AI Isn’t More Intelligent Than an Infant—At Least Not Yet

If you believe an AI model powered by thousands of state-of-the-art chips is intelligent, let me introduce you to the concept of a one-year-old child.

Sure, infants may not write code, tackle complex math, or engage in philosophical debates. However, in contrast to today’s AI models, which require vast amounts of training data and energy comparable to a small nation, babies efficiently learn to navigate their surroundings. They can recognize new objects after just a couple of glimpses and gather knowledge through brief observations and hands-on experiences.

In enhancing AI, the insights drawn from babies—and the structure of their brains—could be invaluable. Developing a more baby-like AI could reduce costs and energy consumption for advanced models, while also enabling AI-driven robots to learn about their environments more naturally.

To delve into this innovative territory, researchers from Meta, Stanford University, the University of Tokyo, and France’s École Normale Supérieure devised a novel test that showcases babies’ learning abilities and encourages AI developers to create corresponding algorithms.

The EgoBabyVLM Challenge assesses how effectively vision language models, or VLMs, which learn from both text and images, can interpret the world through the eyes of a child. It requires a model to describe its surroundings after processing approximately a thousand hours of video captured from cameras worn by infants and toddlers. (Yes, you read that correctly.)

It appears that the leading models struggle significantly when presented with such realistic and chaotic footage, hinting that there may be unique features in the baby brain’s design that allow for rapid learning from minimal information.

Unlike curated datasets, babies learn from a whirlwind of experiences: parents discussing objects that aren’t currently in view, pointing at things with their gaze or gestures, or recounting events from the past or those yet to come, rather than only focusing on the present. Babies acquire knowledge not just through language but also by engaging in rich multimodal and tactile experiences, says Michael Frank, a cognitive scientist at Stanford University who focuses on language acquisition and contributed to the EgoBabyVLM’s creation.

The challenge reveals that in the realm of AI, “it’s evident that there’s more [than just language] that’s essential,” Frank notes.

Language Acquisition

EgoBabyVLM is another example of how researchers are leveraging AI to investigate human intelligence. The BabyLM challenge, launched in 2023, tasked AI models with grasping language syntax using roughly the same volume of data as a 10-year-old absorbs—tens of millions of words, in contrast to trillions employed by AI models. Interestingly, transformer-based AI models, which analyze language by focusing on the relationships between words in various sentences, perform surprisingly well, undermining Noam Chomsky’s theories about the innate wiring of syntax in the human brain.

Ryan Cotterell, a linguist at ETH Zurich who initially developed BabyLM, observes that the scenario differs for comprehending the physical world. “There isn’t going to be a vast collection of human interactions—there’s no internet of human interactions,” he explains.

Joshua Tenenbaum, a cognitive scientist at the Massachusetts Institute of Technology, notes that BabyLM demonstrated that models do not develop “common sense” regarding the physical world, social interactions, or theory of mind.

“Transformers excel at identifying patterns in data,” Tenenbaum points out. “But it seems that purely pattern-learning systems can’t process the kind of information a baby or child encounters and learn to the extent they do.”