A 3-Year-Old Beats GPT-4. Here's Why.

Here's a task I want you to try.

You see a red circle on the left. On the right, it's now a blue circle. Same shape, same size — just a color change. Now you see a red triangle on the left. What goes on the right?

A three-year-old gets this instantly. They've barely mastered shoelaces, but they look at the pattern, extract the rule (color changes, everything else stays put), and apply it to a new object without missing a beat.

GPT-4? Stumbles.

This isn't a trick question or a weird edge case. Yiu et al. (2025) ran exactly this kind of experiment as part of the KiVA benchmark — a suite of visual analogy tasks covering color, size, rotation, reflection, and number transformations, designed specifically to compare preschoolers (ages 3–5) against state-of-the-art multimodal AI models including GPT-o1, GPT-4V, and LLaVA-1.5. The results were, to put it charitably, humbling for the AI side. Children outperformed all tested models at identifying the transformation rule and then extrapolating it to new objects — especially for number, rotation, and reflection.

The models could often tell what changed. They couldn't reliably figure out how to apply it.

That distinction is not a minor footnote. It's the whole game.

Getting Better Isn't the Same as Becoming Competent

There's a tempting story in AI right now: models get better at tasks the more you train them, and with enough fine-tuning on domain-specific data, they develop something like expertise. A general model becomes a legal assistant, a medical summarizer, a code reviewer. That progression sounds intuitive. It borrows the language of skill-building without really committing to it.

Children don't become experts the same way. They don't just get more accurate — they pass through a specific sequence of developmental milestones, each one scaffolding the next. A kid learning number concepts doesn't jump straight from "some" and "many" to multiplication. They go through cardinality, then one-to-one correspondence, then the count sequence. Skip those rungs and the ladder doesn't hold.

This is the finding that jumped out at me from DevBench, a multimodal developmental benchmark for language learning published at NeurIPS 2024 (Tan et al., 2024). The researchers compared vision-language models against children and adults across seven tasks — lexical, syntactic, and semantic. The key insight: models that perform better on a task do, in fact, produce outputs that more closely resemble human behavioral patterns. So the correlation is real. The problem is that they systematically fail to replicate the ordering of human developmental milestones. A model might crush a task that children master late while fumbling on something children figure out at age two.

That's not a child becoming an expert. That's a very different thing wearing the same results on a leaderboard.

Fine-Tuning Is Optimization, Not Development

Here's where I want to push back on a lazy analogy that floats around AI discourse.

People sometimes describe fine-tuning — taking a pre-trained model and training it further on domain-specific data — as if it's analogous to the way a novice becomes an expert. You've got the broad base of knowledge, and now you're specializing. Like a medical student becoming a cardiologist.

The analogy is tempting. It's also mostly wrong.

When children build expertise, they're not just adding data to a fixed architecture. They're building internal representations — chunking sequences of moves into single cognitive units, developing error detection, reorganizing what they already know to accommodate what they're learning. A chess novice sees individual pieces. An expert sees patterns, threats, and gambits as single perceptual objects. The architecture is changing, not just the weights.

Fine-tuning, in the technical sense, is gradient descent applied to a pre-trained model on a new dataset. The model gets better at the target task. But it doesn't necessarily learn why it's better, or build the hierarchical representations that make expertise generalizable. This is part of why fine-tuned models can be brittle in ways that genuine human expertise isn't — they've been optimized, but not reorganized.

Portelance and Jasbi (2024) make a related point in their review of what neural networks can actually tell us about language acquisition. They're careful about the analogy in both directions. Neural networks can help generate hypotheses about what children are doing — they're useful scientific tools. But their failures are equally informative: what models can't do with language input that children master easily points toward something specifically human in the learning process. The analogy illuminates exactly at its breaking points.

Child-Sized Data Doesn't Fix It

You might wonder: what if we just constrained AI to learn the way children do — same amount of input, same kind of exposure?

Researchers have been running that experiment. The BabyLM Challenge asks teams to train language models on data budgets matching a child's real linguistic exposure — either 10 million or 100 million words — instead of the trillions of tokens frontier models consume (Hu et al., 2024). The 2024 edition drew 162 model submissions from around the world.

The results were interesting and slightly deflating in equal measure. Hybrid architectures worked better than pure approaches. Curriculum learning helped on some benchmarks. But here's the finding I keep coming back to: child-directed speech alone — the actual data children hear — underperformed richer literary datasets, even at small scales.

Which is a bit embarrassing, if you think about it. Babies learn from child-directed speech just fine. Models trained on the same input do worse than models trained on Jane Austen and Wikipedia.

Something is doing a lot of heavy lifting in the developing brain that the model doesn't have. And crucially, it's not just a data problem. It's a representation problem, an architecture problem, a rich-physical-world-experience problem.

The Boring Accurate Middle

I don't want to overstate the gap. Models do learn. Fine-tuning does confer real capabilities. The KiVA benchmark is a targeted adversarial test, not a general intelligence measure.

But I think the novice-to-expert framing reveals something genuinely useful: the process matters, not just the outcome. Children don't just end up competent — they get there along a particular route, and that route shapes what they know and how they know it. Generalization, analogy, rule extraction — these emerge from the architecture of how development unfolds, milestone by milestone.

Fine-tuning is a remarkable engineering tool. It is not that.

When a three-year-old figures out that the blue circle was once red and then applies that logic to a blue triangle without anyone telling them to — that's the output of years of embodied, sequential, error-driven learning that we don't fully understand yet. It happens to look effortless. It isn't.

The gap between that and running gradient descent on a new corpus is still, in 2025, very wide. And I'd rather we said so clearly than kept reaching for the expertise metaphor because it sounds good in a pitch deck.