4.2 Intelligences That Learn

What do a dog fetching a stick and a Michelin-starred chef have in common?

Both improve through experience: they try, fail, adjust, and gradually refine their behaviour. For us animals, this way of learning is instinctive. If a behaviour brings a benefit, we repeat it. If it fails, we abandon it.

And machines?

Machines can also learn from experience through machine learning — a family of methods that allow a system to improve autonomously. One of these approaches closely resembles the way animals learn: reinforcement learning.

Here, the machine tries strategies, makes mistakes, adjusts its parameters, and tries again — until it finds one that works.

There are similarities — but humans and machines do not learn in the same way.

We understand context. We see a room, a ledge, an obstacle, and immediately know what to avoid. A machine does not start with meaning.

When an artificial intelligence learns to play Snake, it first makes absurd errors: if the apple is right in front of the snake’s head, it might turn right. It does not know — and never will — what “right,” “left,” “avoid,” or “meal” mean. It simply explores countless possibilities and selects the behaviour that maximises the score.

It needs far more attempts than we do — but it can afford them thanks to its enormous processing power.