The Giant Hippocampus: When AI Treats the Brain Like a Block of Homogeneous Tofu
> *"If you ask a neuroscientist what the cortex looks like, they will describe a mosaic: primary visual cortex packed with a dense granular layer for spatial encoding; motor selection areas dominated by thick layers five and six for temporal integration. But if you ask an AI researcher what the most advanced model looks like, they will describe one thing, repeated infinitely: a Transformer—identical in structure whether processing 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 no one would enlarge a concert hall's blueprint and declare it the ultimate form of all buildings.
Yet this is exactly what AI has done for the past decade. The Transformer, born in 2017 for machine translation, uses attention to let models look back and forth across a sequence. Clever—but then things spiraled: the same structure was applied to image classification (Vision Transformer), speech recognition (Audio Spectrogram Transformer), protein folding, weather forecasting. A design built for sequence modeling was forced into nearly every perceptual modality.
Jaeho Seol, author of *The Giant Hippocampus*, offers a pointed metaphor: if early neuroscience wrongly reduced the complex cortex to a single language area (the "giant Broca fallacy"), today's AI is treating a sequence-memory system (a functional analog of the hippocampus) as a universal key to general intelligence.
---
🔬 The Cortex Is a Mosaic, Not Homogeneous Tofu
In 1909, Korbinian Brodmann published his famous cortical map. By examining the density, arrangement, and layering of neuronal cell bodies, he showed the cortex divides into roughly 50 functionally distinct areas, each with its own cytoarchitecture. Modern tools sharpened the picture: the Julich-Brain project provides probabilistic 3D atlases of each area's boundaries, and Patch-seq simultaneously records single neurons' morphology, electrophysiology, and gene expression. The repeated conclusion: different cognitive functions are implemented by qualitatively different structures, not scaled copies of one template.
Two concrete contrasts:
- Primary visual cortex (V1) is like a precision instrument factory. Its massively developed layer 4 receives visual input from the lateral geniculate nucleus with a fine retinotopic map—built not for general computation but to extract edges, orientation, and color while preserving spatial precision.
- Middle temporal area (MT/V5), a motion-sensitive area in the dorsal stream, has layering better suited to integrating signals across space and time—neurons tuned to motion direction, speed, and disparity. It is an orchestrator of dynamic sequences, not a recorder of static images.
- 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.
Both are visual areas, but the structural differences are not decorative. V1's thick layer 4 makes it a faithful recorder; MT's deep output layers make it a temporal integrator. Form is a prior commitment to function.
---
🏗️ CNNs: The Last Correct Architectural Practice
AI briefly got this right. Convolutional neural networks borrowed directly from visual cortex: local receptive fields, hierarchical depth, translation equivariance—structural priors encoded by designers, not learned from data. CNNs achieved excellent image recognition with far less data than later Transformers, because they were designed to solve the right problem—a key cut for a specific lock.
The ViT paper (Dosovitskiy et al., 2021) candidly admitted this: trained from scratch on mid-sized datasets, ViT underperforms similarly sized ResNets, due to lacking "the inductive biases inherent to convolution, such as translation equivariance and locality." In other words, Transformers must spend data rediscovering what CNNs know by construction. Structural priors are not constraints; they are levers.
---
🎰 The Hardware Lottery and Structural Monoculture
Why does the Transformer dominate everything? The paper's diagnosis: the hardware lottery. GPUs—especially tensor cores optimized for dense matrix multiplication—make Transformers the path of least engineering resistance. Not because they are best in principle for all tasks, but because they best exploit existing compute infrastructure.
Mixture-of-Experts is often cited as architectural diversity, but the paper retorts: MoE partitions parameters among structurally identical experts. It is quantitative division of labor, not qualitative differentiation—100 identical workers split into ten groups. Real diversity needs different structures: vision needs local receptive fields and spatial hierarchy; audition needs varying spectro-temporal resolution; working memory needs recurrent maintenance and gating; executive control needs competitive action selection and inhibition. Transformers struggle with these not because they are "too small," but because their inductive biases mismatch these computations.
---
🧩 Transformer = Hippocampus, Not Cortex
The paper's core claim: functionally, the Transformer is best understood as an analog of the hippocampal formation, not a general cortex. The hippocampus binds distributed perceptual elements (who, what, when, where) into coherent episodic representations and supports sequence prediction. Attention does exactly this: it relates any two positions in a sequence, regardless of distance—a form of all-to-all relational binding.
Powerful, but specific. Applying Transformers to every modality is like building an entire city's infrastructure in the shape of a memory palace. The author calls this "The Giant Hippocampus": a field standardized on a sequence-memory structure, then applied to every task evolution never shaped the hippocampus for—spectro-temporal speech processing, reward-gated executive control, multisensory binding, working memory.
---
🌐 Heterogeneous Topological Networks: An Alternative Blueprint
The paper proposes a concrete alternative: the Heterogeneous Topological Network (HTN), a system of systems. Picture a real city: a "visual district" (convolutional pathways with spatial locality), an "auditory district" (spectro-temporal modules with frequency selectivity), a "memory core" (Transformer-like episodic binding), "executive gating" (basal-ganglia-inspired action selection), and "working memory buffers" (prefrontal-inspired recurrent maintenance).
Modules communicate through standardized interfaces, like power grids across countries. Each keeps the inductive biases its computation requires. This is not a metaphor—it is a design discipline: specify modular constraints and functional decomposition before training begins, based on structural evidence. Architecture is a prior input to training, not reverse-engineered from monolithic behavior afterward.
---
🎯 Why It Matters
The paper's importance lies less in its specific architecture (HTN's details remain open) than in reframing the problem. The mainstream narrative says scale is everything. But if the base architecture's inductive biases mismatch the target computation, scale merely brute-forces design flaws. A century of neuroscience shows biological intelligence succeeds not through one giant general-purpose processor, but through many specialized processors working together through carefully designed interfaces.
Perhaps the next breakthrough will not come from stacking Transformers higher, but from asking a more fundamental question: what structure does the problem we want to solve actually require?
---
References