Think that supercomputer churning out poetry is genius? Wait until you meet a one-year-old.
Babies can’t write code. They certainly aren’t debating metaphysics. But here’s the rub: today’s AI devours training data like it’s going out of style, guzzling electricity comparable to a small nation’s output. Meanwhile, a human infant figures out reality with terrifying efficiency. One or two looks at an object. Boom. Identification secured. Learning happens through fleeting observation. Physical touch. It’s efficient. It’s biological. It beats our silicon brains hands down.
Researchers are starting to admit something uncomfortable. The secret to better, cheaper AI might lie in baby brains.
To explore this bold new frontier, we needed a test that actually pushed models to think like toddlers.
The EgoBabyVLM Reality Check
Meta, Stanford, the University of Tokyo, and École Normale Supéérieure in France got together. They built a challenge. EgoBabyVLM.
It forces vision language models—VLMs that digest both text and images—to make sense of the world the way a baby does. The dataset? About a thousand hours of video recorded from cameras strapped directly to infants’ heads. Yes, really.
And the results? Disappointing for the AI industry.
Current top-tier models fail. Miserably. They choke on the messy, chaotic, first-person footage of actual human experience. The failure suggests a hard truth. Something about the human brain’s architecture lets it learn rapidly from tiny scraps of info. Our algorithms can’t touch that.
Messy Data, Rich Learning
Curated datasets are a crutch. Babies don’t use crutches.
They learn from a kaleidoscope. Parents pointing at invisible objects. Gestures. Talking about things that haven’t happened yet. Past memories. It’s not just words. It’s tactile. It’s multimodal. Michael Frank, a Stanford cognitive scientist involved in the project, put it simply: language alone isn’t the answer.
“It’s clear that there’s more that’s needed,” he said.
What Syntax Got Right (And Wrong)
EgoBabyVLM isn’t alone. Scientists love using AI as a mirror for human intelligence. Remember BabyLM from 2023? That challenge forced AI to learn syntax using data amounts comparable to what a ten-year-old processes. Tens of millions of words. Not trillions.
Transformers—the backbone of modern AI—actually pulled it off. They showed that pure attention mechanisms could grasp word relationships without the massive data hogs we’re used to. This slapped the face of Noam Chomsky. Maybe syntax isn’t hardwired. Maybe it’s learned.
But language is easy. The physical world? Not so much.
Ryan Cotterell from ETH Zurich pointed out a gap. “There isn’t going to be a large corpus,” he said. “No internet of human interactions.”
Joshua Tenenbaum at MIT sees the limits. Transformers find patterns. Sure. But they don’t acquire common sense. They don’t get social dynamics. They lack a theory of mind.
Transformers are very good at finding patterns, but they can’t turn a baby’s raw input into the complex understanding a child develops.
Tenenbaum asks the big question. Did evolution optimize us with built-in shortcuts? Or is the brain just complex machinery with inherent structure? He leans toward structure. A lot of it.
In 2024 someone showed a basic VLM could identify a ball just from one infant’s head-cam feed. It was impressive. It was also primitive. It wasn’t reasoning. It was pattern matching. Brendan Lake from Princeton notes the gap remains.
“The mystery is how children get the full capabilities by age 2.”
Toward a Causal Mind
The EgoBabyVLM authors think the path forward borrows from neuroscience. We need models that pay attention longer. We need them to interpret social cues. Not just pixel arrangements.
Frank already proved a point. Earlier this year his team tested a new model focused on causality. It tracked visual and temporal relationships using that same baby-video data. The model learned object dynamics. How things affect each other over time.
It worked.
It was much better than standard models at physical reasoning.
The possibility hangs there. Models biased to learn physics and social links quickly. They could be efficient learners. Real learners. Not just parrots of data.
Lake is excited.
“I’m excited to see what kinds new architectures researchers come up with,” he says.
New approaches are coming. Or they will. The question isn’t whether AI can learn like a baby. It’s whether we have the patience to teach it.
Who’s holding the camera? 🍼
