Biology & Philosophy

, Volume 26, Issue 4, pp 489–500

Darwinian spaces: Peter Godfrey-Smith on selection and evolution


    • Philosophy and Tempo and ModeAustralian National University
    • Philosophy and Tempo and ModeVictoria University of Wellington
Review Essay

DOI: 10.1007/s10539-010-9244-0

Cite this article as:
Sterelny, K. Biol Philos (2011) 26: 489. doi:10.1007/s10539-010-9244-0

Darwinian populations

Darwinian Populations and Natural Selection is a superb work, introducing an important analytical technique, and applying it to a range of difficult and contested issues within evolutionary theory. The core idea of the book is that of a Darwinian Population (and, derivatively, a Darwinian individual). In developing this concept, Peter Godfrey-Smith builds on a long-standing tradition within evolutionary theory, that of developing a spare, bare-bones specification of the machine of selective change, both to better articulate the core causal processes of biological evolution, and to explore the possibility that selective regimes explain change in other domains (see also Okasha 2006). Godfrey-Smith’s ingredients are familiar: Darwinian Populations are populations of interacting individuals that (potentially) reproduce; that vary from one another in ways that sometimes influence their reproductive potential; and when reproduction does take place, offspring resemble their parents.

Nothing new here, but Godfrey-Smith handles these ingredients in a distinctive way. First, everyone recognises that heritability—the tendency of offspring to resemble their parents—comes in degrees. Godfrey-Smith explicitly extends this recognition of gradients to the other constituent features of Darwinian Populations. He shows that there are clear and less clear cases of reproduction. For example, with asexual clone lines, the distinction between growth and reproduction is complex and contested (Dawkins 1982). Similarly, with social groups, there is no obvious dividing line between cases in which the individuals within the group have reproduced, and cases in which the group itself has founded colonies. So Godfrey-Smith devotes two splendid chapters to exploring and systematising some of these difficult cases. Fitness, and fitness differences are equivocal too, both because fitness differences can be counted in different ways, and, arguably, because differences in actual reproductive success only count as fitness differences if they have systematic and repeatable causes. Finally, there are unmistakable and marginal cases of a group being a population at all. Individuals are members of a single population in virtue of their between-member interactions, and those interactions vary in their intensity and consistency. So there are clear and marginal cases of populations.

Second, Godfrey-Smith explicitly distinguishes between clear, but minimal Darwinian Populations, and paradigm Darwinian Populations. Over the last four billion years, evolution in Darwinian Populations has resulted in a fabulously disparate biota, and, often, an astonishingly complex and intricately adapted biota. But not all Darwinian Populations have the potential for extensive change, or change resulting in complex adaptations. For example, the cells of a worker ant constitute a Darwinian Population. They interact. They reproduce, and daughter cells resemble their parent. They vary from one another (including genetically, as a consequence of somatic mutations), and some of those variations affect reproductive potential. Yet the potential for change is very limited. Paradigm Darwinian Populations, then, are those that possess the additional characteristics that allow selection-driven evolution to be fecund; the characteristics that make the evolution of disparity and complexity possible. A major project of the work is to identify these additional characteristics. While Godfrey-Smith does not use the term, in part Darwinian Populations and Natural Selection is an essay on evolvability.

This is a very fruitful way of conceptualising the field. The focus on gradients takes pointless hair-splitting about particular, difficult cases off the table: we do not have to decide (say) that a wolf packs meets some threshold so the pack itself counts as reproducing. But it does much more: it focuses our attention on the dynamics of evolvability. In a lineage of socially interacting agents, like our wolves, the degree of integration, of reproductive specialisation, of ancestor–descendant similarity need not stay fixed over time. A population of packs might be a very marginal case of a Darwinian Population at one point of time; a clear but minimal case at another; a paradigm Darwinian Population at a third. Populations can both Darwinise and de-Darwinise. Godfrey-Smith (following up (McShea 2002)) points out that as Darwinian individuals become parts of larger, paradigmatically Darwinian individuals, they themselves tend to become simpler, and less obviously members of Darwinian populations. An ant is a more obvious member of a Darwinian Population than an ant cell. The gradient conception of Darwinizing characteristics thus focuses our attention on the causes and consequences of such changes in population-level characteristics; on the flow and ebb of evolvability.

Moreover, theories of evolvability have tended to focus on characteristics of individual agents; typically, on features of their developmental system. But while agent level traits are clearly relevant to evolvability, it is a feature of populations and lineages of populations, not of individual agents. So Godfrey-Smith frames the project asking the right questions, focusing on the features of populations that explain the potential for rich evolutionary change. Finally, Godfrey-Smith treats these gradient characteristics of Darwinian Populations as dimensions, and that enables him to represent these cases spatially. For example, the volvocacean aquatic algae are mostly single-celled eukaryotes. But a number of multi-celled lineages have evolved, and these differ in complexity, reproductive specialisation and internal differentiation. Some of volvocaceans seem to be clear cases of complex Darwinian individuality; others are perhaps more naturally seen as temporary alliances between simpler, single-celled individuals (Kirk 2005; Michod 2011). Godfrey-Smith’s spatial treatment is heuristically very valuable here and elsewhere, because it enables him to exhibit these lineages’ similarities and differences clearly and compactly. A major virtue of this book is Godfrey-Smith’s development of a clear and user-friendly way of highlighting and visualising the key phenomena.

Darwinian populations and gene selection

In building his models of the Darwinian machine, Godfrey-Smith builds on a tradition in evolutionary theory beginning with Darwin himself. But its best known expression is in the work of Richard Lewontin (1970). This tradition takes individual organisms forming a population as the base case, and generalises from that case, abstracting away from details of agent and population. Godfrey-Smith sees the importance of this project, but is acutely sensitive to its limits. He points out that these abstract general recipes for evolution by natural selection cannot deliver on all their ambitions, for there is a fundamental tension between generality and simplicity. Simple, general formulations cannot capture all of their intended targets. So, for example, he points out that Lewontin’s recipe does not capture cases in which the members of a population vary only in the rapidity with which they produce offspring (rather than varying in the number they produce). This is not logic-chopping. While no real population is likely to vary only in this characteristic, arguably selection for maximising the rate of reproduction is the most pervasive from of selection on bacterial populations: see Lane (2005). Even if Lane overstates his case, if As reproduce more rapidly than Bs, As will increase their representation in the population over time. There is no simple, literally true, specification of the necessary and sufficient conditions of evolution by natural selection.

Godfrey-Smith suggests that we do better in seeing formulations like those of Lewontin as models: as a descriptions of an ideal, simple system that is strikingly and importantly similar to many real cases. Thus while seeing the limitations of classical recipes for selective change, Godfrey-Smith suggests ways these recipes can be both enriched and reinterpreted to lay bare the structure of uncontroversial but difficult cases of Darwinian machines in action (for example, evolution amongst the multicelled plants), and to identify more cryptic Darwinian machines in action. He is markedly less charitable to an alternative, more recent attempt to specify a general recipe for evolutionary change, a recipe built around the idea of a replicator. George Williams, Richard Dawkins, and David Hull have argued that evolution is best seen as competition between lineages of active replicators (Williams 1966; Dawkins 1982; Hull 1988, 1989). These are structures that (1) pass on their organization to descendants with high fidelity, through some copying process; and (2) influence their own prospects of being copied. Active replicators exert power over their local environment, and the nature and extent of that power determines their fitness. Over much of the history of life, the default strategy of successful replicators has been that of forming ensembles which guide the construction of vehicles (or “interactors”, in Hull’s terminology). These vehicles interact with the environment in ways which, if successful, lead to a further cycle of replication and vehicle building. So replicator lineages normally grow through the co-operative construction of biological machines. But while this is their default strategy, it is not the only route to replicator success.

Just as the classical recipe takes organisms in competition to be the base case, obviously this replicator-centred view is abstracted from the gene. Godfrey-Smith is deeply sceptical. As a former defender of this view, I shall attempt a defence of my former time slice. After some preliminary sparing, the core argument is developed in Chap. 7, and I read that argument as follows:
  1. 1.

    Godfrey-Smith begins with the basic Lewontin formulation: evolution by natural selection is a consequence of populations in which we find reproduction, heritability and fitness differences. But Godfrey-Smith refines, revises and enriches that formulation in developing his ideas of minimal and paradigm Darwinian Populations.

  2. 2.

    He shows that this conception is theoretically fruitful, by showing how we can use it to organise and systematise a menagerie of difficult cases involving putative multi-level selection, reproduction and individuality. There is, for example, a terrific discussion of plant reproduction. Plants are appallingly behaved (Clarke 2011). Many reproduce by some form of vegetative propagation, so physiological and genetic criteria for individuality come part. They do not have a germ-line that develops early and is isolated from the rest of the organism, so mutations in the course of growth can result in morphological modules which are both genetically different from other regions in the plant, and which have the capacity to reproduce independently. So they are often chimeras. Many ferns reproduce via a haploid stage which grows into a complex multi-celled form, before producing gametes which disperse and, if lucky, fuse with another gamete. Counting plants, and hence estimating plant fitness, is a horrible problem. Godfrey-Smith decomposes the intuitive, folk biological concept of reproduction into three factors: the integration of the putative individual; the extent to which it is genetically homogenous; the extent to which it develops through a narrow bottleneck. In doing so, he gives us a framework to compare the different cases.

    This theoretical productivity step in the argument is crucial: the replicator-vehicle framework was formulated an alternative to, and replacement for, organism-centred conceptions of evolution. To the extent that a Darwinian Population is an idealised, abstracted model of organisms in interaction, of course it does not fit interacting gene lineages. The gene-selection model could not be an advance, if it did, and the replicator-vehicle recipe would not be saying anything new.

  3. 3.

    He shows that the genes in an organism (or in a population of organisms) do not form a clear, unmistakeable Darwinian Population. In those cases in which genes are well individuated in terms of their physical organization and action on the world (for example, coding for a specific protein), they have a shared fate. For genes are well-individuated in prokaryotes and only prokaryotes. But in those prokaryotes, genes succeed or fail together. So we do not see the fitness differences characteristic of unmistakeable Darwinian Populations. Meiosis and cross-over in eukaryotes releases genes from their shared fate, but such genes are much less well-individuated in terms of cohesive physical structure and phenotypic action. The eukaryote challenge is the real driving point at this stage of the argument: a replicator-vehicle model can take the single bacterial chromosome to be a single replicator. But Godfrey-Smith argues that the eukaryote genome does not consist in a temporary alliance of clearly identifiable, autonomous replicators.

  4. 4.

    Godfrey-Smith then suggests that many of the standard gene’s eye cases are better seen as selection on individual organisms, but for genetic properties of those organisms. Supposed selection on giraffe-inhabiting genes for long necks is really selection in favour of giraffes that carry that gene. Thus he points out that even gene selectionists count fitness by counting giraffes with the long-neck gene. Even by gene-selection lights, a long-neck gene does not increase its fitness by increasing its copy-number, because long-necked animas need to grow more neck cells.

  5. 5.

    Outlaw genes—meiotic drivers, sex ratio distorters and the like—are appropriately described as cases of gene selection. For there is variance in the fitness of germ-line genes in individual organisms. But these are exceptional, cases of Darwinian populations. They may even be marginal cases, for though the phenotypic upshot of selfish genetic elements is determinate, their boundaries are often indeterminate, and, as with other eukaryotic genes, heritability is compromised by cross-over.


The bottom line: selection between genes is not even one paradigm case of evolution by natural selection, let alone the base case that should organise our conception of evolution. This is a powerful line of thought. But it does not give replicators their due.

Extended phenotypes

One idea behind the replicator-vehicle model is that genes—replicators—do not always leverage themselves into the next generation via their effects on the phenotypes of well-defined Darwinian individuals. Dawkins’ TheExtended Phenotype introduced this idea, and it has been further developed (in different terminology) in the niche construction literature (Odling-Smee et al. 2003). To borrow Godfrey-Smith’s term, there is a menagerie of such cases. These include cases in which organisms physically modify their environment; they include genes whose adaptive function is to establish stable patterns of social relationships (like those found in many species of co-operatively breeding Australian birds). They include symbiotic alliances, in which each partner is selectively important to the others, but where the lineages have not fused, like eukaryotic cells, into a single individual. Bullhorn acacias, for example, have physical structures that house and feed ants, their symbiotic partners, and those ants engage in guarding and anti-herbivore patrols that are adaptive for the ants only via their benefit to their symbiotic partners (Janzen 1966). Yet we still clearly have independent lineages here. They also include the seriously spooky cases of parasite control of host phenotypes (Combes 2005); for example, the parasitic barnacles that feminise their host's morphology, and hijack their behaviour (Gould 1996).

These extended phenotype cases have two distinctive features. First: they are all cases of selection acting powerfully to shape the biological world. These extended phenotype adaptations are not produced in minimal, low-powered, or marginal Darwinian regimes. Second, if we recast these evolutionary episodes into the framework of Darwinian Populations, we have to choose between two suboptimal representations. One represents these evolutionary histories as the histories of paradigm Darwinian Populations, but the actual adaptive effect that drives change is suppressed: it is represented by its proximal shadow. We represent the ants’ behaviours, but not their effect on acacia growth, health and productivity, and hence their indirect effects on the ants’ home; we represent the chemical messengers that the parasitic barnacle produces, but not its effects on the host phenotype, effects which lead to host care of the barnacles’ own eggs. Alternatively, we represent the adaptive levers, but the lever is acting on aberrant or marginal cases of Darwinian individuals. Our Darwinian Populations will consist of rabbits plus their burrows; ant-acacia partnerships; crabs feminised by their barnacle parasites.

The extended phenotype menagerie shows that paradigm cases of selective shaping do not line up 1:1 with selection driving change in paradigm Darwinian individuals in paradigm Darwinian Populations. Marginal and unclear cases of Darwinian Populations are not always cases where the Darwinian machine is running less powerfully. So if the point of distinguishing between minimal and paradigm populations just is to distinguish those regimes in which the generation of novelty, complexity, or additional disparity is possible, the distinction is only partially doing its job. Novelty and complexity is sometimes not novelty and complexity in a paradigm Darwinian individual. The replicator-vehicle model captures that fact.

Replication and cumulative selection

I also doubt that Godfrey-Smith does justice to the links between replication, high fidelity inheritance, and evolutionary potential. It is a truism of evolutionary biology that adaptive complexity can be built only by cumulative selection of small variations on existing phenotypes. A large, undirected change from an existing phenotype is almost certain to be maladaptive. But a small variation on an existing phenotype has some reasonable chance of improving it. Complexity is built incrementally. In Climbing Mount Improbable (especially), Dawkins connects this truism to the replicator-vehicle framework. Cumulative selection requires high fidelity inheritance, as small variations from existing types must be preserved if they are to be available as a basis for further improvement (Dawkins 1996). High fidelity in turn depends on replication; a discrete, perhaps even digitalised, template copying process (Dawkins 1983). Replication, Dawkins argues, is the only mechanism that could ground high fidelity inheritance. It is fundamental to the evolution of complex, antecedently improbable systems.

I am not sure where Godfrey-Smith disagrees with this line of analysis. He rightly points out that defenders of the replicator-vehicle model are just mistaken when they claim that replication is essential for any form of evolution by natural selection, and equally mistaken in thinking that evolutionary change depends on the persistence of types across generations. Height in a population can evolve over time even if no two individuals ever have the same height; it suffices that offspring resemble their parents more than they resemble others (p. 33). It is also true that we can indeed imagine paradigm Darwinian Populations with no replication. Godfrey-Smith imagines one in which genes direct protein synthesis, but once the organism develops, the genes are metabolised, but reverse translated from proteins immediately prior to reproduction. In such a biology, there would then be no unbroken chain of DNA copies from parent to offspring. We can indeed imagine such a case, but what has imagination got to do with it? Godfrey-Smith agrees that the imagined system is probably not biochemically possible. The fact that we do not see DNA to protein to DNA cycles is no mere quirk of evolutionary contingency.

So the Williams-Dawkins Conjecture is this: (1) richly evolvable selective regimes depend on high fidelity inheritance; (2) all actual and feasible cases of high fidelity inheritance depend directly or indirectly on replication. Indirect cases are scaffolded by Dennett’s cranes (Dennett 1995): complex adaptations whose evolution does depend directly on replicator-based inheritance. As Godfrey-Smith notes, high fidelity cultural inheritance in the human lineage probably does not depend on replication (a point that Dan Sperber has made forcefully: Sperber (1996). But it does depend on complex psychological adaptations whose evolution did depend on replication. Any form of high fidelity inheritance seems to require some combination of a benign environment and complex machinery; machinery unlikely to be available for free. That is why the origin of life presents such a brutally difficult chicken and egg problem. Hi fidelity without replication exacerbates that problem; you need even more machinery. Using a structure as input to a template copying process must be simpler that using a representation of that structure as input; template copying stores information in the world. The machinery which replicates DNA is complex enough. Imagine what you would need to add, if as well you needed molecular machines which would translate unfolded, complex three dimensional protein shapes into a 2D linear order of amino acids, before these were reverse translated into DNA. If the Williams-Dawkins Conjecture is right, even though Godfrey-Smith is indeed right that some evolutionary regimes do not involve replication, replication is fundamental to evolution in a way his framework does not capture.

Suppose Godfrey-Smith was to accept much of this. What then? The importance of replication could be built into his picture very simply. In graphing paradigm Darwinian Populations in selection space, one dimension represents heritability. Godfrey-Smith could simply stipulate that high heritability values depend on high fidelity replication, thus explicitly recognising the pivotal role of replication.

Capturing the extended phenotype zoo is less straightforward. But the model-based strategy defended in the book does not demand exclusivity. Godfrey-Smith writes:

One option is to give a literally true description of the important features of all cases of the phenomenon. The other option is to give a detailed description of one class of cases, usually a relatively simple case, and use it as a basis for understanding the others. … Understanding is achieved via a similarity relation between the simple cases we have explicitly analysed and the more complex ones (p. 26).

This meta-theoretical perspective does not require a single choice of model. So Godfrey-Smith could argue that while many central episodes of evolutionary history are best understood as evolutionary change in, from, or to paradigm Darwinian Populations, when the adaptive effects of genes and gene changes are distributed over several biological individuals, or when they affect coupled organism-environment systems, the Replicator-Vehicle model is more illuminating. We would get a more complex picture and we would need some account of the interaction of the two kinds of case. But the move is available.

Selection spaces

A crucial part of Godfrey-Smith’s analysis is his distinction between marginal, minimal and paradigm Darwinian Populations. These distinctions are drawn by locating populations in an abstract, five-dimensional selection space. The dimensions are (a) H: the fidelity of reproduction; (b) V: the supply of variation; (c) α: the intensity and ubiquity of competition for reproduction through the population; (d) C: the smoothness of the fitness landscape; and (e) S: a parameter that represents the extent to which fitness differences in the population depend on differences in the intrinsic features of the individuals in the population. The importance of V and H is obvious. The role of C has also been identified in the literature. Cumulative evolution requires high fidelity inheritance. But it also requires similar phenotypes to be similar in fitness. If the fitness landscape is rugged, with very similar individuals have very different reproductive fates, there will be fitness trenches between (say) rudimentary and more complex eyes (Kauffman 1993). Thus Nilsson’s and Pelger’s famous model of eye evolution assumes that small increments in visual resolution corresponds to small increments in fitness (Nilsson and Pelger 1994).

Godfrey-Smith’s α takes up a neglected issue is evolutionary theory, that of identifying populations. The possums on Black Mountain (in Canberra) interact only occasionally with those on nearby Mount Ainslie, because a chunk of urban Canberra divides them. So do we have one population or two? Parameter α is a first-pass attempt to address this issue. A group of individuals is more paradigmatically a population, the more reproductive success for one reduces the prospects of the others. High α accentuates potential fitness gradients. Suppose a new variant in the population is unusually fit. If α is high, interaction is pervasive, and that distributes the effect of the advantage across the population. But as Godfrey-Smith acknowledges, this is at best a first pass attempt at capturing a very complex set of possibilities. So, for example, it does not capture interaction chains: common when the radius of individual influence is small compared to the suitable habitat through which individuals are spread. A tree in an extensive forest competes with its neighbours, and those neighbours compete with their further neighbours, and so on. But our first tree, and one growing a kilometre away, have negligible effects on each others’ prospects. The same will be true of many of the small invertebrates living on and by the tree. Sometimes evolution acts in networks or webs rather than relatively discrete populations, and α does not capture this difference. So we should see α as a temporary placeholder for the complex and varying ways in which a species is typically composed of a variable and overlapping mosaic of populations, quasi-populations and networks of interaction.

The importance of H, V, and C is widely recognised. As just noted, there has been much less work exploring the different ways populations are integrated, and Godfrey-Smith is right to move that issue to centre stage. He is also innovative in arguing for the special importance of intrinsic fitness differences. But this is an innovation I resist. Intrinsic properties contrast to relational ones, and Godfrey-Smith argues that while relational properties make a difference to fitness, and can be inherited, evolutionary possibilities based on relation-based fitness differences are limited. His example is location. Location matters: an animal’s exposure to lighting strikes, storms, and other unpredictable disturbances will often depend on location. In many animals, the distinction between germ-line and somatic cells, likewise, depends on accidents of initial location, and this dampens down the effect of within-organism evolution in the population of cells. A variant that increases its frequency in the cell population does not thereby increase its representation in the sample of cells from which new organisms derive. Moreover, location can be inherited. Many organisms reproduce close to the place they were born. That is true even of relatively mobile animals. The birthplace of a brushtail possum will resemble that of its parents.

However, Godfrey-Smith points out that while the location of possums can evolve, evolutionary possibility is very limited. If a fortunate possum finds itself in a favourable area, the demographic focus of the population can change; a greater proportion of the population comes to live hither rather than yon. A population can explore its physical environment, concentrating in favourable areas; thinning out in less favourable ones. But that is about all: “if extrinsic features are most of what matters to realised fitness—if intrinsic character is not very important—then other than this physical wandering, not much can happen (p. 55).” Thus basing the germ-soma distinction on location within the developing embryo powers down Darwinian processes at the level of cells. Fitness differences depend on extrinsic traits of the cell, and cell-level evolution goes nowhere (pp. 102–103).

I am not convinced. First, I wonder whether the concept of an intrinsic trait is well-defined. What of those traits whose development is richly dependent on environmental input? The food preferences of a rat, for example, depend on maternal imprinting and early experience, as well as the rat’s own genetic endowment. So are preferences intrinsic? Think too of traits that guide interaction with the environment. Consider a gerygone’s disposition to build a camouflaged, pouched nest. If the gerygone’s nest building were powered by an internal template that specified location, construction technique, materials and the like, the nest building disposition would be an intrinsic trait. But what if, as is likely, gerygones “store information in the world”. The inner template guides only an initial starting point and choice of materials. The growing nest then plays an essential role in stimulating and guiding further steps in the process. Elaborate courtship displays, like those of the great crested grebe, are quite likely to be like this too. Neither sex has a mental template of the whole routine. Rather, each bird has a linked set of responses triggered by the previous step of its partner. These are cases in which the agent in question has a disposition or capacity that can be specified without reference to its environment. But the categorical basis of the disposition is not internal to the agent, as the categorical basis of the disposition of a crystal glass to shatter is internal to the glass. Rather, it depends on the way the agent is embedded in its environment. If this is how displays are organised and nests are built, are these dispositions intrinsic traits?

Second, I think there is an important difference between relations that merely happen to an agent, and relations that an agent actively makes and maintains. So the extended phenotype cases that fuel my sympathy for the replicator-vehicle model make me sceptical of the special importance of intrinsic difference makers for fitness. Termites live in elaborate physical structures, and in networks of complex social relationships. These are not intrinsic properties of the termite (they might be intrinsic properties of termite mounds, but even if a mound is a Darwinian individual, so is a termite). Yet these are clearly the result of elaborate, rich evolutionary trajectories. Some ancestral proto-termites were fitter than others in their population, because of their social relations with other termites, and through their physical organization of their surrounds. In general, when one Darwinian individual is part of a larger Darwinian individual, the inside individual’s fitness will depend heavily on evolved and elaborate relational properties. The same is true of other extended phenotype standards: beaver dam and lodge systems; parasite control over host behaviour and morphology. There might not, for example, be much current, heritable variation in beaver dam and lodge systems. But if these evolved, as they surely did, by cumulative selection, then some proto-beavers built less good dam and lodge systems, and others built better ones, and were fitter as a consequence. They were fitter because they lived in lodges that resisted attack better, or because their dams were better sited, deeper or more secure. Living in a deep, well-protected lodge and dam system is not an intrinsic feature of a beaver. But such differences were fitness difference makers, and they lead to important evolutionary changes.

That said, the parameter S gets at something: the connection between developmental control and heritability. Cumulative selection requires high heritability, and high heritability requires fine-grained control of the developing phenotype. The more development is affected by environmental noise and disturbance, the less reliably offspring resemble their parents. The boundary of the agent is often the boundary of reliable developmental control: a homeostatically buffered, predictable inner environment gives way to a wild and intractable outer one. So relational properties will develop reliably and predictably—and hence can be tuned and modified by selection—only in special cases; cases in which agents can actively manage and protect them. But the niche construction-extended phenotype literature suggests that these cases are neither rare, marginal or inconsequential.

Here is another way to put the idea. Godfrey-Smith is right to identity the intrinsic-extrinsic difference as important to evolutionary potential. But the distinction is important when we explain the cause of parent-offspring similarity, not to the similarities themselves. When traits are heritable largely because of extrinsic factors that happen to impinge on genealogical lineages within a population in repeated ways—where they happen to be born; disturbance regimes that regularly act on the phenotypes of some members of the population but not others—nothing much will happen. Evolutionary potential is much greater when parent-offspring similarity is largely due to factors intrinsic to parent and offspring. As we see above, intrinsic versus extrinsic causes of heritable differences roughly line up with intrinsic versus extrinsic traits. But the correlation is rough. So while S tracks an important control on the supply of heritable variation, I am yet to be persuaded that intrinsicality is important in itself.

In brief: I have reservations about some of the uses to which Godfrey-Smith puts his analytic apparatus. But in many cases, he uses it very illuminatingly: for example, in the analysis of difficult cases of reproduction and individuality; and in the analysis of multi-level systems. Most importantly, I have nothing but admiration for the toolbox he has built for all of us.

Copyright information

© Springer Science+Business Media B.V. 2011