AI Still Fails Tests a 1-Year-Old Passes

A 1-year-old still outperforms today’s most advanced AI at one of the hardest problems in computing: learning efficiently from the real world.

That is the premise behind EgoBabyVLM Challenge, a new benchmark created by researchers at Meta, Stanford University, the University of Tokyo, and France’s École Normale Supérieure. The test asks vision-language models (VLMs) to interpret roughly 1,000 hours of video captured from cameras strapped to the heads of infants and toddlers. According to Wired, leading models perform badly on the messy, fragmented footage.

The result points to a basic gap between current AI and human learning. Babies can identify new objects after seeing them once or twice, and they learn through brief observation and physical interaction. By contrast, modern AI systems rely on vast curated datasets and enormous computing resources.

As Michael Frank, a cognitive scientist at Stanford University involved in the benchmark, puts it:

Recommended reading

Mira Murati’s lab unveils Inkling, a 975B open model

“it’s clear that there’s more [than just language] that’s needed,”

What EgoBabyVLM reveals about AI learning

Unlike internet-scale training data, an infant’s experience is chaotic. Parents refer to objects that are no longer visible, point with their gaze or gestures, and talk about past or future events rather than only what is directly in view. Frank says babies also learn through rich multimodal and tactile experience, not language alone.

That helps explain why progress in language has not translated cleanly into physical understanding. A separate benchmark, BabyLM, introduced in 2023, asked models to learn syntax from about the amount of language a 10-year-old encounters—tens of millions of words instead of the trillions used for large AI models. Ryan Cotterell of ETH Zurich, who developed BabyLM, found transformer-based systems can do surprisingly well on syntax.

But physical and social reasoning appear much harder. As Joshua Tenenbaum of MIT told Wired:

“Transformers are very good at finding patterns in data. 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.”

Baby-like architectures are now the research target

Researchers are increasingly asking whether human and animal brains benefit from built-in structures shaped by evolution, rather than relying on general-purpose pattern matching alone. Tenenbaum noted that the brain has extensive built-in architecture, and that remains a central debate in cognitive science and neuroscience.

There are early signs of progress. In 2024, researchers showed that a basic VLM could learn simple concepts such as what a ball is from recordings taken from the head of a single infant. But Brendan Lake of Princeton University said that remains far from the kind of sophisticated reasoning children show by age 2.

The EgoBabyVLM authors argue that ideas from cognitive science and neuroscience could help, including models that sustain attention over longer periods and interpret social cues. Frank and colleagues also reported earlier this year that a different model, designed to learn causality and visual and temporal relationships, performed much better on the same baby-head video data.

For researchers chasing more efficient systems, that is the real appeal: a model that learns more like a baby might need far less data and energy than today’s frontier AI.