Despite the immense computing power behind today's generative AI models, a new test from researchers at Meta, Stanford University, the University of Tokyo, and France's École Normale Supérieure reveals that these systems still cannot match the learning efficiency of a one-year-old baby, according to WIRED.
The EgoBabyVLM Challenge
The EgoBabyVLM Challenge judges how well vision language models (VLMs), which learn from both text and imagery, make sense of the world as a baby sees it. The test requires a model to describe the world after ingesting about a thousand hours of video collected from cameras strapped to the heads of infants and toddlers. According to WIRED, "it turns out that the cutting-edge models fail miserably when fed this realistic and messy footage, which suggests there may be something different about the design of the baby brain that enables it to learn so rapidly from so little information."
Babies learn from a kaleidoscopic view: parents talking about objects no longer visible, indicating things using gaze or gestures, or discussing past or future events. Michael Frank, a cognitive scientist at Stanford who specializes in language learning and was involved with EgoBabyVLM's development, said: "it's clear that there's more [than just language] that's needed."
BabyLM and Language Learning
EgoBabyVLM follows a related challenge called BabyLM, introduced in 2023, which tasked AI models with learning the syntax of language using about the same amount of data a 10-year-old takes in—tens of millions of words, compared to trillions for AI models. Remarkably, transformer-based AI models can do this quite well, a finding that challenges Noam Chomsky's ideas about syntax being hardwired into the human brain. However, Ryan Cotterell, a linguist at ETH Zurich who first developed BabyLM, noted that understanding the physical world is different. "There isn't going to be a large corpus of human interactions—there's no internet of human interactions," he said.
Joshua Tenenbaum, a cognitive scientist at the Massachusetts Institute of Technology, observed that BabyLM showed models do not acquire "common sense" about the physical world, social dynamics, or theory of mind. "Transformers are very good at finding patterns in data," Tenenbaum said. "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."
Implications for Enterprise AI
For enterprise decision-makers investing in AI, the findings carry practical weight. According to WIRED, building a more baby-like version of AI "could make frontier models less costly and less energy intensive." It might also be valuable if AI-powered robots are to learn about their environments in a more natural way. This is particularly relevant for industries like logistics and manufacturing, where robots often operate in unpredictable settings. Current large models require massive datasets and energy; a baby-like learning algorithm could dramatically lower the bar for deployment.
Sources:
