Even the Most Advanced Neural Networks Are Not Yet Smarter Than a One-Year-Old — and That Is No Exaggeration

Despite the impressive progress of artificial intelligence, modern neural networks still lag far behind young children in one of the most fundamental aspects of intelligence — the ability to quickly understand how the world around them works and learn from minimal experience. This is the conclusion increasingly reached by experts in cognitive science, neurobiology and machine learning.

The paradox is that language models can write computer code, pass difficult exams, engage in meaningful conversations and analyse vast amounts of information. However, when AI is placed in a new, unfamiliar environment that requires an understanding of physical laws, cause-and-effect relationships and interaction with objects, its capabilities decline sharply. Where a one-year-old begins to figure things out independently in literally a few days or weeks, artificial intelligence requires millions of training examples and enormous computing power.

Why children learn more effectively than AI

The main difference lies in the way they learn.

An infant receives information through multiple channels at once: sight, hearing, touch, movement, balance, social interaction and emotional feedback. Every action becomes an experiment. If a child drops a toy, they observe how it falls. If they push an object, they understand that it moves. If they smile at an adult, they receive a response. In this way, the brain continuously builds an internal model of the world around it.

What is more, a child needs very little data to form such representations.

Modern neural networks work in a completely different way. They are trained primarily on vast amounts of ready-made information — texts, images, videos or audio recordings. To learn a pattern, models often need millions of examples, whereas a child may sometimes need only a few observations.

That is why experts say that modern AI is excellent at recognising statistical patterns but is far worse at understanding the very nature of what is happening.

The huge difference in efficiency

The gap becomes especially clear when comparing the amount of data used.

During the first year of life, a child sees only a small part of the world around them, yet this is enough to form complex ideas about objects, people, space and cause-and-effect relationships.

Large language models, by contrast, are trained on virtually the entire available internet, libraries, scientific publications and other digital sources. Thousands of powerful graphics processors consuming enormous amounts of electricity are used to create them.

Despite this, AI still makes mistakes that a child would typically stop making after their first months of becoming familiar with the world around them.

What is the main problem with modern models

One of the key reasons is considered to be the lack of genuine interaction with the physical world.

Most existing models effectively «sit inside a computer». They analyse data, but cannot themselves explore their surroundings, experiment, make mistakes and gain new experience in the way humans do.

At the same time, active interaction with the world is considered the foundation of human intelligence.

That is why an increasing number of research groups are working in a field that combines robotics, neurobiology and artificial intelligence. Their goal is to create systems that can learn not only from ready-made data but also through their own experience.

What scientists propose

Today, researchers are increasingly turning to the principles of how the human brain works.

Among the most promising areas are:

learning through interaction with the environment, rather than only from pre-prepared data;

developing models capable of independently exploring new situations;

combining vision, hearing, tactile sensations and movement into a unified perceptual system;

creating mechanisms for long-term memory and the accumulation of experience;

developing what is known as a "world model" — an internal model of how reality is structured.

These approaches could make future AI systems significantly more flexible and more economical to train.

The next stage in the development of artificial intelligence

Experts increasingly agree that further progress in AI will depend not so much on increasing the volume of data or the power of graphics cards as on a deeper understanding of how the human brain learns.

If engineers manage to reproduce even some of the mechanisms that allow infants to master the world around them incredibly quickly, this could become the next major breakthrough in the development of artificial intelligence.

That is why neurobiology and cognitive psychology are becoming no less important to the future of AI than programming and computer technology. Perhaps the path to creating truly general artificial intelligence lies not in increasingly powerful computers, but in understanding how several billion neurons in a child's brain transform limited life experience into comprehensive knowledge about the world.