Haueis and Colaço correctly address the importance of metabolic constraints on brain processing and brain modeling, and illustrate this with a number of valuable examples.

However, their target article does not appropriately address the influence of metabolic constraints on the nature of human cognition, even though that would be one of its most important topics. In the evolution of human cognition, there would have been an interaction between metabolic constraints and the development of cognitive abilities as required by human-level cognition.

A well-known view is that human-level cognition depends on compositional processing, and the computational architecture that could achieve this is the Von Neumann Architecture (VNA) or variants thereof (Fodor & Pylyshyn, Reference Fodor and Pylyshyn1988; Newell, Reference Newell1990). But the VNA is highly energy inefficient. The target article indeed refers to this and suggests that analog AI chips could solve this energy problem. But this solution throws away the baby with the bathwater. To see why, we have to look a bit at the structure of the VNA, both from an energy and a cognitive perspective.

Basically, the VNA consists of a central processing unit (CPU) connected to a data-bus, which is connected to memory, e.g., given by registers. These are shielded off from the data-bus; otherwise, they would interfere with processing. In the execution of a program, they can be opened by the CPU using “address” codes. In this way, data from memory can be transported to the CPU for processing, or data from the CPU can be transported to memory. This creates the well-known “bottleneck” of the VNA, because at any given time only one of these operations can be carried out (Backus, Reference Backus1978).

The structure of the VNA is important because it relates directly to its energy consumption. Computations in the CPU do not consume most of its energy. Instead, the main energy consumption derives from data transfer over the data-bus (Hess, Reference Hess2025). This begs the question of whether this data transfer is actually needed. The VNA uses it because it is designed as a universally programmable computer. Hence, it also uses it for numerical computations as found in vector and matrix operations. Here, indeed, energy will be reduced by using, e.g., analog AI chips (Mehonic & Kenyon, Reference Mehonic and Kenyon2022).

But these processors do not (yet) provide the compositional processing underlying human cognition, as exemplified in language. For example, a sentence is composed out of words. The sentence structure expresses specific relations between these words, which can be retrieved with relational or “who does what to whom” questions. As noted, a standard view on cognition is that we represent and process sentences in this way as well. An important argument for this was given by Newell (Reference Newell1990), based on the notion that the amount of information stored at any given location in any architecture is limited due to physical reasons.

Hence, information needed for more complex processing will be distributed over an architecture. An example is given by the observation that information related to words is distributed over large parts of the cortex (Huth et al., Reference Huth, de Heer, Griffiths, Theunissen and Gallant2016). Yet, words can be used on the fly to compose and process arbitrary sentences. According to Newell, this requires an architecture that has access to distant and distributed information. Access then allows the transfer of distant information to integrate it with local processing. Much in the same way as the VNA transfers data to process it in the CPU. Yet, the access and data transfer underlying compositional processing would increase energy consumption, as it does in the VNA.

But perhaps compositional processing would not be needed for higher-level cognitive processing. An alternative could be given by large language models (LLMs). They do not represent and process sentences in a compositional way but are trained directly on a huge set of sentences. However, they need to be trained on more sentences than humans will see in their lifetime (Linzen, Reference Linzen, Jurafsky, Chai, Schluter and Tetreault2020). The energy costs of this are even more staggering (Ji & Jiang, Reference Ji and Jiang2026). Perhaps this could be reduced with other kinds of processors (Mehonic & Kenyon, Reference Mehonic and Kenyon2022). But given the number of sentences required for training, it is questionable whether this would result in significant energy reductions.

Moreover, LLMs exhibit a number of problems from a cognitive perspective. For example, they do not reliably answer relational questions, addressing the structure of a sentence, even for sentences that they have learned (Berglund et al., Reference Berglund, Tong, Kaufmann, Balesni, Stickland, Korbak and Evans2024). In contrast, children learn language with a minute subset of sentences compared to LLMs (Saxton, Reference Saxton2010), and they can answer relational questions because they learn and process language in a compositional manner (Marcus et al., Reference Marcus, Vijayan, Bandi Rao and Vishton1999).

Hence, on the one hand, compositional processing reduces energy consumption because it does not require learning, e.g., massive numbers of sentences. But on the other hand, the transfer of information needed for compositional processing increases energy consumption in architectures such as the VNA.

This contrast requires the development of cognitive architectures that are capable of compositional processing in an energy-reduced manner. Promising examples are based on forms of neurosymbolic processing (Hitzler, Sarker, & Eberhart, Reference Hitzler, Sarker and Eberhart2023), or processing akin to that. They could achieve compositional processing in a neural manner. But the energy consumption of information transfer needs to be reduced as well.

A solution for this problem is access to distant information without having to transfer it for processing, as with distributed information that is interconnected in a small-world network structure (Watts & Strogatz, Reference Watts and Strogatz1998). The connection paths in this structure allow access to distant information and can be used to interconnect local processing at different sites, e.g., by shifting attention from one site to another (van der Velde, Reference Van der Velde, Hitzler, Sarker and Eberhart2023). This would reduce energy consumption because it does not require information transfer.

To summarize, the focus on metabolic constraints is important. But on its own it is not sufficient and could even lead modeling in the wrong direction. To be useful, it has to be combined with a clear understanding of its effects on human-level cognitive architecture.