From Cortex to Cerebellum: Teaching Machines to Think Like Children

People keep asking me what I've been up to lately. The honest answer: I've been going down a rabbit hole so deep that the only way out is to write about it. After taking a course in computational neuroscience, something clicked — and the way I see AI models has never been the same since.

Two Systems, One Brain

You've probably encountered Daniel Kahneman's famous "Thinking, Fast and Slow" framework. System 1 is fast, effortless, intuitive. System 2 is slow, deliberate, computational. Kahneman was talking about human cognition, but I've become obsessed with what this looks like in the brain's actual hardware.

The human brain can be loosely divided into three evolutionary layers. The oldest, inherited from our reptilian ancestors, contains the cerebellum — it governs motor control and is wired directly to the spine. The middle layer, from our mammalian lineage, houses structures like the hippocampus and governs emotion. The newest, outermost layer — uniquely human — is the cerebral cortex, the seat of complex reasoning and language.

Yes, this is a simplification. Yes, neuroscientists will wince. But it's a beautifully useful simplification, so let's keep moving.

Here's the analogy that keeps me up at night: modern LLMs map almost perfectly onto the cortex. They are System 2 at industrial scale — massive neural networks processing language with probabilistic, associative machinery. The cerebellum equivalent in AI? That's where things get interesting.

Intelligence is the ability to consciously raise your entropy to think in System 2 and solve problems — then lower it to compress the learnings into intuition and push them into System 1.

Statistical Parrots vs. World Models

Ask an LLM: "If I hold an ice cube over fire, it will..." — it'll answer "melt." Correctly. But not because it understands thermodynamics. It answers correctly because that word appeared most frequently after that sequence of tokens in its training corpus. It's performing very sophisticated autocomplete.

Contrast this with what Yann LeCun calls world models — architectures like JEPA that attempt to build an internal representation of physics. Such a system would know that matter has phases, that heat drives transitions between them, and that solid → liquid is called melting. It doesn't retrieve the answer; it derives it.

LLMs are brilliant parrots. World models are toddlers who have actually touched a hot stove. Guess who learns faster next time.

This distinction — parroting vs. understanding — is at the core of what I'd call AI 1.0 vs. AI 2.0. AI 1.0 was the arms race for scale: the largest models, the most parameters, the most megawatts. AI 2.0 is the pivot toward efficiency, distillation, and intuition.

The XOR Problem: A Case Study in Intuition

Let me make this concrete with an example that has become something of an obsession for me: the XOR function.

XOR is notoriously unfriendly to neural networks. A standard feedforward net needs multiple hidden layers just to learn this simple binary function. Given two input bits, XOR outputs **1** only when they differ: (0,0)→0, (0,1)→1, (1,0)→1, (1,1)→0. A cortex-style neural net would look something like this:

The cortex approach: 2 inputs → 4 hidden nodes → 1 output. Matrix multiplications, activations, the works. Elegant, but expensive.

At inference time, this architecture costs 8 multiplications, 7 additions, and 4 nonlinear activations (tanh). Respectable. But is it the most efficient representation of XOR?

Here's what a cerebellum would do. XOR, at its core, is just parity detection: count the number of 1s in the input. If the count is odd, output 1; if even, output 0. Now encode this geometrically:

Start with a point on the "even" side. For each input bit: if it's **1**, flip the point to the opposite side; if it's **0**, leave it. After processing all bits, read off which side you're on.

That's it. Two operations. The answer is geometrically encoded. The computational cost isn't 8 multiplications — it's essentially counting. This is what I mean by intuition: the rule has been compressed to its geometric essence, and inference becomes a series of lightweight, deterministic acts.

This also generalizes to any number of bits. XOR of 100 inputs? Still just parity. The adult cortex is sweating over tensor math while the cerebellum child is just flipping a coin — and winning every time.

The Sleep Hypothesis

There's a mysterious process that happens during sleep. The brain cleanses itself — cerebrospinal fluid flushes through, metabolic waste is cleared — and simultaneously, the events of the day are replayed and compressed. Episodic memories become semantic ones. Skills become reflexes. Knowledge migrates from the cortex to the cerebellum.

This "distillation during sleep" is still largely a mystery physiologically, but conceptually it's the most beautiful idea in all of neuroscience: the brain spends its downtime converting expensive knowledge into cheap intuition. System 2 runs its lessons through some unknown compression algorithm, and System 1 wakes up a little wiser.

The cerebellum doesn't store facts. It stores shapes. It encodes the geometry of skill, the topology of habit, the manifold of intuition.

AI 2.0: The Race for Intuition

The AI race of the past five years was about scale. Who can train the largest model? Who can afford the most compute? GPT-4, Gemini, Claude — these are cerebral cortices in silicon: immensely capable, immensely expensive.

But the next race is different. The humanoid robotics boom has changed the constraints entirely. A robot can't afford to run a transformer inference pass every time it needs to decide whether to catch a falling glass. It needs something closer to the cerebellum: fast, cheap, nearly hardwired.

This is the emerging art of model distillation — the process of compressing a large, expensive model into a smaller, faster, more specialized one that captures the intuition behind a task without the full computational weight of the original. Not local intelligence in the old sense (brittle, narrow, unable to generalize), but compressed intelligence: still capable of perceiving the world meaningfully, but encoded in a circuit-like architecture that runs on a battery.

The question that keeps AI researchers up at night: what are the right distillation techniques? How do you take a 300-billion-parameter model and extract the geometric essence of what it knows — the sphere-and-flip representation of XOR, the phase-ladder representation of thermodynamics — and hard-wire it into something that can run in real time on a robot's onboard compute?

A robot falling to the ground doesn't need eloquence. It needs reflex. There's a difference between knowing how to write a poem about catching yourself and actually catching yourself.

What's Next

I've been building a library of examples — what I call Intuition Distillation case studies — where I map the high-entropy, cortex-level computation of a task onto the low-entropy, cerebellum-level geometric representation. XOR is just the first. The same framework applies to modular arithmetic (navigation on a circle), coupled oscillator synchronization (geodesics on a torus), and even biomechanical health metrics like scapulohumeral rhythm.

The deeper pattern is this: the cerebellum first identifies the right manifold — a sphere, a circle, a torus — and then hard-wires a simple navigation act on that manifold. The answer isn't computed; it's read off from a position in space.

I believe a formal distillation technique can be built from these primitives. A way to go from "neural net with hidden layers" to "point on a manifold with an act." From System 2 to System 1. From cortex to cerebellum.

Stay tuned. The rabbit hole goes further than you'd think.

---

**Author's Note.** This article relies heavily on metaphor and abstraction to convey its core ideas. The analogies are intended as intuitive lenses, not literal one-to-one mappings, and should not be quoted out of context. I am fully aware that both brain physiology and large language models are far more nuanced and complex than the simplified framing presented here.