When the System Breaks, the Architecture Shows

Somewhere in a second-grade classroom right now, a child who can recite poetry, build elaborate block towers, and navigate the social dynamics of the lunch table is staring at the number 7 and feeling nothing quantitative at all. The symbol is there. The magnitude isn't. Seven doesn't arrive heavier than three, or lighter than thirty. The numerals float on the page like decorative glyphs in a language nobody ever translated.

This is developmental dyscalculia — a specific learning disability affecting roughly 5–7% of children — and for most of its history, we could describe it from the outside without quite understanding what was actually broken on the inside. General attention? Working memory? Processing speed? The usual suspects, assembled in the usual lineup.

In 2025, a team at Stanford decided to build the failure instead.

Rather than observing dyscalculia children and cataloguing what was missing, Strock et al. (2025) constructed a deep neural network designed to mimic the dorsal visual pathway — the circuit the brain uses for spatial and numerical reasoning — and then broke it in precisely the ways that dyscalculia breaks the biological version. The digital twin reproduced the behavioral and neural signatures of children with dyscalculia almost exactly. More importantly, it told them where the problem lives: not in attention, not in working memory, but in the formation of number-selective representations in parietal cortex analogs. The abstract concept of magnitude — the thing that makes 7 feel heavier than 3 — requires a specialized representational structure in a specific neural circuit, and in dyscalculia, that structure fails to form properly.

This is the logic of lesion studies, applied at a new resolution. Break something precisely; learn what it was for.

---

Developmental science has always learned fastest from where development goes differently — not as a rhetorical claim but as a methodological one. Simon Baron-Cohen's foundational work on autism in the late 1980s didn't just describe a clinical syndrome; it revealed that the capacity to model other minds is a discrete, separable cognitive module that can be selectively spared or impaired independently of general intelligence. Children who could solve complex puzzles couldn't reliably predict that another person might believe something false. The inference that should be automatic — she doesn't know what I know — simply didn't fire. Autism made the mind's social architecture legible by showing it could come apart.

A 2025 paper in npj Artificial Intelligence made a curious echo of that discovery, but from the machine side. Wu et al. (2025) set out to locate exactly where theory-of-mind capabilities are encoded in large language models — to find, in a sense, the neural correlates in silicon. The answer was startling: ToM-like behavior in LLMs is encoded in an extraordinarily sparse subset of the model's parameters. Perturbing as little as 0.001% of those specific weights significantly degrades performance on theory-of-mind tasks. And these sensitive parameters cluster in one place — the positional encoding module, particularly in architectures using Rotary Position Embedding.

Zero-point-zero-zero-one percent. The social reasoning capacity of a large language model hangs on a few thousand parameters out of hundreds of billions.

This does not settle whether LLMs have theory of mind in any meaningful sense — the question of whether they genuinely model others' beliefs or simply pattern-match on the textual surface features of social reasoning remains genuinely open. But the sparseness is itself informative. In the human brain, theory of mind is distributed across a network: the temporoparietal junction, medial prefrontal cortex, the superior temporal sulcus, the amygdala for emotional inference. It is a system with multiple components that can be separately damaged, as autism research so painstakingly revealed. In a language model, the "equivalent" — if that word applies at all — is a whisper concentrated in how the model tracks positional relationships between tokens.

Different architectures. Different failure modes. The same job done in almost incomparably different ways.

---

Underneath dyscalculia, autism spectrum conditions, language disorders, and the whole taxonomy of developmental divergence runs a more fundamental question: what is the actual engine?

Romberg and Saffran (2025) argue in a comprehensive review that the answer is statistical learning — the ability to extract regularities from environmental input using transition probabilities. An 8-month-old discovers that in the stream of syllables washing over them, certain sounds follow other sounds more predictably than others, and from those regularities, they carve out words. The mechanism is domain-general, operating across auditory, visual, and motor modalities alike. It bootstraps word segmentation, syntactic acquisition, and social prediction. It is not one tool among many, they argue, but the foundational computational engine that makes everything above it possible — a base layer the rest of development builds on.

Which makes it striking that so many neurodevelopmental conditions involve disruptions specifically to statistical learning mechanisms. Dyslexia involves difficulty tracking the phonological transition probabilities that give rise to the sub-lexical structure of words. Specific language impairment involves impaired extraction of grammatical regularities from the speech stream. Some researchers argue that ADHD partly involves dysregulated temporal prediction — difficulty with the sequential structure of time itself.

When the foundational statistical machinery falters, the whole developmental hierarchy above it becomes unstable. Not broken at the top; undermined at the base.

AI systems are statistical learning engines by design — transformers are, at a certain level of abstraction, extraordinarily efficient transition-probability machines. But as Yiu, Kosoy, and Gopnik (2024) document with systematic precision, the way AI systems use those statistics is fundamentally different from how children do. AI models are cultural transmission engines: they reproduce the statistical patterns in their training data with extraordinary fidelity, but they lack the capacity for genuine causal understanding, selective imitation, and innovation that define children's learning. A child doesn't just track what follows what — they ask why, they discard causally irrelevant steps, they invent new solutions when imitation fails.

There is a pointed irony here: the child with autism who over-imitates, mirroring procedures without extracting their causal logic, is doing something that looks more like what a language model does than what typical children do. Pattern-following without the pragmatic filter that extracts intent. The deviation from typical development ends up, accidentally, describing the norm for machine learning.

The nature of the divergence teaches you something about the nature of the norm.

---

What we have, then, is a methodology that has operated quietly in developmental science for decades and is only now becoming legible in AI research. You build a model of learning. You break it in specific ways — or you find cases where it broke on its own. You study the failure mode. You learn what the component was for by understanding what goes wrong without it.

Strock et al. (2025) built a failing model and learned that numerical magnitude is not a general-purpose computation — it is a specialized structure that forms in a specific place through a specific developmental process. Wu et al. (2025) found that social reasoning in LLMs is concentrated in a location nobody would have predicted, encoded in a way nobody would have designed deliberately. Romberg and Saffran (2025) traced the whole hierarchy of language development back to a statistical foundation that, when disrupted, takes everything above it down.

The question that opens up at the end of all this is whether we can run the logic in reverse. If studying atypical development reveals the architecture of typical minds, what do the specific failure modes of large language models — hallucination, contextual amnesia, brittle generalization, overconfident pattern-following — reveal about the architecture of machine learning itself? Are these the lesion studies of AI, if we know how to read them?

And if we built a system that failed in precisely the same ways that children's development fails — the same circumstances that produce dyscalculia analogs, the same conditions that disrupt statistical learning's foundational machinery — would we have found something like a genuine correspondence between mind and machine?

Or would we discover that the breakage patterns are too unlike to illuminate each other?