Loads of people are talking about artificial intelligence, which is basically getting computers or machines to do things that would normally require a person to do.

That's because there's been big improvements in how AI systems work since last year.

But artificial intelligence is not new on its own – it's been used for a long time on websites like Youtube and Netflix, to suggest what videos you might like to watch, it's also used by many websites to display advertising.

The basics of how AI works

A big part of artificial intelligence is about getting computers to recognise patterns.

Scientists show computers examples of things, and ask them to guess things about the examples. Then they correct the computers about their guesses until they start to make better guesses.

We’ll use the example of cars. Scientists showed computers millions of examples of pictures of cars, until the computer started to correctly guess when they saw pictures of cars.

The computers started to recognise that cars tend to have wheels and windows, and a bonnet at the front, and a booth at the back.

Because of that, when someone now asks an artificial intelligence computer system to draw pictures of cars, the system is able to.

We showed an AI system this picture that we drew, and it recognised as a sketch of a car.

Then we asked it to improve the image several times.

We typed in words like 'sketch’, then ‘colourful’ and ‘watercolour’ and ‘mosaic’ and eventually got this version back 15 minutes later.

What is remarkable is that the AI system recognised the image first of all as a car, but then also that it understood the words used in the requests to improve the image.

Systems like ChatGPT work in a kind of similar way, except they produce sentences and paragraphs rather than pictures.

Just like the image systems know that cars have wheels because they’ve seen millions of examples of cars, systems that produce text know how words and sentences tend to relate to each other because they’ve been shown a lot of sentences.

The way they originally learned the relationships within languages was amazing. Computers figured out a kind of maths for words.

For example, they learned the relationship between the word king and queen was this sum: KING – MAN + WOMAN = QUEEN

That’s really interesting, because it’s similar to how people learn about words too, whether we realise it or not. Take the sentence "the cat likes milk."

Everyone who reads it will understand the relationships between the words, and you’ll know you do by trying to rewrite another sentence that means the same thing.

Perhaps you might write "milk is something that the cat likes." That shows your brain knows about the links between the words.

Computers figured out that too, after seeing millions of examples. Then, more recently, scientists found a way to give the computers billions of examples.

Basically, to make systems like ChatGPT, they found a way to make computers read almost all the stuff on the internet, and figure out as much as they could about all the sentences there, and break down the links between the words.

Now, when someone types a question into it, what ChatGPT does is respond with what it has learned about words and relationships between them.

Just like the system that knew a car should have wheels and windows, systems like ChatGPT know about how a lot of words and sentences should go together on a lot of topics and about a lot of things.

And that can be amazing. However, the things ChatGPT writes might not be correct, often because stuff on the internet can be inaccurate or wrong.

Yet what scientists are surprised about, and why they think it’s so interesting, is that they didn’t expect computers to be quite this good at understanding human language mainly from being shown so many examples.

They thought they’d have to figure out more complicated ways to explain other stuff to computers before they got this good.

They’re amazed that these artificial intelligence computers are as clever as they seem to be. Some even think computers now are able to reason and be logical in a way they didn’t expect to see at all yet.

Even though it takes millions of examples for a computer to learn what a child can understand after maybe ten examples, the computers are still able to learn a lot, more than expected – to become artificially intelligent – and that’s why there’s a lot of excitement about AI these days.