The Cognitive Lego: Why the Brain Outperforms Transformers at Flexibility

The quest to build a "thinking machine" has often been a race toward scale: more parameters, more compute, more data. Yet, while Large Language Models (LLMs) like the transformer architectures powering today's AI have achieved near-human proficiency in static tasks, they remain notoriously brittle when faced with the need for true generalization.

A new study published in Neuron this July offers a compelling explanation for this gap. By identifying a method to disentangle the brain's "reusable building blocks," researchers have provided fresh evidence that biological intelligence functions less like a monolithic processor and more like a dynamic assembly line of "cognitive Legos."

The "Building Blocks" Discovery

In a study led by researchers at Northwestern Medicine, the team introduced Sparse Component Analysis (SCA) to analyze neural populations. For decades, neuroscience has struggled with the "mixed signals" problem: individual neurons often participate in multiple computations simultaneously, making it difficult to isolate specific functional roles. [1]

The SCA method succeeded in separating these mixed signals into distinct, interpretable components. The researchers discovered that the brain does not "learn" a brand-new neural pattern for every new action. Instead, it maintains a library of reusable building blocks (fundamental neural sub-routines) that it recombines to generate a near-infinite variety of behaviors. [2]

For example, when studying motor cortex activity, the team found that the same neural components used to reach outward are effectively repurposed to pull the arm back. The brain isn't recording two different "movements"; it is calling upon the same foundational motor-primitive sub-routine twice in different contexts. [2]

Biological Recombination vs. Transformer Modularity

To understand why this is a revolutionary insight for AI, we must look at how transformers differ from this biological reality:

1. The Nature of "Modularity"

Transformers: These models are modular in an architectural sense. They are built from stacks of identical transformer blocks (each containing self-attention and MLP layers). While this allows for efficient scaling, it is largely a static arrangement. The "modularity" is baked into the hardware design, not necessarily the internal logic of the weights. [3]
The Brain: Biological modularity appears to be functional and dynamic. It isn't just that the brain is split into regions; it is that the brain can "snap together" transient patterns of activity to solve a novel problem. This is a form of compositionality that remains a primary hurdle for LLMs. [1, 2, 4]

2. Catastrophic Interference

One of the most persistent issues in AI is "catastrophic interference." When you teach a transformer to perform a new task (e.g., coding) without careful fine-tuning, it often "forgets" how to perform previous tasks (e.g., creative writing). [5, 6]

Biological brains avoid this through these reusable building blocks. By recombining existing cognitive "Legos," the brain preserves the integrity of older skills while integrating new ones. The logic is analogous to a computer program: if you have a well-tested calculate_color() function, you don't need to rewrite it to use it in a "sorting" task; you simply map its output to the new objective.

The Scaling Wall: Why Brute Force Is Reaching Its Limit

While engineers have pushed the transformer paradigm to its limits, recent evidence suggests that the era of "more is better" is encountering significant friction. The scaling laws that once guaranteed predictable performance gains are beginning to flatten, signaling a fundamental shift in the development of artificial intelligence. [7]

The Exhaustion of High-Quality Data

The primary fuel for LLMs (high-quality, human-generated text) is a finite resource. Projections suggest that the stock of public internet text could be fully exhausted by 2026 to 2028. As models consume this data multiple times over, they hit diminishing returns, encountering redundant or noisy content that offers little signal for further intelligence. [8]

The Trap of Synthetic Data

To bypass this, researchers have turned to training models on data generated by other AIs. However, this creates a "closed loop" that risks model collapse. When models train recursively on their own outputs, they lose the diversity and "tails" of the original human data distribution, leading to a homogenization of thought and a degradation of reasoning capabilities. Unlike the brain, which continuously incorporates novel, ground-truth sensory experience, LLMs face a crisis of self-reference. [8]

Non-Linear Scaling and Diminishing Returns

The relationship between scale and capability is becoming increasingly expensive. We are observing that:

Compute vs. Utility: A 10x increase in computational power no longer yields a proportional 10x increase in capability. [7]
Architectural Bottlenecks: Transformer models pay a quadratic cost for self-attention, making them inefficient for long-horizon reasoning. As models grow, they are often bottlenecked by memory movement rather than raw processing speed, further complicating the scaling effort. [7]

Why This Matters for Engineering

If we are to move beyond the current plateau of LLMs, we may need to shift our focus from raw scaling to compositional learning.

The takeaway for engineers is clear: the path to "frontier" intelligence may not lie in simply adding more layers to a network, but in designing architectures that allow for the dynamic reuse of internal representations. If we can develop AI architectures that treat their learned features as a library of functions to be called and recombined (rather than a static, unchangeable state) we may finally bridge the gap between "statistical prediction" and "true flexibility."

The brain’s strategy is elegant, efficient, and, most importantly, modular in a way that allows it to learn without overwriting its own history.

References

Zimnik, A., et al. (2026) "Sparse component analysis: A method that uncovers separable computations within neural population activity" Neuron
Dimmer, O. (2026). "New Analytic Method Reveals 'Building Blocks' of Brain Activity." Northwestern Medicine News
Vaswani, A., et al. (2017). "Attention is All You Need." arXiv
Tafazoli, S., et al. (2025). "Building compositional tasks with shared neural subspaces." Nature
Imanov, A. (2026). "Mechanistic Foundations of Forgetting in Transformer LLMs." Emergent Mind
Fernández-Peña, C. (2026). "Autonomic C1 neurons promote anxiety via activation of vlPAG." Neuron
Haider, Z. (2026). "Scaling Laws Are Slowing Down so Now What?" Medium
Shen, T. (2026). "Will LLMs Scaling Hit the Wall? Breaking Barriers via Distributed Resources on Massive Edge Devices" arXiv)