Integrating metabolism into cognitive science is essential. Beyond a mere constraint, metabolism may be constitutive of cognition itself. This requires expanding metabolism beyond neuronal energy use to include glial contributions, compartmentalized and dynamic processes, and broader brain–body interactions (e.g., interoception). Capturing this complexity will require more refined, integrative models, potentially leveraging artificial intelligence to unify multiscale biological data.

Multiple realizability (MR) is not necessarily unclear nor does it purely operate at the computational level. To understand potential relationships between MR and other constraints, such as metabolic, we formalize possible meanings of function in cognitive science. We build on these to formalize MR, thus resolving its apparent vagaries. Importantly, MR formalisms meaningfully guide and constrain theory building.

Haueis and Colaço argue that motivational/computational models of effort should incorporate metabolic factors. But metabolic and motivational accounts predict the same behavioral patterns, making their contributions hard to distinguish empirically. Moreover, philosophy provides independent reasons favoring motivational accounts: effort is unified across domains and normatively significant in ways metabolic factors cannot capture.

We argue that under minimal thermodynamic and control-theoretic assumptions, metabolic constraints imply generative cognition. We formalize cognition as a resource-bounded controller of probability flows, where metabolic budgets define path-space divergence constraints. This reframes inference as a physically instantiated process governed by metabolic regulation, implying that progress in the cognitive sciences requires grounding cognitive models in nonequilibrium thermodynamics and dynamical systems.

Haueis and Colaço point out that mammalian brains are much more energy efficient than AI supercomputers because the former use orders of magnitude less energy. I argue this issue is more nuanced, and the “efficiency” is scale dependent. Contrary to popular belief, I show that the human brain uses much more energy per operation than the most powerful AI supercomputers.

We build upon Haueis and Colaço’s account in two ways. First, we argue that the backgrounding of metabolism was a pragmatic response to a data granularity problem faced by traditional cognitive modelers. Then, we suggest that equivalence demonstrations offer a promising pathway for incrementally introducing metabolic considerations into extant modeling practice.

Explicitly modeling metabolic constraints remains rare in cognitive neuroscience due to methodological, empirical, and conceptual challenges. Although metabolic limits already shape the biological data that many models are fitted to, direct measurements of energetic costs remain scarce and difficult to relate to model variables. We argue that recent experimental and computational advances may soon enable more explicit integration of metabolism into cognitive models.

Haueis and Colaço compellingly argue for integrating metabolic considerations into cognitive modeling. We suggest that current discussions suffer from a scale mismatch: metabolic constraints are typically framed at global or microscopic levels, whereas cognitive models may operate at a systems level. We propose mesoscale metabolic properties of functional neural networks as a more promising interface for constraining cognitive models.

This commentary emphasizes the importance of how we individuate cognitive capacities, operations, and vehicles. It challenges the target article’s reliance on mechanistic notions of computation and medium-independence and uses examples from the memory literature to suggest that the role played by metabolic considerations in cognitive models will depend on questions of taxonomy and cognitive ontology.

Building on Haueis and Colaço, we argue that metabolic considerations should also constrain cognitive architecture. Globally shared and dynamic resources make strong modularity difficult to justify. Rather than fixed features, modules may reflect local optima within a metabolically constrained landscape – stable under typical conditions but liable to break down under perturbation. Metabolic perturbations thus provide diagnostic tools for architectural assumptions.

Haueis and Colaço claim that metabolic considerations function as explanatory constraints. I argue the explanatory contribution of metabolic constraints for synaptic models of memory is not well delineated. While metabolic constraints determine which synaptic architectures are biologically possible, they do not structure the outcome of synaptic consolidation. Metabolic constraints are explanatory since they account for the background conditions of memory formation.

Survival depends on efficient energy management, and neural systems are highly metabolically demanding biological structures. We fully endorse Haueis and Colaço’s call to integrate metabolism as a fundamental constraint on cognitive modelling. However, by focusing primarily on the neuronal level, their analysis underplays system-level approaches that have already incorporated metabolic constraints into successful and predictive accounts of cognition.

Metabolic considerations are argued to drive cognitive models. I explore the related high-level hypothesis that the brain favours the most anatomically specialised system that still meets demands for computational flexibility in a given task. The hypothesis points to specific claims across a hierarchy of specialisation, from inter-neuronal connection structure to neuronal cell structure, through to non-neuronal cells and analogue mechanisms.

Although abstract computational models in neuroscience omit biochemical detail, metabolic considerations still apply through abstract measures of efficient information processing. Against Haueis and Colaço’s claim that such measures are too generic to discriminate among competing models, I argue that coding efficiency and related measures are commonly applied to evaluate and generate specific cognitive models.

The target article underscores that metabolic considerations can accommodate the standard functionalist account of cognition. However, the energetic makeup of (biological) information processing systems in fact supports the claim of medium dependence of cognitive capacities. We explore two case studies from active matter physics and basal cognition to highlight the philosophical implications of the physicality of information processing.

Haueis and Colaço’s framework can be inscribed into a positive theory of metabolic rationality where cognitive strategy selection is governed by allostatic energy management across sequential decisions under uncertainty. Heuristics emerge as solutions to a metabolic opportunity cost problem with distinct empirical predictions.

The central distinction – between evaluative and generative metabolic considerations – appears less theoretically secure than Haueis and Colaço suggest. First, the generative examples may ultimately reduce to evaluative reasoning applied earlier in the modeling process. Second, the principal generative examples do not clearly show metabolic considerations functioning within a cognitive model.

This commentary extends recent accounts of metabolic constraints in cognitive modelling by considering their role in brain development. We argue that shifting energy availability both constrains and drives the emergence of functional specialisation and network organisation. Integrating interactive specialisation, we show how metabolic factors shape competitive and cooperative interactions, and how atypical energy metabolism may contribute to neurodivergent developmental trajectories.

We reorient Haueis and Colaço’s metabolic framework of cognitive modeling by drawing on neuroscientific evidence from animal motor control, reinforcement learning, and decision-making to show that cognition emerged to serve the metabolic demands of the whole organism, rather than those of the brain alone. We also illustrate the generative value of our reframing to model cognition in artificial systems.

The authors argue convincingly that metabolic considerations are highly relevant to evaluating and generating models of cognition. However, they understate their relevance to the mind–body problem. Metabolic reasoning about energy consumption should be combined with other constraints of speed, space, and history, providing an alternative to multiple realizability and medium independence: coherent materialism.