We endorse much of the picture presented by the author, particularly since it avoids inferential pitfalls such as over-interpreting the minds of other species, or populations, as “archaic” or “primitive” compared to our own cognitive style. In our commentary, we draw attention to a different possible role for specific materialities. The idea we wish to explore is that different materialities (fragile, enduring, and mixed) may behave differently as levers for ongoing cognitive change and evolution. Our proposal integrates that of Stibbard-Hawkes.

Following the author, we see the mind, any mind, as formed by a constant dialectic process in which embodied, active agents make the most of whatever resources (persisting or perishable) happen to be available. Such exploitative activities fall naturally out of a view that has guided much of our own recent work – the so-called “Active Inference” view prominent in recent neuro-computational work (Parr, Pezzulo, & Friston, Reference Parr, Pezzulo and Friston2022). According to active inference, predictive brains seek to minimise “prediction errors” (the differences between their predictions and current waves of sensory evidence). But predictive brains also bring about actions, designed both to achieve practical goals (rendering certain predicted future states actual) and – crucially for our suggestion – to improve their own states of information in ways that will aid current and future success (e.g., Clark, Reference Clark2015; Mirza, Adams, Mathys, & Friston, Reference Mirza, Adams, Mathys and Friston2016). This is because, in order to minimise prediction errors over longer timescales, predictive brains learn to forage for information, structuring their worlds in epistemically useful ways (Clark, Reference Clark2023; Parr & Friston, Reference Parr and Friston2017). Instead of being merely a bearer of genetically transmitted abilities and a passive recipient of information, such brains proactively predict and by action structure the world so as to curate useful streams of sensory information. The upshot is that action-exploiting predictive brains will repeatedly outsource work using whatever social or material resources might be locally available. Digital and analogue calendars can both help us remember an appointment, and tallies made on clay or an electronic calculator can both help us in calculations. Particularly in the context of nomadic, hunter-gatherer populations, such active inference agents might well exploit – for epistemic gain – perishable but easy-to-find resources.

This suggests a double perspective on the minds of our distant ancestors. On the one hand, weaving baskets and huts, and making clothes and weapons from bio-organic materials (branches, leaves, etc.) may indeed require similar cognitive complexity to making a tool from a stone. On the other hand, engaging and exploiting these different materials might drive and intergenerationally alter minds and worldviews in different ways. The question then becomes not just that of determining which cognitive capacities were required for producing various (perishable and enduring) technologies but also one of understanding how different material technologies might slowly shape human cognitive functions and understanding. For example, a stone artefact is, by definition, more durable than one made of leaves or branches. Agents could encounter old stone tools or the remains of settlements little affected by the passage of time, even if made by extinct populations or by themselves when, at some long ago time, they crossed the same territory. Such encounters might aid in developing ideas such as “permanence” and “family inheritance.” For a worked example involving the possible cognitive role of persisting structures and items, see Sutton (Reference Sutton and Hodder2020). By contrast, highly decomposable artefacts and relatively impermanent settlements might struggle to usher in this kind of temporally deep understanding.

Different materials might also impact cognitive processing in other ways. Consider the upshot of an incoming error signal (such as a mistake in some artefact's production). A stone tool and a woven container made of leaves and branches could both be fixed by working upon the affordances brought about by mistake. But in an attempt to avoid more serious damage such as breaking the core of a stone tool, an agent might be more motivated to take precautions and plan further ahead when using those materials. The production of woven objects might favour different (but equally useful) skills such as learning to replace some branches with different ones, or fully disassembling the artefact – unmaking it – and then making it again, recycling the same materials, shifting attention from the object to the process. When an environment provides both perishable and enduring materials, predictive brains enjoy multiple ways to offload work, minimise prediction error, and structure future beliefs. The interaction with, and modification of different materials allows us to train and tune our own cognitive functions in different ways, and to build various forms of epistemically rich environments. This eventually transforms cognitive processes in ways that may yet reflect the various physicalities provided by different materials.

In our own work, we have been investigating these alternative landscapes using the toolkit of active inference. We have shown, for example, how experience with a specific materiality (a decorated pot) might itself alter how agents approach a brand-new problem domain (Constant et al., Reference Constant, Tschantz, Millidge, Criado-Boado, Martinez, Müeller and Clark2021). We have also explored how achieved understanding becomes gradually “uploaded” into human-structured worlds, by encoding information and directing attention, via culturally agreed practices such as stopping at red traffic lights (Constant, Clark, Kirchhoff, & Friston, Reference Constant, Clark, Kirchhoff and Friston2022). Recently, we have modelled a “toy version” of the knapping process. In our model, changes in the shape of the tool act to encode useful information by cueing attention and engaging action in different ways, reflecting alterations to the flow of predictions and prediction errors during the activity.

We applaud Stibbard-Hawkes for drawing attention to the importance of more ephemeral materials and alternative technologies. By approaching material culture using the promising toolkit of “Active Inference” we may respect this insight while exploring new ways of putting computational flesh on the idea that material culture does not just reflect but also helps to bring about the variable suites of competencies that we call “minds.”