The Giant Hippocampus: When AI Treats the Brain as a Block of Homogeneous Tofu
> *"If you ask a neuroscientist what the cortex looks like, he will describe a mosaic: primary visual cortex packed with a dense layer 4 for spatial encoding; motor selection areas dominated by thick layers 5 and 6 for temporal integration. But if you ask an AI researcher what the state-of-the-art model looks like, he will describe one thing, repeated infinitely: a Transformer, using the exact same structure whether it processes text, pixels, or speech."*
---
🏛️ A Misread Cathedral
Imagine walking into a Gothic cathedral—soaring nave, stained glass, pointed arches. Now walk into a modern concert hall with acoustic panels and seats in the round. Both are "architecture," but nobody would scale up the concert hall's blueprint and call it the ultimate form of all buildings.
Yet this is what the AI field has been doing for the past decade.
The Transformer, born in 2017, was originally designed for machine translation, using attention to look back and forward through a sentence. That was clever. But then things spiraled: the same structure was applied to image classification (Vision Transformer), speech recognition (Audio Spectrogram Transformer), protein folding, weather prediction... a design built for sequence modeling was force-fitted into nearly every perceptual modality.
The paper *The Giant Hippocampus*, by Jaeho Seol, uses a sharp metaphor: AI is committing a mistake symmetric to the historical "giant Broca's area fallacy"—just as early neuroscience reduced the complex cortex to a single language area, today's AI treats a sequence-memory system (a functional analogue of the hippocampus) as a master key to general intelligence.
---
🔬 The Cortex Is a Mosaic, Not Homogeneous Tofu
To appreciate the force of this critique, return to the hard evidence of neuroscience.
In 1909, German anatomist Korbinian Brodmann published his famous cortical map. Rather than using a microscope on cells, he observed the density, arrangement, and layering of neuronal somata in sections. The result was startling: the cortex is not a block of homogeneous tofu but roughly 50 functionally distinct areas, each with a unique cytoarchitecture—like buildings in different regions having different styles, materials, and purposes.
Modern techniques sharpened this further. The Julich-Brain project uses probabilistic 3D atlases to precisely define cortical boundaries in standard stereotaxic space. Patch-seq lets scientists record a single neuron's morphology, electrophysiology, and gene expression simultaneously. These studies repeatedly confirm one conclusion: different cognitive functions are implemented by qualitatively different structures, not scaled versions of one template.
Two concrete contrasts:
Primary visual cortex (V1) is like a precision instrument factory. Its layer 4 (granular layer) is extremely developed, receiving visual input from the lateral geniculate nucleus while preserving a fine retinotopic map. Neurons are laid out like a high-precision map, each position corresponding to a point in the visual field. This structure exists not for "general computation" but to solve a specific problem: extracting edges, orientation, and color while preserving spatial precision.
Middle temporal area (MT/V5) is entirely different. As the motion-sensitive hub of the dorsal visual stream, its laminar structure is suited to integrating signals across space and time. Its neurons are sensitive to motion direction, speed, and parallax, coordinating information across multiple time windows. Its "architectural design" resembles a symphony conductor's podium—not processing static images but orchestrating dynamic sequences.
Both areas participate in vision, but their structural differences are not decorative. V1's thick layer 4 makes it a faithful recorder of information; MT's deep-layer output structure makes it a temporal integrator. Form is an a priori promise of function.
---
🏗️ CNNs: The Last Correct Architectural Practice
Interestingly, AI once briefly got this right.
Convolutional neural networks were designed by directly borrowing the visual cortex's hierarchy: local receptive fields modeled the spatial confinement of V1 neurons, hierarchical depth corresponded to progressive abstraction from simple features to complex patterns, and translation equivariance built in robustness to object position changes. These were not learned from data—they were "structural priors" encoded by the architect in advance.
The result: CNNs achieved excellent image recognition performance, and with far less data than later Transformers. Not because CNNs were "smarter," but because they were designed to solve the right problem—like a key cut for a specific lock, easier to open the door than a master key.
The Vision Transformer paper (Dosovitskiy et al., 2021) candidly admitted the disadvantage: trained from scratch on mid-sized datasets, ViT underperformed ResNet of comparable scale. The authors attributed this to "the lack of convolutional inductive biases such as locality and translation equivariance." In other words, the Transformer must spend data rediscovering what the CNN knows out of the box.
That is a crucial clue: structural priors are not constraints but levers. The right architecture makes learning easier, not harder.
---
🎰 The Hardware Lottery and Structural Monoculture
So why does the Transformer dominate everything?
The paper offers a pointed diagnosis: the "hardware lottery." The massively parallel compute of GPUs—especially tensor cores optimized for dense matrix multiplication—made the Transformer the path of least engineering resistance. Not because it is in principle best suited to all tasks, but because it best exploits the existing compute infrastructure.
It is like a city with only a subway system, so all transit planning revolves around subway stations. Subways are efficient, but not the only vehicles. Some places need buses, some need bikes, some need walking. When a city invests everything in the subway and declares "the subway is the ultimate form of transportation," it is defining the problem space by engineering constraints.
Mixture-of-Experts (MoE) is often cited as evidence of architectural diversity, but the paper notes sharply: MoE merely partitions parameters among structurally identical experts. It is not qualitative differentiation but quantitative division of labor—like splitting 100 identical workers into 10 groups with different tasks, all trained identically with the same tools.
True diversity requires different structures. Vision needs local receptive fields and spatial hierarchies. Audition needs varying spectral-temporal resolution. Working memory needs recurrent maintenance and gating. Executive control needs competitive action selection and inhibition. These are not things a single Transformer block handles elegantly—not because the Transformer "isn't big enough," but because its structural priors mismatch these computations.
---
🧩 Transformer = Hippocampus, Not Cortex
The paper's core argument culminates here: from a functionalist perspective, the Transformer is best understood as a functional analogue of the hippocampal formation, not a general-purpose cortex.
The hippocampus is the brain's structure for relational binding and episodic memory. It binds dispersed perceptual elements (who, what, when, where) into coherent event representations and supports sequence prediction over those representations. The Transformer's attention does exactly this: it associates any two positions in a sequence regardless of distance, forming an "all-to-all" relational binding.
Powerful—but specific. The hippocampus is not the visual system, the auditory system, or the motor control system. It is a memory-prediction system. Applying the Transformer to all modalities is like building an entire city's infrastructure in the shape of a hippocampus—libraries, factories, hospitals, and schools all designed as memory palaces. It works (with enough data and scale), but is it an elegant solution?
The author calls this phenomenon "The Giant Hippocampus": the whole AI field standardized on a structure designed for sequence memory, then applied it to every task the hippocampus was never evolved to handle: spectral-temporal speech processing, reward-gated executive control, multisensory binding, working memory maintenance.
---
🌐 Heterogeneous Topological Networks: An Alternative Blueprint
The paper does not stop at critique. It proposes a concrete alternative: the Heterogeneous Topological Network (HTN), a "system of systems."
Imagine a real city: a "visual district" (convolutional pathways with built-in spatial locality), an "auditory district" (spectral-temporal modules with built-in frequency selectivity), a dedicated "memory core" (a Transformer-like episodic binding system), "executive gating" (a basal-ganglia-inspired action selection mechanism), and "working memory buffers" (prefrontal-inspired recurrent maintenance).
These modules are not haphazardly stitched together. They communicate through standardized interfaces—like power grids in different countries interconnected by unified standards. Each module keeps the inductive biases its computation requires, instead of being forced into a shared substrate.
This is not a cognitive-science metaphor. It is a design discipline: specify modular constraints and functional decomposition from structural evidence before training begins. Architecture is not something reverse-engineered after the fact from monolithic behavior—it is an a priori input to training.
---
🎯 Why This Matters
The importance of this paper lies not in a specific architecture (HTN's details remain to be explored) but in reframing the problem.
The mainstream narrative in AI today: scale is everything. More parameters, more data, more compute will eventually yield emergent general intelligence. But this narrative ignores a question: if the base architecture's structural priors mismatch the target computation, scale is just brute force compensating for design flaws.
A century of neuroscience evidence tells us: biological intelligence succeeds not because it has one giant general-purpose processor, but because it has many specialized processors, each optimized for a particular computation, working together through carefully designed interfaces.
Perhaps the next breakthrough in AI will not come from stacking Transformers higher, but from asking a more fundamental question: what structure does the problem we're trying to solve actually require?
---
References
- Seol, J. (2026). The Giant Hippocampus: From Structural Monoculture to a System of Systems. arXiv:2607.19973.
- Brodmann, K. (1909). Vergleichende Lokalisationslehre der Grosshirnrinde.
- Amunts, K., et al. (2020). The Julich-Brain atlas. Science, 369(6506), 1035-1037.
- Dosovitskiy, A., et al. (2021). An Image is Worth 16x16 Words. ICLR 2021.
- Jonas, E., & Kording, K. (2017). Could a Neuroscientist Understand a Microprocessor? PLOS Computational Biology.
- Marr, D. (1982). Vision: A Computational Investigation into the Human Representation and Processing of Visual Information.