AI Isn’t Smarter Than a Baby Yet – New Research Highlights Differences
The new EgoBabyVLM benchmark reveals that advanced vision-language models fail to learn from infant-like headcam video, underscoring that AI still lacks the efficient learning capabilities of a one-year-old.
While artificial intelligence models run on thousands of cutting-edge chips and consume vast amounts of data, a one-year-old human baby learns with remarkable efficiency: identifying new objects after one or two glimpses, using brief observations and physical interaction.
Researchers from Meta, Stanford University, the University of Tokyo, and France's École Normale Supérieure developed the EgoBabyVLM test to compare AI learning to that of infants. The test feeds vision-language models (VLMs) about a thousand hours of video recorded from cameras strapped to the heads of babies and toddlers. Cutting-edge models perform poorly on this realistic, messy footage, suggesting fundamental differences in how baby brains learn from minimal information.
"It's clear that there's more [than just language] that's needed," says Michael Frank, a cognitive scientist at Stanford specializing in language learning who was involved in developing EgoBabyVLM. Babies learn not only from language but also from a rich multimodal and tactile experience: parents talking about objects no longer visible, indicating things with gaze or gestures, and discussing past or future events rather than the present.
The earlier BabyLM challenge from 2023 tested whether AI models could learn language syntax using about the same amount of data a 10-year-old takes in — tens of millions of words versus trillions for AI models. Transformer-based models performed surprisingly well, challenging Noam Chomsky's ideas about innate syntax. However, the situation differs for understanding the physical world. "There isn't going to be a large corpus of human interactions — there's no internet of human interactions," says Ryan Cotterell, a linguist at ETH Zurich who developed BabyLM.
Joshua Tenenbaum, a cognitive scientist at MIT, notes that BabyLM showed models fail to acquire "common sense" about the physical world, social dynamics, or theory of mind. "Transformers are very good at finding patterns in data," says Tenenbaum. "But it does seem that just pure pattern learning systems are not able to take the kind of data that a baby or a child receives and learn all the things that they do."
In 2024, researchers showed a basic VLM could learn simple concepts like what a ball is by consuming data from a single infant's headcam. But this falls far short of sophisticated reasoning. "The mystery is how children get to the full capabilities that they have even at the age of 2," says Brendan Lake, a cognitive scientist at Princeton University involved in that project.
The EgoBabyVLM authors suggest borrowing ideas from cognitive science and neuroscience to develop more humanlike learning algorithms, including models that can pay attention over longer periods and interpret social cues. Frank and colleagues have already tested a new model adept at learning causality and visual-temporal relationships from the same infant video data, which learned object dynamics much more effectively — a foundation for physical reasoning.
"EgoBabyVLM is a wonderful challenge," says Lake. "I'm excited to see what kinds of new architectures, approaches, and ingredients researchers come up with."
