000 years, humans became sedentary rather than transient. Most abandoned foraging for other economic activities, with much co-operation mediated by market mechanisms. We began living in much larger, much more stratified groups, with formal institutions and top-down decision making. Sexual relations changed, with reproductive skew becoming an important factor in many human cultures. Technology, including specialised information technology, elaborated, and the pace of technological change ultimately quickened so that real change took place within a single generation. We live in a new world (as did many of our more recent ancestors). If we really had stone-age minds in an electronics-age world, we would be crippled by adaptive lag. But we remain competent in responding to many of these novel challenges; for example, most of us work competently in formal institutions. The modularity hypothesis embraces the centrality of learning, but modules channel learning. To the extent that the modularity hypothesis explains competent response to information-hungry problems by appeal to pre-loaded information, it is poorly posed to explain competent response to evolutionarily novel challenges.

Competence in the face of evolutionarily novel problems depends on skill. Skills are phenomenologically akin to modules: they are fast, automatic, and task specific. Without the special training of a professional linguist, it is impossible to hear speech in a familiar language as mere noise: I cannot but hear English speech as words, sentences, conversations. Likewise, I cannot but see written English words as words; I cannot see them as mere shapes. But reading is a response to a novel feature of the world; language in a new medium. Moreover writing makes a new kind of  communication possible:  decontextualised, often one-off communication with strangers, sometimes displaced in space and time. Once the capacity is acquired, reading seems as natural as listening. But like many skills, reading depends on a long learning history in organised developmental environments. The problem is to characterise this environment and explain its evolution.

The evolved apprentice.

On both the modularity hypothesis and the apprentice learning alternative, human minds are evolved learning machines. But on the apprentice model, our minds are adapted to evolutionarily salient channels, sources and contexts of learning (and teaching), as much as (or more than) our minds are evolved to learn about specific factual domains. Human lifeways came to depend on social learning, for co-operative foraging depends on synthesising social, ecological and technical information. So successful social foraging — especially in high risk environments — depends on each partner having a well-tuned sense of the skills and intentions of the others. It depends  on communication and planning. But it also depends on rich, accurate local knowledge; on a detailed grip of the natural history of target species; on locally made technology used with great expertise. No generation acquires these informational resources from scratch: the cognitive capital on which successful foraging depends is acquired by cross-generation information pooling. The informational resources one generation inherits will be modified (as conditions change and through innovation) and transmitted for further modification to the next.

High volume, high fidelity, inter-generational cultural learning coevolves with social foraging. There is feedback. As the fidelity of social learning improves, social foraging becomes more effective, for technology and technique improve across the generations. As social foraging becomes more profitable, adults can more effectively support the next generation while they acquire skills and information. This loop depends on the fact that humans organise the learning world of the next generation. Humans (like many other organisms) modify their own environment. One important form of human niche construction is informational engineering. Humans of one generation act in ways that transform the learning environment of the next generation. Cultural learning is obviously central to human social life. But most cultural learning is hybrid learning; it is culturally enhanced trial and error learning. Very few humans acquire significant life skills just from instruction and demonstration; very few learn skills by unassisted exploration. Human children explore and experiment on their physical and social environment. But they often explore environments which have been shaped to make it easier or safer for them to acquire critical capacities.

Apprentice learning is a good model of this form of cultural learning. Apprentice learning is a very powerful mode of social learning, making possible the reliable acquisition of complex and difficult skills. Apprentice learning is hybrid learning: an apprentice combines information from the social world with information from the physical-biological environment. It is learning by doing, but in supervised and enriched learning environments. Apprentices are assigned tasks that are both appropriate to their current skill levels, but which stretch them, and which scaffold the acquisition of new skills. In their work they are surrounded by props: tools; completed and partially completed artefacts; raw materials in various stages of preparation; errors — examples of what can go wrong. Often they have sources of advice and demonstration, for learning is often social and collaborative. Apprentice learning depends on individual cognitive adaptations, including those for social learning,  but it depends as well on these adaptively structured learning environments.

In my view, craft expertise — skill sets of the kind forager lives depend on — are fine-tuned at a generation, and reliably transmitted across generations, by this mode of organised human learning environments. Novices learn by doing in an environment seeded with informational resources and with their learning trajectory partially organised by experts. The expert organize the trial and error learning of the less expert by a combination of (i) task decomposition: (ii) ordering skill acquisition, so each step prepares the next; (iii) well chosen exemplars. Such expert-structured and supervised learning by doing  is very powerful, as craft apprenticeship learning shows. To the extent that skill acquisition in forager societies is similar to this mode of hybrid learning, it makes possible high volume, high fidelity social learning. This model acknowledges the importance of individual cognitive adaptations, but equally calls attention to the role of adapted environments.

As I see it, the apprentice learning model has important virtues. First: it identifies a form of learning that can be assembled incrementally. As Eva Jablonka has shown, the reliable transmission of skill can begin as a side-effect of adult activity, without adult teaching or adaptations for social learning in the young. For an innovation that becomes central to adult economic activity as a side-effect transforms the local environment that juveniles explore, positively biasing learning probabilities. Once established, skill transmission then brings with it selection for cognitive and social changes that increase the reliability or reduce the cost of learning. But rudimentary but reliable skill transmission does not presuppose the presence of such adaptations. Second, apprentice learning is known to support high fidelity, high bandwidth knowledge flow. Until recently, much technical competence depended on such learning. Third, this organisation of learning remains powerful in changing environments, so long as the pace of change is not catastrophically rapid. Fourth, the model fits the ethnographic and archaeological data. For example, in many forager societies children’s toys and games practice crucial skills, and those societies organise and enhance children’s participation in economic activity. All this supports the transmission of traditional craft skills. Finally, this basic model has broad application, beyond formal apprenticeship in highly skilled craft guilds. For instance, I have argued elsewhere that it fits the cross-generation transmission of norms and values. There were no Palaeolithic craft guilds akin to those of early modern Europe, but there are quite striking similarities between skill transmission in formal apprenticeships and skill transmission in traditional society. They both depend on socially organised and adapted learning environments which marry the power of trial and error learning to that of cultural transmission. Such environments make possible the transmission of high fidelity, high volume information across the generations. That in turn makes possible the reliable acquisition of complex, learning-dependent competences in a changing world.

10 comments to The Evolved Apprentice

  • Interesting model. How can it be falsified?

  • Clem Weidenbenner

    Kim said:
    If we really had stone-age minds in an electronics-age world, we would be crippled by adaptive lag.

    Doesn’t this beg the question of how an electronics-age world could come to exist among stone-age minds in the first place? I suppose an advanced life form from another world might visit, but shy of that… Still the concept of an adaptive lag might actually be a potential testing point – in regard to Bjorn’s question. I’m not a cultural anthropologist so this notion could simply be a naive one, but are there any extant groups with survival skill sets (Aboriginal perhaps?) so far removed from our “modern” (electronic-age) ones that they might be tested for an adaptive lag?

  • Paul Sheldon Davies

    Affective Apprentices: A Third Framing Idea?

    Mammal species differ in cortical structures but hardly at all in mid- and lower-brain structures. The latter are largely homologous with shared systemic functions. In humans, as in rats and cats, electrical stimulation of specific tracts in the lateral hypothalamus arouses the same autonomic responses and also the same proximate behavior – vigorous exploration of the environment. Jaak Panksepp calls this the SEEKING system and hypothesizes that it is, among other things, the neural basis for foraging behavior.

    SEEKING, along with three other systems identified by Panksepp, are fundamental to mammalian sociality. The PANIC system, for instance, can be triggered by stimulating homologous subcortical structures in mammal brains, and these stimulations elicit the very same behaviors as when infants are separated from caregivers – distress cries, spikes in stress hormones, etc. This system is also triggered when adult humans suffer severe loneliness or grief and is quieted when attachments are reestablished – when social homeostasis obtains – or when chemical surrogates (such as heroin) are administered. But PANIC also moves us to seek and maintain social attachments; it is an anticipatory system against the pain of future loss. It is, in consequence, a robust processor and evaluator of social information, and a trigger for gathering more information. So our ancestors 400,000 years ago foraged together for more than food; they were warding off the pain of hunger and of isolation.

    These and other ancient affective systems are primitives of the mammalian mind. They are ancestral capacities from which current cortical capacities evolved and without which cortical structures lose their efficacy. If, for example, we remove the cortex of non-human animals early in life, they nevertheless exhibit integrated affective behavior later in life. Humans who do not develop a cortex are much the same; despite devastating cognitive deficits, their behavior is affectively coherent.

    We cannot hope to understand cortically-based capacities without understanding the pervasive and foundational effects of mammalian affective capacities. Nothing in mammalian psychology makes sense except in light of evolved affective capacities – including our capacities to respond to novelty in a changing world. We are evolved apprentices by virtue of being affective apprentices.

  • Catherine Driscoll

    I think this is a really interesting article, and mostly I agree with Kim about the importance of social learning in human history and the way that learning environments get structured by teachers. There are, however, two of Kim’s points to which I want to respond. I’ll put these in separate posts.

    1. Domain specificity, novelty and “information load”
    Kim argues that modules are poor handlers of novelty – modular minds “should be crippled by adaptive lag” in the face of “electronic age technology and institutions”. It’s not obvious to me that a modular mind of the sort actually thought to exist by some evolutionary social scientists would be routinely unable to handle novelty. The “domain” of a module can be almost anything – including some problem or set of properties of an environment abstract enough to be present in both the Stone Age and the Electronic Age. Face recognizers and cheater detectors – classic central processing “modules” – are interestingly adaptive even in modern societies, where there are still faces and cheaters. Modules only become maladaptive where the modern environment changes in a way that cuts across their capacity; Kim’s main example, formal institutions, isn’t obviously a case where this should happen. Stone Age civilizations may not have had big complicated formal institutions like universities, governments and so on, but they certainly had simpler social institutions with the attendant symbolism, roles and social norms. Modules of a relatively abstract sort, able to handle he abstract features of these sorts of institutions, might well be able to extend themselves to handle modern ones. Some of our failures to handle the electronic age are consistent with the view that our intelligence is a bit modular – ask any iPad designer about creating technology for creatures whose minds are designed for interaction with tools via touch, for pushing, pulling and dragging objects that obey certain elementary physical laws.

    Another misunderstanding about minds composed of domain specific mechanisms is that they require problem solving and information management to be piecemeal, with one module handling one problem at a time. This is simply not true – modules in the sense of domain specific mechanisms might handle more complex problems by interacting together – by sharing components of a larger task and engaging in complex feedback interactions (this post is already too long so I’ll give an example only if asked). With enough modules, modules handling abstract enough representations, and a few domain general mechanisms for handling storage and memory of socially learnt information, modular minds might also handle rapidly changing environments.
    Consequently, unlike Kim, I’m still agnostic about whether our minds are partly or largely modular. They might be.

  • Catherine Driscoll

    2. Social learning and nativism
    Kim clearly isn’t keen on views of the mind that require a lot of innate information, although he does say here that evolved apprentices do need some innate knowledge. I think this needs more emphasis. Kim’s apprentice, even learning in her scaffolded learning environment, needs a lot of innate information to do what she needs to do successfully. Imagine I am a child living with some adults in just such an environment and I am, at some point, going to learn how to carve a wooden spear head. Suppose I undergo the sorts of experiences Kim describes above – I see adults interacting with objects in the right way – I regularly get to watch Mary carving a spear. I also regularly see the materials for making spears about, and I see some partially completed spears. What do I need to be able to do in order to use this environment to learn to make a spear? I need:

    To understand that the environment around here is full of “things” with which it is possible for me to physically interact; that “things” behave in certain ways. That there are thing-types: “knives”, “spears”, “pieces of wood”.

    To understand (when watching Mary carve a spear with a knife) that there are two things (a knife and a spear), and Mary is doing something with one thing to the other thing. That Mary can do something with one thing to the other thing of the same type, on multiple occasions (that there are such things as action types and that “carving” is an action type). I need to understand which bits of what Mary actually does are relevant to something counting as a token of the action type – stabbing herself in the finger with the knife and yelling “ow!” isn’t relevant, nor is the way that Mary moves her elbows, but the direction in which she moves the knife is relevant.

    To understand that Mary intends something by the knife and wood activity, and what it is she might intend (she is trying to make a spear). That sometimes she might not intend anything (she’s just whittling).

    To understand why what Mary is doing matters, and why it is relevant to me (spears might help me live successfully in this environment, just as they do for Mary). That because it is relevant to me, it should be something to which I should attend; I should be motivated to copy Mary.
    To understand that Mary is worth paying more attention to than John, and this is because John’s spears are less good than Mary’s with respect to the features that matter to me.

    To understand that all these funny looking objects are partially made spears; they are on the way to being spears; that these are not just weirdly caved sticks, or mistakes, or sticks that got like that accidentally. That there is a difference between made artifacts and accidentally caused things.

    And so on ad nauseam.

    Many of these things are not things that a scaffolded learning environment can teach a child. They are things that the child has to be able to understand or learn, by themselves, in order to make use of a scaffolded environment. They are skills that the successful use of a scaffolded environment is predicated upon. How much of all this requires genuinely innate knowledge, and how much can be done using simple innate learning biases (such as attention-based and motivation biases) is up for grabs. But it does suggest that “evolved apprenticeship” requires an interaction between innate knowledge or biases and cross-generationally structured learning.

  • I agree with large chunks of the picture of hominid evolution drawn by Kim Sterelny. He is right to emphasize the importance of trans-generational, high-volume, high-fidelity social learning, and the apprentice learning model, based on the idea that expert teachers design learning environments in which such social learning takes place, contributes to our understanding of how this form of social learning might have evolved.

    This model also fits well with the hypothesis of an evolved pedagogy put forward by Csibra and Gergely (2009) and with what is known about the evolution of childhood and the juvenile period (for review, see Kaplan et al., 2000 and for an evolutionary model, see Kaplan & Robson, 2002). (Incidentally, Kim, how do you view your hypothesis in relation to these hypotheses and models?)

    That said, I have reservations about the contrast Kim draws between the massive modularity hypothesis and the apprentice learning model. It is misleading to treat (even for expository purposes) the modularist conception of the evolution of the mind (when it is not reduced to a convenient strawman) and the apprentice learning model as incompatible alternatives.

    First, the fact that we acquire much knowledge and skills through this form of social learning is compatible with the existence of numerous modules fulfilling various evolutionary important functions (mate choice, predator detection, cheater (bastard?) detection, and so on). That modules are not the whole story does not mean that they are not an important part of the story.

    Second, as most (if not all) modularists maintain, modules are typically learning systems because the information required to fulfill their function (e.g., the information needed to detect predators) was not stable over evolutionary time. For this very reason, Clark Barrett (e.g., 2005) insists that the hypothesized cognitive system involved in learning to identify predators is fundamentally a learning system that helps children learn which animals are predators in their environments. The existence of modules is thus not incompatible with the importance of learning and of teaching. Quite the contrary in fact: the apprentice model complements nicely the modularist approach. Because in many (but not in all) domains information was not stable over evolutionary time, we have evolved task-specific learning systems, and we have also evolved a disposition to create learning environments for the acquisition of the relevant information.

    Whether or not (and in what way) learning is central to particular evolved cognitive system) depends on the nature of the functions these evolved cognitive systems fulfill, and on the nature of the information that is needed to fulfill these functions (Fessler & Machery, forthcoming). This varies from domain to domain and from function to function. Little understanding is gained by opposing the modularist approach to the apprentice model.

    Kim acknowledges that modules are not incompatible with learning, but goes on all the same opposing the modularist approach with the apprentice learning model. The reason is that Kim does not see how evolved modules could underlie much of our behavior since (1) modules are inflexibly geared toward solving those past problems that were relevant during human evolution, (2) our environment has changed in tremendous ways (“many central aspects of the human world have changed fundamentally”), and (3) we are adapted to our environments (“we remain competent in responding to many of these novel challenges”). He concludes that “to the extent that the modularity hypothesis explains competent response to information-hungry problems by appeal to pre-loaded information, it is poorly posed to explain competent response to evolutionarily novel challenges.”

    My comments above about the place of learning in the modularist approach suggests that (1) is dubious, but in spite of the suggestive examples given by Kim (2) is also controversial. In many domains (mate choice, group identification, etc.), one does not need to look very hard behind the variable aspects of human life to identify constancies. People all over the world pay much attention to group membership and advertise their membership in groups by means of ethnic markers, the pragmatics of conversation are, as far as we know, the same all over the world, and so on. Indeed, one of the mechanisms mentioned by Kim to characterize the apprentice hypothesis—niche construction—explains why there is so much stability behind the appearances of variability. Because humans build their social world and to some extent their physical world, they create similar social organizations across generations and cultures exactly as beavers make sure that the environment is adapted to their needs by creating dams. For instance, as Richerson and Boyd (1999) have shown, one finds again and again large groups of thousands of individuals in a range of social and historical contexts.

    Barrett, H. C. (2005). Adaptations to predators and prey. In D. M. Buss (ed.), The Handbook of Evolutionary Psychology (pp. 200-223). Hoboken, NJ: John Wiley & Sons.
    Csibra, G., & Gergely, G. (2009). Natural pedagogy. Trends in Cognitive Sciences, 13, 148-154.
    Fessler, D., & Machery, E. (Forthcoming). Culture and cognition. In E. Margolis, S. Laurence, & S. Stich (Eds.), Oxford Handbook of Philosophy and Cognitive Science. Oxford University Press.
    Kaplan, H. S., & Robson, A. J. (2002) The emergence of humans: The coevolution of intelligence and longevity with intergenerational transfers. PNAS, 99, 10221-10226.
    Kaplan, H. S., Hill, K. R., Lancaster, J. B., & Hurtado, A. M. (2000). A theory of human life history evolution: Diet, intelligence, and longevity. Evolutionary Anthropology, 9, 156-185.
    Richerson, P. J., and Boyd, R. (1999). The evolutionary dynamics of a crude super organism. Human Nature, 10, 253-289.

    Edouard Machery

  • Some of us are more externalist than others and either a complement to or a consequence of our externalism is a skepticism about poverty of stimulus arguments. In the bad old days when behaviorism was the only externalist view in town the concept of the environment was weak and underdeveloped. Kim Sterelny’s externalist approach blends insights from behavioral ecology, niche construction, social learning theory and multi-level inheritance theory, among others. Sterelny brings these resources together to help us understand that we can explain rapid changes in human cognitive competence during evolutionary history without having to resort to a dizzying array of special purpose, internal adaptive mental modules, each of which requires an account of their evolution. As I see it, the key to his approach is a careful consideration of environments.

    Environments are dynamic or stable, transient or permanent, but also, as other evolutionary theorists have pointed out, selectively relevant or irrelevant. What Sterelny helps us to see is that in lots of cases, it is a lot easier — in terms of energy expenditure, time spent and so on – to get a reasonably plastic individual to perform a demanding cognitive task by structuring their environment than it is to wait around the requisite time for their ilk to evolve a ready to order cognitive mechanism for the task. The relevant environmental structure has many dimensions. When I first walked into my grandfather’s tool shed in his yard, I was about six years old. The shed was full of all manner of foreign objects with plenty of potential. My grandfather’s demonstration of how to use a plane, a saw and a chisel provided crucial additional structure to the environment. My subsequent practice under his watchful eye got the process of skill transmission in motion.

    All this being said, I do not see Sterenly, here or anywhere else, ruling out evolved internal structures. As he says here “if the information a child needs about her world is stable over evolutionarily significant time frames, selection can build that information into human minds.” What he emphasizes is that this approach will not explain all of human cognitive evolution. A vast amount left unexplained by modularity theorists will succumb to a combination of approaches that place heavy emphasis on carefully characterizing relevant environments.

  • Daniel Kelly

    The notion of a skill looms large in Sterelny’s provocative post and its intriguing alternative to extreme nativist views about the human mind, and as such, that notion promises to figure prominently in attempts to formulate the sorts of more fine-grained hypotheses that might be testable, and which would help distinguish his evolved apprentice model from its competitors. Despite its import, though, I am not sure how he would characterize skills, and so hope to smoke him out a bit on this issue.

    One clear feature has to do with their sources: much or most of the information relevant to a skill is acquired, or derived from the (often structured) social environment, rather than being innately specified or brought to bear on the task of learning. For example, contrasting comprehension of spoken language with comprehension of written language, he holds that “like many skills, reading depends on a long learning history in organized developmental environments.” Skills are also the types of things that can be subject to cultural evolution, since they are transmitted from one generation to the next, like beliefs, values, norms, stories, and so forth.

    Sterelny also seems to contrast skills with modules. He holds that “Skills are phenomenologically akin to modules: they are fast, automatic, and task specific.” While I take the gist of this last point, I’m unsure how to flesh it out. Are skills and modules antithetical to each other? One would think not, for a skill might be subserved or implemented by one or several modules. Take even the example of reading: surely, exercising this skill engages a slew of cognitive mechanisms that are paradigmatic modules, namely ones that subserve vision. Likewise, our ability to predict and make sense of others by ascribing to them mental states like beliefs and desires, and making inferences about the connections between those mental states and behavior – what is sometimes called our capacity for mindreading – might naturally be described as a skill, at least given the colloquial usage of that term. However, while different theorists might agree on how to characterize this skill, they might (and have, and do) at the same time disagree about the nature of the cognitive architecture that undergirds it, including whether or not that architecture is modular, or how much of it is innately specified and how much is learned, etc.

    I have used different words – “subserve”, “implement” “engage” “undergird” – to point to one crude but viable model with which to understand the relationship between a skill and a module. Surely this needs to be fleshed out more, but given the contrast Sterelny implies between modules and skills, it does not seem to be one that he would be willing to embrace.

  • Kim Sterelny


    Testing the Model. I agree that the issue of testability is of great significance, though I would not agree that testability is best understood as falsifiability, in Popper’s sense. So in a forthcoming paper in Philosophical Transactions (From Hominins to Humans: How Sapiens Became Behaviourally Modern), I devote the final section to testability, testing the idea both against the anthropological and the archaeological record. I argue, for example, that on the evolved apprentice model, the human capacity to maintain and expand cognitive capital depends on both the intrinsic cognitive capacities of the human mind, and the organisation of the learning environment. On cognitive breakthrough models we should expect to see pulses and plateaus; on the evolved apprentice model, partial reversals are to be expected, and that indeed is what the record shows.

    Sheldon Davies

    Neural bases of behaviour. This essay is a synopsis of a monograph, and in the first draft I had a quite long section arguing that the evolved apprentice model fitted neuroscientific data quite well. I have cut that section completely. There is, indeed, quite a lot of evidence of neural plasticity and of environmental effects on neural organisation (e.g. work on the effects of environmental enrichment in animals). But these results are no more than consist with the evolved apprentice model; they do not directly support it. Moreover, since I am persuaded that gene-culture coevolution has long been important, and that there are important environment-cortical organisation interactions in development, I would be extremely cautious about following Sheldon Davies’ lead in making claims about cortical organisation of hominins of 400 kya. Even if we could clearly establish the function of some cortical system in living populations of humans (and that is far from trivial), we cannot safely assume that homologous structures in ancestral populations have the same structure.

    Affect. I agree with Sheldon Davies that to understand the descent of mind, we have to understand the evolutionary transformations of affect and motivation, not just of cognitive systems narrowly understood. But I would not regard this as especially controversial now. Sarah Hrdy (in Mothers and Others), in arguing for her “reproductive co-operation” model of human evolution makes the transformation of emotion and affect central. Michael Tomasello and his colleagues have recently focuses on the pivotal role (as they see it) of joint attention and collective intentionality, and for them, collective intentionality has an important motivational element; it is not just cognitive.

    Catherine Driscoll

    Novelty and Modules. Unsurprisingly, I am unconvinced by Catherine on this. First: it is not enough that the problem be stable in the Pleistocene to (late) Holocene transition. The kind of information needed to solve it must be stable, too, if prewiring is to explain our cognitive competence. And while the problem of detecting cheaters may well be stable, I think I is most unlikely that the way cheats are exposed is stable — at any useful level of abstraction — in that transition. (I agree face recognition may well be a module, though it is probably perceptual rather than cognitive). Second: the more one think of a set of modules as operating simultaneously, sharing data, sharing components, the less obvious one still has the explanatory clout that the appeal to modules was supposed to buy: their speed, their independence (and on some views of modules, their double dissociablity). They are no longer autonomous systems. Of course, I accept that modularity is a matter of degree. But this suggestion risks saving the term at the cost of its original substance.

    Social Learning and Nativism. I am not sure how fundamentally I disagree with Catherine here, because I am not sure of just how much she includes in “many of these things are not things a scaffolded environment can teach a child”. Obviously: there is initial structure in human minds, and that is different from the initial structure of infant chimps and infant gorillas. There is a distinctive organismal contribution, and this no-one denies; Steve Downes points this out too. In my essay (and still more, in my other recent work), I have been concerned to emphasise how rich the stimulus is; children develop in a very information rich world, and often a world organised to make that information salient to them. Salience, of course, is sensitive to initial, and then typically developing, structure. Pointing is salient to young humans but not young chimps. So yes: there is initial structure, and sometimes that initial structure will develop adaptively relatively independently of organised informational inputs (quite likely, in the case of “folk physics”). In many cases, it will not. Part of the argument, then, is about cases, and about how much structure, and how specific it is to specific learning tasks. But that is not all the argument there is. For it is a further claim that this initial structure is well characterised as pre-installed information. Pat Bateson and Paul Griffiths, in their different ways, are two sceptics about this further claim. That further claim might be true, but it certainly does not follow immediately from the claim that initial structure is necessary.

    Edouard Machery

    I agree with Edouard, of course, that hybrid models are possible; indeed, they are plausible. As I mentioned above, I think a modular view of folk physics is very plausible; a modular view of language is still a viable option, though the case for it is much less overwhelming than it once seemed to be. I also agree that modules can be seen, and often are seen, as learning machines. The question, then, is how much modules as learning machines shape or constrain the direction of learning. For modular hypotheses to explain our cognitive competence in the face of high information load problems, it seems to me that they have to shape or constrain learning very significantly: we are good at cheater detection, mate choice, predator avoidance, because our innate, pre-wired, modules tells us what we need to learn. But if that is right, it seems to me that the social problems of large scale, hierarchically organised, societies really should generate massive adaptive lag problems. In contrast to Edouard, I do not think that the problem of mate choice is roughly the same in small scale, egalitarian societies in which almost everyone knows almost everything about everyone else, and hierarchical, informationally opaque large scale social worlds. I guess the difference between Edouard and me here is one of degree: I see some elements of stability, but very large elements of quite fundamental change. He sees the opposite.

    Stephen Downes

    Steve is right of course: emphasising the information rich nature of the environment, and emphasising the fact that this richness is no accident, by no means rules out innate structure. I know Steve is sceptical about the Baldwin Effect, but I am less so. Learning has costs, so where fitness critical features of the environment are stable, I would expect a total or partial canalisation of development; I would expect it to become less dependent on specific learning experiences, even if competent response was originally mediated by learning.

    Daniel Kelly

    Skills and modules. I take the paradigm cognitive module (language, theory of mind, moral cognition) to develop rapidly; to be relatively independent of rich, specific informational inputs; to be relatively invariant across the (relevant) population. Arguments for modularity, in this sense, often depend on some combination of invariance, poverty of the stimulus, and early development. I take skills to contrast in all three dimensions: they are not invariant within or across population; they often develop slowly (as in the “10,000 hour rule”); they often require rich and highly specific inputs. That said, of course they can both depend on, and interact with, modules, both perceptual and cognitive. If folk physics is a genuine module, many artisan skills depend on it (and, of course, on many perceptual modules). As modules typically develop early and skills develop slowly, this relationship is not symmetrical. If it really were a module, mate choice might be an exception, since it is relevant only to adult life (perhaps there are others like this).

  • This conversation, while ending here, continues on Facebook. Join us there by logging in to your Facebook account and proceeding to our group: