The Brain Prunes. AI Hoards.

At around age two, you hit your all-time synaptic peak. Your brain has packed in roughly 100 trillion connections — more than it will ever have again, more than you'll ever "need," more than any engineer would knowingly spec into a system. Then, for the next two decades, your brain destroys them.

Not by accident. Not due to neglect. Deliberately.

This is synaptic pruning, and it's one of the most counterintuitive things the developing brain does. The naive assumption is that more connections = more intelligence. More wiring = more capacity. It's the intuition baked into every "use it or lose it" warning your parents gave you. But that's not quite how it works.

The brain doesn't keep everything. It keeps the right things. And figuring out which connections to axe — and when — turns out to be one of biology's most elegant computational tricks.

Pruning Season Isn't Over at Two

Here's what people often miss: synaptic pruning isn't a one-time childhood event. It's a prolonged, hierarchical process that runs well into your mid-twenties.

A 2025 paper in Neuropsychopharmacology makes this case carefully. Critical periods — those windows of heightened neural plasticity where experience carves itself into the brain's architecture — don't open and close all at once. They unfold in a specific order (Neuropsychopharmacology, 2025a): primary sensorimotor cortices first, higher-order association cortices later. The visual and auditory cortex fine-tune themselves relatively early. The prefrontal cortex — home to planning, impulse control, and abstract reasoning — stays plastic well into young adulthood.

So when you're wondering why a 15-year-old can learn a new language with accent-free fluency that a 35-year-old will struggle to match, but also makes bafflingly impulsive decisions — that's the hierarchy in action. Different systems, different timelines, different pruning schedules.

It's Not Random. It's Experience-Shaped.

The brilliant, maddening thing about pruning is that it isn't a preset countdown clock. It's responsive. Connections that fire together survive together; connections that don't get used get trimmed.

But it goes deeper than that. A 2025 review in Neuropsychopharmacology on epigenetic regulation reveals that the molecular machinery governing plasticity is itself shaped by experience (Neuropsychopharmacology, 2025b). Through DNA methylation, histone modification, and chromatin remodeling, the brain doesn't just encode what you learned — it encodes how efficiently it should continue learning. Early-life adversity, for instance, can permanently alter the expression of plasticity-regulating genes, shifting the very timing windows through which pruning occurs.

Think about that for a second. The brain doesn't just have a learning algorithm. It has a meta-learning algorithm that adjusts the learning algorithm based on what kind of world it grew up in.

This is, to put it mildly, not how gradient descent works.

AI Does the Opposite

Modern neural networks are, in a word, hoarders. A typical large language model has hundreds of billions of parameters. Training is additive — weights get nudged, but rarely deleted. The network accumulates knowledge by adjusting the strength of connections, not by eliminating them.

There is a field called model compression and network pruning — the idea that you can train a large model and then strip out the weights that don't matter. The "lottery ticket hypothesis" suggests that inside every overparameterized network is a small, sparse subnetwork that could have done the job all along. It's a compelling idea. It's also, notably, something you have to engineer manually, after training, as a deliberate optimization step — not something that emerges from the learning process itself.

The brain prunes continuously, adaptively, in response to use patterns, regulated by epigenetic mechanisms that track developmental history. AI squishes weights toward zero if you explicitly tell it to.

These are not the same thing.

A Clue Hiding in LLM Internals

Here's where it gets interesting, and where I'll admit the story gets more speculative.

A 2025 paper in npj Artificial Intelligence by Wu et al. set out to understand where Theory of Mind capabilities actually live in a large language model's weights. ToM — the ability to reason about what other minds believe, want, or know — shows up in children around age four, if you're keeping score.

What Wu et al. found was startling: ToM-like behavior in LLMs is encoded in a remarkably sparse fraction of the model's parameters — just 0.001% (Wu et al., 2025). Perturb those specific weights, and ToM performance collapses. Leave everything else alone and it hums right along. These sensitive parameters, it turns out, are concentrated in the positional encoding module — a tiny functional neighborhood inside an otherwise vast network.

This isn't pruning. The model still carries all its weights; nothing got deleted. But it's functionally adjacent to the insight behind pruning: most of the model doesn't actually do the ToM work. The capability is concentrated, localized, and surprisingly sparse.

The brain would recognize this logic. Synaptic pruning produces a similar outcome — not by trimming a dense network to find the sparse core, but by starting overconnected and sculpting toward specificity through use.

The LLM backed into sparsity through sheer scale. The brain sculpted toward it through development. Same destination, very different roads.

What to Make of All This

I don't want to overstate the analogy. Synaptic pruning is a specific biological mechanism involving glial cells, immune signaling, activity-dependent competition between synapses, and developmental timing windows governed by epigenetics. Network pruning is a post-training optimization technique run by engineers with compute budgets. These are not the same thing, and collapsing the distinction helps no one.

But the principle is worth sitting with. The brain treats forgetting as a feature. Eliminating unused connections isn't cognitive loss — it's cognitive refinement. What survives the pruning is faster, more efficient, better tuned to the actual statistical structure of the environment the system developed in.

AI, by default, treats capacity as free. Keep everything. Let the weights sort it out. The result is models that are remarkably powerful and also strangely brittle — systems that can write you a sonnet in iambic pentameter but whose most socially sophisticated capabilities, it turns out, fit in 0.001% of their parameters.

Maybe the question isn't how to build AI that knows more. Maybe it's how to build AI that knows how to forget.

The developing brain figured this out somewhere between toddlerhood and your 25th birthday. It's still a little ahead of the field.