Haueis and Colaço (H&C) provide an important corrective to the common tendency to background or ignore metabolic considerations in cognitive modeling. Given evolution by natural selection and the fact that, all else being equal, evolution favors more efficient solutions, metabolic factors should clearly be taken into account if one’s goal is to construct biologically realistic models of cognition (Niven & Laughlin, Reference Niven and Laughlin2008). Indeed, H&C discuss the relevance of evolution in several places (Sections 2.1.2 and 2.2.2).

It is therefore surprising that H&C focus only on the evaluation and generation of exclusively neural models. That is, nervous systems did not evolve in isolation but in the context of organisms that were already embodied and situated in structured environmental niches, and ignoring these facts is bound to distort our understanding of how behavior and cognition come about (Chiel & Beer, Reference Chiel and Beer1997; Jékely, Keijzer, & Godfrey-Smith, Reference Jékely, Keijzer and Godfrey-Smith2015). H&C furthermore focus, without substantive argument, only on computational models, in which cognition is characterized as the construction, processing, and transmission of information.

H&C’s discussion of the implications of metabolic considerations in philosophical debates about multiple realizability could be understood to leave open the possibility of non-neural realizers of cognition (e.g., “The metabolic demands of the brain differ from the energy consumption of other systems,” p. 14, emphasis added) but is not explicitly framed as such. Alternatives to computationalist cognitive modeling are mentioned but only in passing (Section 2.1.1), and the suggestion is made that the specific ontology of models (e.g., computational vs. dynamical) is not particularly relevant to their arguments (see also p. 15). I think that, on the contrary, this distinction is in fact crucial to evaluating and generating cognitive models in light of metabolic considerations.

Ecological psychology as well as related approaches like enactivism (Chemero, Reference Chemero2009) and its more recent offspring, ecological neuroscience, take a rather different approach to cognitive modeling (de Wit, Reference de Wit2024; Falandays et al., Reference Falandays, Yoshimi, Warren and Spivey2024; Favela, Reference Favela2024). One motivation behind these approaches is the assumption that brain-body-environment-spanning solutions to cognitive problems are likely to be energetically cheaper than exclusively neural solutions and therefore favored during natural selection whenever possible (Anderson, Richardson, & Chemero, Reference Anderson, Richardson and Chemero2012; Barrett, Reference Barrett2011; Gibson, Reference Gibson1966; Reed, Reference Reed1996). Some or all of the information needed to perform a particular cognitive capacity may be found in organisms’ richly structured environments and made available through bodily action and thus need not be reconstructed internally (Warren, Reference Warren2021). Exploitation of a species’ particular embodiment is another way in which the performance of a cognitive capacity can be simplified or even solved (Chiel et al., Reference Chiel, Ting, Ekeberg and Hartmann2009; Pfeifer, Iida, & Lungarella, Reference Pfeifer, Iida and Lungarella2014).

In H&C’s taxonomy of how metabolic considerations can be incorporated, from not addressing them at all to estimating the specific numeric values of the energetic costs associated with a model of a particular cognitive capacity, researchers in the ecological tradition can be classified as presenting substantive arguments for the importance of metabolic considerations when modeling cognitive capacities but otherwise largely backgrounding metabolism. In order to strengthen the competitiveness of ecological models relative to neurocentric, computational models, it would be beneficial to heed H&C’s call and move from logical argument to an explicit integration of metabolic considerations in as specific a sense as possible (as discussed in Section 2.1.5). H&C provide clear guidelines for evaluative metabolic considerations in neurocentric, computational modeling. What would such an enterprise look like for ecological modeling?

Here, I can only speculate and gesture at some potentially relevant prior work. Collins et al. (Reference Collins, Ruina, Tedrake and Wisse2005) compared the efficiency of human walking with passive dynamic walkers, which solve walking entirely via interactions between body morphology and environment, and with Honda’s ASIMO robot, which solves it via central control of joint angles. They found remarkable similarity between human walking and passive dynamic walking, both of which were an order of magnitude more efficient than ASIMO. MacIver, Patankar, and Shirgaonkar (Reference MacIver, Patankar and Shirgaonkar2010) showed in electric fish that reducing swimming efficiency to improve sensing increases prey encounter rates, which more than makes up for the additional energy expended in swimming inefficiently. Finally, as with neural tissue, there are also methods for estimating metabolic cost in non-neural tissue, including skeletal muscle tissue (Javed et al., Reference Javed, He, Davidson, Thornton, Albu, Boxt, Krasnow, Elia, Kang, Heshka and Gallagher2010) and even bone (Zhang, Modlesky, & McCully, Reference Zhang, Modlesky and McCully2021).

I end by raising two problems for cognitive modeling as a whole when broadening the space of possible models to also include more thoroughly evolutionary-informed, ecological models. First, for a variety of reasons, evolution by natural selection need not always produce perfectly optimal solutions (Withagen & Chemero, Reference Withagen and Chemero2009; Withagen, de Wit, & van der Kamp, Reference Withagen, van der Kamp and de Wit2018). This complicates model evaluation. It is important to develop additional evaluative criteria that are sensitive to the possibility that a less efficient (but in theory metabolically viable) model may nevertheless be more plausible than a more efficient model. Second, once exclusively neural and ecological models are both included in the space of possible models, it will be important to develop ways to directly compare these model classes with very different metabolic considerations. For example, metabolic considerations related to the cost of deploying non-neural tissue in support of a cognitive capacity will by definition not be present in neurocentric models of cognition. Conversely, the number of bits sent through a neural pathway per ATP used will not be a meaningful metric for ecological models in which environmental ecological information is generated and detected via bodily action rather than reconstructed via internal processing.