What is AI, really?

These words get thrown around like they mean the same thing. They don't. They fit inside each other, like Russian dolls.¹

×

The biggest doll is artificial intelligence: any machine doing something that looks smart. Old chess computers count, even though people wrote their rules. IBM's Deep Blue, which beat the world chess champion in 1997, searched 200 million positions a second and judged them with chess knowledge its team built in, with advice from grandmasters.²

Inside that is machine learning. Instead of writing the rules, you show the computer thousands of examples and let it work out the rules itself. In 2012, one learned to recognise a thousand kinds of object from more than a million labelled photographs, without anyone describing what each object looks like.³

Inside that are neural networks, the best-known way to do that learning. You use them every day without knowing it. Apple's Face ID uses them to tell a real face from a mask. Google has used them since at least 2013 to let you search your photos for things like flowers or cars. Spotify's researchers have built them to predict which songs you'll want to hear next. We'll open one up in a minute.¹⁴⁵⁶

And one particular kind of neural network is the large language model, or LLM. The name tells you what it is. Large, because it has billions of dials: GPT-3, in 2020, had 175 billion. Language, because it learned from text. Model, because it's a mathematical model of language: a huge set of numbers capturing how words tend to follow one another. ChatGPT, Claude and Gemini are all built on large language models.^7891011

Wrapped around all of it is ordinary software: the app that takes your words in and shows you the answer. The app can add features of its own, such as remembering things you've told it before.¹²

So is a neural network a brain? Not really. Think of a bird and an aeroplane. Planes were inspired by birds, but they don't flap. Neural networks borrowed one big idea from the brain and left the rest behind.¹³

×

Here's the idea. Your brain contains cells called neurons, which pass tiny electrical and chemical signals to each other.¹⁴ In 1949, the psychologist Donald Hebb proposed that when one cell repeatedly helps to fire another, the link between them grows stronger.¹³

Picture a toddler. They see a dog, and at that same moment a parent says "dog". The cells for the sight and the cells for the word fire together. Again and again. Each time, the link between them grows a little thicker. Until one day, just seeing a dog makes the word pop into their head.

×

That's learning. Not storing a file, but strengthening the right connections. The neuroscientist Carla Shatz later put it in a phrase that stuck: neurons that fire together wire together.¹⁵¹³

Neural networks copy that idea, with maths instead of chemistry: what they learn lives in the strength of their connections. Hebb's rule is the most basic way to adjust that strength, though today's AI uses a more powerful method, which comes next.¹³¹⁶

×

Each artificial neuron is incredibly simple. It takes in signals from the neurons before it, weighs them, adds them up, and passes a result on.¹⁷

The connections between neurons are where the magic lives. Each one has a strength, like a dial; engineers call it a weight. Turn it up and that signal matters more. Turn it down and it barely counts.¹⁷

A modern AI has billions of these dials. And here's the strange part: no single dial, or single neuron, means "dog" or "France". When researchers watch the neurons at work, they see a long list of numbers with no obvious meaning. Each idea is spread across many neurons, and each neuron takes part in many ideas. Researchers are only beginning to decode them.⁹

The text it learned from isn't kept inside as a copy. Every sentence it read just nudged the dials a tiny bit, and those nudges added up. Like eggs in a baked cake: they've changed every bite, but you can't point to where they are. That's why even the people who build these systems can't fully read what's inside.⁹

×

There is one caveat. A model can absorb some passages so thoroughly that it can repeat them word for word. In 2020, researchers pulled hundreds of verbatim passages from the training text of GPT-2, an earlier language model, just by prompting it to write large amounts of text and sifting through the results.¹⁸

This is why AI isn't like Google Search. Google keeps an index of pages that already exist and shows you the ones that match.¹⁹ The model itself has no pages to find. It writes a brand new answer, based on what its dials learned. Many AI apps can now run a web search and pass the results to the model, but the answer is still written, not copied. Usually it's right. But because it's writing rather than copying, it can sound completely sure of itself and still be wrong.⁸

When you first build one, all those dials are set at random. It knows nothing. So how does it learn? The same way the toddler does: by guessing, and being corrected.¹⁷¹⁶

The toddler points at a cat and says "dog!" The parent says "no, that's a cat." Next time, they're a little less likely to get it wrong.

For an AI, the parent is the text itself. Take a real sentence: "The cat sat on the…" Cover up the next word and ask the AI to guess. It doesn't give one answer. It spreads its bets across the words it knows, like a quiz contestant splitting their chips.⁷²⁰

Then the cover lifts. The real word was "mat", because that's what the writer wrote. Every dial gets nudged, just slightly, so next time a few more chips land on the right word.¹⁶

×

Repeat that billions of times, across a huge slice of everything people have written: GPT-3 practised on about 300 billion word-pieces of text.⁷ Finally, real people rate its answers, and it's nudged towards the ones they prefer. In one study, people preferred the answers of a model trained this way to those of the original GPT-3, more than a hundred times its size.²¹

So, back to our question. What actually happens when you hit send?

First, your words are cut into pieces, like Scrabble tiles. A piece can be a whole word, part of a word or a punctuation mark. In English, a piece is about three-quarters of a word on average.²²

Each piece gets a spot on a map of meaning. Think of a supermarket: apples sit near bananas, shampoo near soap. On this map, France sits near Spain and Germany, and Paris near London and Rome. Nobody arranged it. Training did: in 2013, researchers found their model had arranged countries and their capital cities in a consistent pattern, without ever being told what a capital is.²³

×

Next, the pieces look at each other. It's like a party where every word scans the room for who's relevant. "Capital" is asking: capital of what? "France" is wearing a name badge that says: I'm a place. Match. Researchers call this attention, and it is the heart of the transformer design that today's language models are built on.²⁴²⁰

That happens again, layer after layer, dozens of times: GPT-3 has 96 layers. Like an artist going over a sketch: first the outline, then the shapes, then the detail.⁷²⁰

Then it bets on the next word. By the time it reaches "The capital of France is…", Paris gets almost all the chips. It adds that piece and goes round again. One piece at a time, until it picks an invisible piece that simply means: stop, or it reaches a length limit. It usually picks a likely piece rather than always the top one, which is why the same question can get different answers.²⁵

×

And here's the twist. While answering you, it learned nothing that lasts. Not one dial moved. It can take account of what you said earlier in the conversation, but that lives in the conversation, not in the dials. Your question just flowed through them, like a marble dropped into a board of pegs that training put in place. The pegs guide it to the answer, but they never move.⁷

×

Answering is performance, not practice. Like a musician playing a song they already know: the practice happened months ago.

That doesn't mean your conversation goes nowhere. An app can keep notes about you and add them to later questions, and some companies use conversations to train their next model unless you opt out. Both happen outside the dials that answered you.¹²

Let's put it all together.

AI is a family of ideas, nested like Russian dolls, and the one you chat with is a large language model. That model is a neural network: billions of dials. Those dials were set by practice, guessing the next word and being corrected, like a toddler learning that a cat isn't a dog. The text it practised on isn't kept as a copy; it's baked in, like eggs in a cake. So when you ask a question, the model doesn't look the answer up like Google. Your question flows through the dials, like a marble through pegs, and out comes an answer, one word at a time.

So, what is the AI you chat with, really? It's software running a neural network: billions of dials, tuned by practice, that performs by predicting one word at a time.

Is that thinking? That's still hotly debated. A deeper question is whether such a system could ever be conscious. In 2023, nineteen researchers assessed AI systems against the leading scientific theories of consciousness. Their analysis suggested that no current AI system is conscious, but they found no obvious technical barrier to building one that would meet their criteria.²⁶