Studying how babies learn could help develop more powerful AI models.
This story originally appeared in The Algorithm, a weekly AI newsletter. To receive articles like this in your inbox, Sign up here.
Human babies are fascinating creatures. Despite being completely dependent on their parents for a long time, they are able to accomplish amazing things. Babies have an innate understanding of the physics of our world and can quickly learn new concepts and language, even with limited information. Even the most powerful AI systems we currently have do not have such capabilities. For example, the language models that power systems like ChatGPT are good at predicting the next word in a sentence, but nothing comes close to the common sense of young children.
But what if AI could learn like a baby? AI models are trained on huge data sets consisting of billions of data points. Researchers at New York University want to see what such models can do when trained on a much smaller dataset: the sights and sounds experienced by a single child learning to speak. I did. Surprisingly, their AI learned a lot thanks to a curious baby called Sam.
Don’t be satisfied with only half the story.
Access the latest technology news without a paywall.
Subscribe now
Already a subscriber? Sign in
Researchers strapped a camera to Sam’s head and kept it on and off for a year and a half, from when he was six months old until shortly after his second birthday. With the material he collected, researchers were able to teach neural networks to match words to the objects they represent, Cassandra Willard reports in this article. (It’s worth the click just to see the incredibly cute photos!)

Wai Keen Vong
This research is an example of how babies can teach computers the same way humans do, and ultimately bring us closer to building AI systems as intelligent as humans. It’s just that. Babies have inspired researchers for years. They are keen observers and excellent learners. Babies also learn through trial and error, and humans become smarter as they learn about the world. Developmental psychologists say that babies intuitively understand what will happen next. For example, they know that a ball exists even when hidden from view, that it is solid and cannot suddenly change shape, and that it cannot roll down a continuous trajectory and suddenly teleport to another location.
Researchers at Google DeepMind bring the same “intuitive physics” to AI systems by training a model that learns how objects move by focusing on objects in a video rather than individual pixels. I tried to give it a sense of. They trained their model on hundreds of thousands of videos to learn how objects behave. The theory is that if a baby is suddenly startled by something like a ball flying out of a window, it’s because the object is moving in a way that goes against the baby’s understanding of physics. Researchers at Google DeepMind were also able to get an AI system to show “surprise” when an object moves differently than it has learned.
Yann LeCun, Turing Award winner and chief AI scientist at Meta, has argued that teaching AI systems to observe like children could be the path to more intelligent systems. Humans have a simulation of the world, or a “world model,” in their brains that allows them to intuitively know that the world is three-dimensional and that objects that disappear from view do not actually disappear. He says he can. This allows you to predict where a bouncing ball or speeding bicycle will be in a few seconds. He is passionate about building entirely new AI architectures inspired by the way humans learn. Here, we covered his big bet on the future of AI.
Today’s AI systems excel at narrow tasks, such as playing chess or producing text that sounds like it was written by a human. But compared to the most powerful machine we know, the human brain, these systems are fragile. They lack the common sense that would allow them to operate smoothly in a messy world, make more sophisticated reasoning, and be more helpful to humans. Studying how babies learn could help unlock those abilities.
deeper learning
This robot can tidy up a room without any help
Robots are good at specific tasks. For example, she’s good at picking up and moving things, and she’s also getting better at cooking. But while robots may be able to easily complete such tasks in the lab, getting them to operate in unfamiliar environments with little data available is extremely difficult.
Pick this up: Now, a new system called OK-Robot has the potential to train robots to pick up and move objects in environments they’ve never experienced before. This is an approach that has the potential to bridge the gap between rapidly improving AI models and the capabilities of real robots, as it does not require costly and complex additional training. Learn more about Rhiannon Williams here.
bits and bytes
Semaphore 🤝🏻 Microsoft
The news startup has signed a deal with Microsoft to use the tech giant’s AI chatbot to write articles. For breaking news events, Semafor uses chatbots to search for reports and commentary in different languages from other news sources around the world. This report will be added to a new feed called “Signals.” (Financial Times)
Multinational company loses $26 million to deepfake phone scam
Deepfakes are out of control, and it’s not just porn. Scammers were able to use a deepfake of a multinational company’s chief financial officer to trick employees into transferring company funds to five different bank accounts in Hong Kong, according to reports. South China Morning Post. As this technology improves and becomes more accessible, we can expect to see more stories like this. (Bloomberg)
EU countries give stamp of approval to AI law
With EU countries voting in favor, the AI law is inching closer to coming into force. But in typical EU fashion, there was a lot of last-minute panic and tensions as France tried to water down the bill. Euractic Contains a complete synopsis of the drama. The final step left is for the European Parliament to vote in favor later this spring. His Dragosh Tudorache, one of the leading negotiators of the AI Act, will speak at his EmTech Digital event in London on April 16th and he will be held on the 17th.
Inside the music industry’s high-stakes AI experiments
A fun story that examines the disruption occurring in the music industry and how veteran Lucien Grainge, Chairman and CEO of Universal Music Group, seeks to reap the benefits of technology without sacrificing copyright. (New Yorker)