Production note. This started as a long conversation with Gemini 3.1 Pro on August 10, 2026. I didn’t want to leave it sitting in a chat. An LLM summarized and drafted it; I edited it by hand.
My mother still keeps a little notebook where she writes down every step: “Open WeChat, tap the plus sign top right, choose Pay & Receive.” She’s had that notebook for almost ten years now, and every time the phone updates, she has to rewrite it. My niece, on the other hand, is three. Nobody taught her anything on the iPad — she found her favorite cartoon in two minutes and, while she was at it, went digging through my entire photo library.
Nothing new here. Everyone’s seen some version of this. But the part that actually stopped me was the question underneath it: why? Not the lazy answer — “kids these days are just sharper” — but something more specific. What are young people and older generations actually running, under the hood, that’s so different?
Young people aren’t learning. They’re pattern-matching.#
Strip it down, and what a teenager does with a new device isn’t really “learning” at all. It’s closer to pattern-matching.
Products today all look roughly the same underneath: a gear means settings, a magnifying glass means search, three lines mean menu, swiping left means delete. It’s a whole visual slang, and young people have been swimming in it since they could hold a phone. So when they open an app they’ve never seen, the thought isn’t “what is this, let me study it” — it’s closer to “oh, icon in that spot, that’s settings” — pulling from a grammar they already have and applying it to a new sentence. My mother doesn’t have that grammar. Every new screen is a document she has to translate word by word.
The deeper split is in the mental model itself. Young people default to thinking of software as a tree: get lost, hit back, or just mash the home button and you’re safe — nothing breaks. Switch phones, log into your account, and everything — photos, contacts — is already waiting in the cloud. It’s an abstract, reversible way of seeing the world. Older users are often still running the logic of a radio or an old television: press the wrong button and something might actually break; drop a cup and it doesn’t glue itself back together. So they go looking for a fixed sequence of steps, terrified of deviating — which is exactly why a single redesign wipes out years of hard-won familiarity, and hits them harder than it hits anyone younger.
There’s a more direct psychological layer sitting under all of this too. Young people operate on the assumption that a wrong tap costs nothing, so they tap freely. Older users assume a wrong tap costs something, so they freeze. One group learns by doing; the other learns by avoiding — and learning by doing is just faster, full stop, than waiting for someone to walk you through it. Marc Prensky put a name on this twenty years ago: digital natives and digital immigrants. For the former, digital is a mother tongue. For the latter, it’s a second language they’ll speak with an accent for life — printing out emails to read them, copying phone instructions into a notebook by hand.
So the “instant fluency” isn’t a gift. It’s inventory. Somewhere in a young person’s head sits a skeleton built from hundreds of prior apps and devices, and a new gadget is just a little new flesh grown onto bone that’s already there.
Machine learning has been chewing on the exact same problem#
I mentioned this to a friend who works in ML, and he stopped mid-sentence: that’s literally what we argue about all day.
There’s a wall in machine learning that nobody’s fully gotten past, called generalization: a model can look brilliant on its training data and still completely fall apart the moment it sees something new — the question is whether it can reason its way to something sensible (extrapolation), or just freezes because it only ever memorized (interpolation). Older AI systems trained to play chess could only play chess; change the map in a video game and they were useless. That’s not so different from my mother locking up the moment her phone updates — both cases are a system that memorized the surface, not the underlying rule.
A few of the more interesting research threads from the last few years read almost like an engineering translation of “young person vs. older generation.” Meta-learning tries to make a model learn how to learn — not memorizing the answer to one problem, but acquiring a general instinct for solving problems, so a new one only takes a handful of examples. That’s the same move as a teenager applying a general grammar to a brand-new app. GUI agents (Apple’s Ferret-UI, the “computer use” features showing up in various large models) are this idea made literal — instead of memorizing one app’s specific steps, the model learns abstractions like “what counts as a clickable button” or “what counts as going back.”
Push a layer deeper, and you get world models — the JEPA line of work Yann LeCun has been championing — aimed squarely at the reflex-like understanding older users default to. It’s the same failure mode as an older vision model that learned “snow = husky” purely because huskies in the training set kept showing up in snowy photos — a correlation mistaken for understanding. A world model tries to run a little simulation internally: if I tap this, what probably happens next? That’s mechanically the same thing as the abstract mental model young people carry around for software; it’s just phrased in engineering terms.
The trial-and-error piece maps onto curiosity-driven reinforcement learning — give an agent a reward for being surprised, and it starts poking at things the way a toddler does. Paired with that is work on safe exploration — an agent only dares to gamble freely once it knows it can undo the mistake. Which is really just the engineering version of “worst case, I restart.”
Line all of this up, and it points to something fairly plain: the way today’s best models are trained is starting to look a lot like raising a digital native — massive, task-free pretraining so the model can find the patterns on its own, then a safe environment where it’s actively encouraged to poke around.
But is “learns fast” just another word for “young”?#
Here’s where the harder question shows up: is all of this ultimately just a biological clock? Do young people learn fast simply because their brains are young — and does that window close, hard, at some age?
Neuroscience does have an answer, and it’s a fairly detailed one. A young brain runs in something like sponge mode — dense neuronal connections, new pathways forming easily whenever something novel shows up. Past a certain point, the brain goes through synaptic pruning — to save energy, rarely-used pathways get cut, while the well-worn ones get reinforced into highways. That’s why older adults are often startlingly efficient in the domains they know well — a trade, decades of social experience — while carving a new pathway for something genuinely unfamiliar costs far more than it does for someone younger.
The psychologist Raymond Cattell’s distinction between fluid and crystallized intelligence fits neatly here too. Fluid intelligence — the ability to reason through something new with no prior knowledge to lean on — tends to peak around 20 and drift downward from there. Crystallized intelligence, built from accumulated experience, keeps climbing for most of a lifetime. Facing something genuinely novel draws mostly on fluid intelligence — young people are near their peak, while older adults reach for the crystallized playbook, which often just doesn’t map onto the new thing in front of them.
Add in the difference between implicit learning (a child absorbs a native language without ever studying grammar) and explicit learning (once the prefrontal cortex has matured, new material needs to come as an explicit, attention-demanding rule), plus shrinking working memory, slower neural conduction leading to cognitive overload, and a dopamine system that shifts from an “explore” bias in youth to an “exploit and avoid risk” bias later in life — and you get a biological explanation that sounds remarkably tidy.
And that’s exactly where I want to push back a little.
It’s tidy enough that it’s easy to stop asking a question it should invite: how much of this gap is actually hardware aging, and how much is “never had to learn this,” “no one bothered to teach me,” or “too embarrassed to fail in front of someone”? I know plenty of people in their sixties who took up livestreaming, video editing, or drone flying after retirement and now outpace people half their age — their brains obviously didn’t get twenty years younger. I also know young people who, having had everything handled for them growing up, freeze up in front of anything unfamiliar. Chalking the whole gap up to synaptic pruning and declining fluid intelligence has a way of sounding scientific while quietly turning into fatalism — as if an older person’s struggle with technology were simply decided at birth. That understates how much motivation, self-efficacy, and sustained practice can actually move the needle. The original conversation does note there’s no strict critical period for technology the way there is for language, but the argument keeps leaning, page after page, on “age equals a closing window” anyway. I think that balance is off, and it’s worth saying plainly: biology sets a tendency, not a verdict.
If humans built their capacity to generalize during childhood, should AI get one too?#
Follow the thread far enough and you land here: human infancy looks completely useless on the surface — eat, sleep, cry, repeat — but is it secretly doing something enormous, laying down a general-purpose foundation for everything that comes after?
Some serious scientists think exactly that. I actually went and checked this one, because it’s the kind of claim that’s easy to fabricate a citation for: a 2024 paper in Trends in Cognitive Sciences, by Rhodri Cusack and colleagues (one of them at Google DeepMind), argues that the seemingly helpless early phase of human infancy is functionally training a “foundation model” — no specific task, just absorbing enormous amounts of multimodal data to underwrite everything cognition needs later. That paper is real, and the analogy is the authors’ own, not something invented along the way — which, once I confirmed it, made me trust this particular thread more than I expected to.
Follow that logic forward, and a few genuine research lines have been converging on it over the last couple of years. Drop a reinforcement-learning robot straight into “work mode” — reward it only for putting the cup on the table correctly — and it tends to memorize one exact trajectory; change the color of the table and it breaks. Give it a reward-free sandbox phase driven only by curiosity instead, and it tends to build a world model with more accurate physical intuition, showing noticeably stronger zero-shot generalization once it’s actually put to work. Some teams have gone further and trained vision models directly on footage from cameras worn by real infants — the SAYCam project, for instance — shaky, blurry, constantly interrupted by a baby’s hands grabbing at things. And oddly, models trained on that messy, noisy data end up with more human-like conceptual abstraction than ones trained on clean, curated images. Yoshua Bengio’s curriculum learning runs on the same underlying logic: don’t hand a model the hardest problem first — start simple, then escalate, the same way nobody hands a three-year-old calculus.
I mostly buy this thread — but I want to complicate it a bit too.
Pretraining a language model and an infant exploring the world sound like the same move, but underneath, they’re not. An infant is bootstrapping causal structure and physical intuition from raw, unlabeled sensory noise — nobody’s handed it clean text. A large language model, by contrast, reads text that humans have already written, already compressed, already structured with logic baked in. In a sense it’s inheriting a finished cognitive estate rather than building one from scratch. The real analogue to “infant-style exploration” is embodied agents poking around a physical or simulated environment, driven by curiosity, bumping into things and learning from the surprise — not a language model predicting the next token off a page someone else wrote. Collapsing those two into one story is, I think, the single sloppiest move in this whole discussion — not because the analogy is worthless, but because it’s aimed at the wrong target. And “Developmental AI” gets described here like an established field the industry has already agreed on, when as far as I can tell it’s really a handful of independent lines — curiosity-driven RL, curriculum learning, infant-perspective data — that happen to share a philosophy and have started citing each other. “Several threads converging” is a more honest way to put it than “a mature paradigm.”
Where this leaves me#
Run the whole analogy end to end and it’s genuinely compelling: young people’s instant fluency with new technology comes from internalized grammar and an abstract mental model, backed by real neuroplasticity, fluid intelligence, and a dopamine system tuned toward novelty. Machine learning, trying to escape rote memorization, is deliberately borrowing the same logic — treating task-free pretraining as a kind of childhood, swapping curiosity in for memorization, trying to move models from recalling answers to something closer to actually understanding them.
There’s a nice pattern in evolutionary biology: the smarter a species gets, the longer its childhood tends to stretch — that stretch of time spent protected, unproductive, and mostly just playing around. It’s true of us. It might turn out to be just as true of whatever comes after today’s models.
But reading through the whole arc of this, I kept reminding myself of one thing: a good analogy illuminates a problem, and it’s also very good at quietly doing your causal reasoning for you without asking permission. Young people really do learn faster than older people in some domains, and there’s real biology behind part of that. But a large chunk of the gap is something more human and, importantly, more changeable — motivation, embarrassment, whether anyone was willing to teach you in the first place. That second part is the one actually worth our attention. You can’t do anything about synaptic pruning. Whether someone’s willing to teach, and whether someone’s willing to try — that part is still very much up for grabs.