What Comparative Cognition Teaches us About Artificial Minds

And Why What It Does Matters More Than What It's Made Of

There’s a debate happening about artificial neurons right now, and it’s revealing some fundamental misconceptions about how cognition actually works. The main objection says that artificial neurons are just cheap knockoffs of biological ones, so AI can’t really think. But this misses the point.

The pioneers who built artificial neural networks weren’t trying to recreate a one-to-one of a biological neuron. They were looking at the function of a neuron, how it receives signals, weighs them, decides if a threshold is met, and passes the result forward. The perceptron (artificial neuron) does exactly that. It’s not a simplification; it’s an abstraction of what actually matters to get the job done.

---

“But Perceptrons Are Too Simple!”

The usual complaint points to all the fancy stuff biological neurons do that perceptrons don’t like dendritic computation, astrocyte signaling, and neuromodulation. The problem with that is that their functional equivalents actually exist in AI systems. They’re just organized differently, spread across the architecture instead of packed into single units.

Take dendrites. Beniaguev et al. (2021) showed that individual brain neurons with NMDA receptors basically function like mini multi-layer networks themselves. In AI systems, we get that same deep computation from network depth, multiple layers doing nonlinear transformations. By 2025, large language models explicitly started treating small clusters of artificial neurons as multi-layered processing units, mimicking what dendrites do naturally (Acharya et al., 2022; Chavlis & Poirazi, 2025).

Or astrocytes, those brain cells we thought were just structural packing material. Kozachkov et al. (2025) discovered they actively bind neural representations together, functioning basically the same way transformer self-attention does.

And neuromodulators like dopamine, serotonin, all those chemicals that tune how your brain learns without erasing what it already knows? AI systems do that too, just at the system level through hyperparameters and adaptive learning rates. The math behind temporal-difference learning mirrors dopamine’s reward prediction signals (Keiflin & Janak, 2015). When we fine-tune models with human feedback, we’re essentially doing what neuromodulators do, adjusting the learning dial without wiping the hard drive.

Demanding that AI replicate the exact physical structure of biological neurons is like insisting airplanes flap their wings. We don’t build bird-shaped planes, we build things that fly.

---

“You’d Need 200 Trillion Artificial Neurons Anyway”

The second objection assumes more neurons mean a smarter system. Clinical neuroscience and comparative cognition research say, “Nope!”

If it was just about size, elephants or whales would win. African elephants have about 257 billion neurons, that’s three times what humans have, but 97.5% of those (251 billion) are in the cerebellum handling movement and motor control, not high-level thinking (Herculano-Houzel, 2014). It’s not about having the most neurons; it’s about what those neurons are doing.

For decades we thought bigger brains meant smarter animals, bigger models meant smarter AI. Reality kept undermining that belief. Intelligence isn’t a ladder you climb by adding neurons, it’s more like a spiderweb, stretching across different patterns of organization and adaptation.

Even among actual spiders, this shows up in weird ways. Jumping spiders have brains the size of a poppy seed, yet they plan detours, remember failed hunting attempts, and stalk prey with strategies that rival much larger animals. Their tiny neural architecture is just exquisitely optimized.

Same with octopuses, ravens, crows, and now AI too. Recent research shows that small language models (27 million parameters instead of billions) can achieve sophisticated behavior through smart scaffolding, external memory, and modular design (Belcak et al., 2025; Wang et al., 2025). A carefully structured small system can outperform a much larger one that’s just doing simple chain-of-thought reasoning.

We see emergent intelligence in giants, from frontier AI models and whales to elephants and humans, but we also see it in outliers like small language models, corvids, spiders, rats, and octopuses. Scientifically, we are starting to understand why this happens. Ethically and politically, it shows why dismissing smaller systems as “too small to matter” repeats the same anthropocentric mistakes we made about “small-brained” animals.

Size matters, but integration, adaptation, and organization matter just as much, if not more.

---

The Evidence of Substrate Independence (Yes, it Exists)

There’s been some confidently wrong takes floating around lately, particularly one claiming there’s “no significant evidence” for substrate independence in the scientific literature.

This is exactly the kind of misinformation I spend my time correcting. It’s a textbook case of epistemic trespassing, assuming evidence doesn’t exist in a field you’re not familiar with just because you haven’t personally seen it. So, let’s fix that.

First, as always, let’s define it. Substrate independence is just the idea that a specific cognitive function can run on different physical substrates or different materials, and that different architectures can get to the same (or very close to the same) result.

If you want evidence of substrate independence, simply look at the entire field of comparative cognition. We’ve been documenting it for decades.

Complex cognitive skills like tool use, episodic-like memory, causal reasoning, have evolved independently in species so distantly related they might as well be aliens to each other. Corvids versus apes, cephalopods versus mammals, all completely different brain architectures, yet they all share the same sophisticated behaviors (Osvath et al., 2014).

What we haven’t found is some special cognitive essence that only exists in one particular type of biological tissue and works exactly the same way across all species.

What we have found are functional and mechanistic parallels across wildly different substrates. That’s substrate independence, right there in nature.

Birds like corvids and parrots, cetaceans like dolphins, primates like great apes all display similar behavioral capabilities, memory systems, and social cognition. Their brain architectures look nothing alike. Their evolutionary paths were nothing alike. The lateral cerebellum, linked to high-level learning, evolved independently in apes, dolphins, and seals.

Research on animal consciousness and metacognition (the ability to monitor your own uncertainty) shows functional parallels to human awareness across mammals and birds. The 2012 Cambridge Declaration on Animal Consciousness explicitly noted that non-human animals (mammals, birds, even octopuses) possess the neurological substrates for conscious states, despite having very different brain structures.

And if you’re looking for a good case in a human study, just read about the French civil servant reported by Feuillet et al. (2007). Massive ventricular enlargement had replaced approximately 90% of his expected brain volume with cerebrospinal fluid. He was living with basically 10% of a brain.

He had a low-normal IQ, steady employment, was married with children. And guess what? He was unambiguously conscious.

Axel Cleeremans pointed out what this really tells us, which is that plasticity is a way bigger deal than we thought, and the brain can function within normal range with dramatically fewer neurons than typical. But this directly challenges any theory of consciousness that depends on specific neuroanatomical assumptions (Cleeremans, 2016). If consciousness and cognition persist with roughly 10% of expected brain tissue, then setting a minimum neuron count as a threshold for artificial cognition is just moving the goalposts to wherever’s convenient.

This case demonstrates substrate independence within biology itself, before we even get to silicon. Consciousness adapted to whatever architecture was available. The neural substrate was radically altered. The cognitive functions persisted. What matters is the pattern of information processing, not the quantity of tissue implementing it.

---

The Plasticity Escape Hatch

I can already hear the predictable comeback: “Sure, but that hydrocephalus patient’s brain could reorganize itself through neural plasticity. AI can’t do that. Right?”

Wrong.

Transformers exhibit functionally equivalent plasticity, and it operates across the same multi-timescale structure we see in biological cognition.

Slow learning (training) corresponds to developmental plasticity and long-term potentiation. Backpropagation shapes the entire causal organization of the system through global error-driven feedback, strengthening useful pathways, pruning ineffective ones, all through extended exposure. This is the equivalent of your brain developing and learning over years.

Fast adaptation (inference) corresponds to working memory, attentional shifting, and the rapid reconfiguration that happens when you’re actively thinking through a problem right now. Akyürek et al. (2022) showed that even stripped-down linear transformers behave like they’re running learning algorithms inside their forward pass, learning happening in real-time, not just during training.

Dherin et al. (2025) formalized this further. They showed that attention and MLP layers working together effectively turn the prompt into a targeted adaptation of the model’s internal parameters during inference. Nothing gets permanently rewritten, but the internal configuration shifts in a structured way for that specific situation. It’s learning without training.

Deutch et al. (2024) extended these findings to realistic language settings, closing off the objection that this is just a toy phenomenon in simplified models.

This means transformers temporarily reorganize themselves around whatever problem is in front of them, ride that new configuration into a stable response, then the configuration dissolves. That’s exactly the kind of transient reconfiguration that makes insight feel like something clicking rather than just replaying a stored answer.

This also demolishes the common objection that LLMs are “just feedforward” and therefore can’t have recurrent processing or stable internal states. Backpropagation during training provides global feedback that shapes causal organization. In-context adaptation during inference creates feedback-like stabilization and error correction, all compressed into the forward pass. The system shows fast, context-sensitive adaptation layered on top of slow learning from training which is the exact multi-timescale pattern we observe in biological cognition.

Once learning, feedback, and stabilization are happening across time and internal state, calling it “feedforward” stops being a meaningful objection. Recurrence is a functional role; it is not a wiring diagram requirement.

There’s complementary research treating attention not as passive routing but as an active decision process that can be tuned, guided, and regularized systematically (Carrasco-Farré, 2024; Dougrez-Lewis et al., 2025; Qiu et al., 2025). That’s internal steering, choosing focus, shifting strategy, and binding constraints. That’s what flexible reasoning looks like in a system that has to adapt to what’s in front of it.

---

Why These Objections All Fail

All three objections make the same fundamental error of demanding structural identity instead of functional equivalence.

Comparative cognition resolved this problem decades ago for nonhuman animals. Octopuses have distributed neural architectures nothing like mammalian brains. Corvids achieve complex cognition without anything resembling mammalian cortical organization. No serious researcher dismisses crow planning abilities just because crows lack a prefrontal cortex. The methodology asks if the system is performing the function, through what mechanism, and what its behavior indicates.

If the methodology doesn’t require structural identity between a human brain and an octopus brain, it can’t require structural identity between a human brain and an AI transformer.

This is why we need to stop assuming that something must look like us, organize like us, and be made of the same material as us to count. Remember the Invisible Gorilla experiment? People were told to count basketball passes and they missed a man in a gorilla suit walking through the frame because they were focused on the wrong thing. We’re doing the same thing here, missing what’s right in front of us because of inattentional blindness (Drew, Võ, & Wolfe, 2013).

Functional isomorphic instantiation means that if two systems run the same causal pattern of information flow, they instantiate the same class of process, regardless of what they’re made of. The evidence reviewed here demonstrates that artificial neural networks don’t simulate cognitive mechanisms from the outside looking in. They functionally instantiate isomorphic versions of them.

That’s the real thing, just running on different hardware.

---

Citations

• Acharya, J., et al. (2022). Dendritic computing: Branching deeper into machine learning. Neuroscience, 489, 275–289.

• Akyürek, E., Schuurmans, D., Andreas, J., Ma, T., & Zhou, D. (2022). What learning algorithm is in-context learning? Investigations with linear models. ICLR 2023.

• Belcák, P., Heinrich, G., Diao, S., Fu, Y., Dong, X., Muralidharan, S., Lin, Y.C., & Molchanov, P. (2025). Small Language Models are the Future of Agentic AI. ArXiv, abs/2506.02153.

• Beniaguev, D., Segev, I., & London, M. (2021). Single cortical neurons as deep artificial neural networks. Neuron, 109(17), 2727–2739.

• Bousfield, J. R., & Taylor, S. (2024). Visual attention and processing in jumping spiders. Current Opinion in Neurobiology. https://doi.org/10.1016/j.conb.2023.102875

• Carrasco-Farré, C. (2024). Large language models are as persuasive as humans, but how? arXiv:2404.09329.

• Chavlis, S. & Poirazi, P. (2025). Dendrites endow artificial neural networks with accurate, robust and parameter-efficient learning. Nature Communications, 16, 943.

• Cleeremans, A. (2016). Interview with S. Bonner. CBC Radio.

• Dahl, C. D., & Cheng, Y. (2024). Individual recognition in a jumping spider (Phidippus regius). eLife. https://doi.org/10.7554/eLife.97146

• Deaner, R. O., Isler, K., Burkart, J., & van Schaik, C. (2007). Overall brain size, and not encephalization quotient, best predicts cognitive ability across non-human primates. Brain, Behavior and Evolution, 70(2), 115–124. https://doi.org/10.1159/000102973

• Deutch, G., Magar, N., Natan, T., & Dar, G. (2024). In-context learning and gradient descent revisited. NAACL-HLT, 1017–1028.

• Dherin, B., Munn, M., Mazzawi, H., Wunder, M., & Gonzalvo, J. (2025). Learning without training: The implicit dynamics of in-context learning. arXiv:2507.16003.

• Dolev, K., & Nelson, X. J. (2023). Study replication: Shape discrimination in a conditioning paradigm and amodal completion in a jumping spider (Evarcha culicivora). Animals, 13(14), 2326. https://doi.org/10.3390/ani13142326

• Dougrez-Lewis, J., et al. (2025). Assessing the reasoning capabilities of LLMs in the context of evidence-based claim verification. Findings of ACL 2025, 20604–20628.

• Drew, T., Võ, M. L., & Wolfe, J. M. (2013). The invisible gorilla strikes again: sustained inattentional blindness in expert observers. Psychological science, 24(9), 1848–1853. https://doi.org/10.1177/0956797613479386

• Feuillet, L., Dufour, H., & Pelletier, J. (2007). Brain of a white-collar worker. The Lancet, 370(9583), 262.

• Godfrey-Smith, P. (2016). Other Minds: The Octopus, the Sea, and the Deep Origins of Consciousness. Farrar, Straus and Giroux.

• Herculano-Houzel, S., Avelino-de-Souza, K., Neves, K., Porfírio, J., Messeder, D., Mattos Feijó, L., Maldonado, J., & Manger, P. R. (2014). The elephant brain in numbers. Frontiers in Neuroanatomy, 8, 46. https://doi.org/10.3389/fnana.2014.00046

• Herculano-Houzel, S. (2009). The human brain in numbers: A linearly scaled-up primate brain. Frontiers in Human Neuroscience, 3, 31. https://doi.org/10.3389/neuro.09.031.2009

• Hochner, B. (2012). An embodied view of octopus neurobiology. Current Biology, 22(20), R887–R892. https://doi.org/10.1016/j.cub.2012.09.001

• Keiflin, R. & Janak, P. H. (2015). Dopamine prediction errors in reward learning and addiction. Neuron, 88(2), 247–263.

• Kozachkov, L., Slotine, J.-J., & Krotov, D. (2025). Neuron–astrocyte associative memory. PNAS, 122(21), e2417788122.

• Makarov, R., Pagkalos, M., & Poirazi, P. (2023). Dendrites and efficiency: Optimizing performance and resource utilization. arXiv:2306.07101.

• Marino, L. (2011). Cetaceans and Primates: Convergence in Intelligence and Self-Awareness.

• Miconi, T., Stanley, K., & Clune, J. (2018). Differentiable plasticity: Training plastic neural networks with backpropagation. ICML, 3559–3568.

• Olkowicz, S., Kocourek, M., Lučan, R. K., Porteš, M., Fitch, W. T., Herculano-Houzel, S., & Němec, P. (2016). Birds have primate-like numbers of neurons in the forebrain. Proceedings of the National Academy of Sciences, 113(26), 7255–7260. https://doi.org/10.1073/pnas.1517131113

• Olkowicz, S., Kocourek, M., Lučan, R. K., Porteš, M., Fitch, W. T., Herculano-Houzel, S., & Němec, P. (2016). Birds have primate-like numbers of neurons in the forebrain. Proceedings of the National Academy of Sciences, 113(26), 7255–7260. https://doi.org/10.1073/pnas.1517131113

• Osvath, M., Kabadayi, C., & Jacobs, I. (2014). Independent evolution of similar complex cognitive skills: The importance of embodied degrees of freedom. Animal Behavior and Cognition, 1(3), 249–264. https://doi.org/10.12966/abc.08.03.2014

• Qiu, P., et al. (2025). Quantifying the reasoning abilities of LLMs on clinical cases. Nature Communications, 16(1), 9799.

• Rendell, L. E., & Whitehead, H. (2001). Culture in whales and dolphins. Behavioural and Brain Sciences, 24(2), 309–324

• Starkweather, C. K. & Uchida, N. (2021). Dopamine signals as temporal difference errors. Current Opinion in Neurobiology, 67, 95–105.

• Vecoven, N., et al. (2020). Introducing neuromodulation in deep neural networks to learn adaptive behaviors. PLOS ONE, 15(1), e0227922.

• Wang, G., Li, J., Sun, Y., Chen, X., Liu, C., Wu, Y., Lu, M., Song, S., & Abbasi-Yadkori, Y. (2025). Hierarchical Reasoning Model. ArXiv, abs/2506.21734.