Enable
22 November 2026
18 minute read
How AI grew up
The bedtime story behind your chatbot.
In short
AI writes one word at a time, predicting what comes next. It took more than 75 years to get good at it: learning from examples rather than rules, fed by internet data and gaming chips, organised by the 2017 Transformer and taught manners by people, which gave us ChatGPT in 2022. Since then it has learned to see, listen, think first, use tools and work in teams. It can look scary, but it is worth understanding: listen to it, and check it.
For transparency: the images in this post were generated with AI to help illustrate the technical ideas.
“One more story, a new one. About a dinosaur.”
Your six-year-old is dinosaur mad, and you have run out of stories. So you open a chat tool like ChatGPT, Claude or Gemini and type:
“Write a five-minute bedtime story about a baby T. rex called Rory who wants the biggest teeth in the forest. Make it gentle, and end with Rory all grown up, surrounded by his friends.”
A few seconds later, the story appears, word by word. Rory has a mum, a muddy pond, teeth that are still a bit wobbly and a happy ending. It is better than anything you could have made up at seven o’clock.

Seven o’clock, out of stories, and a chat tool to the rescue.
Here is the twist. The AI that just wrote a story about a little dinosaur growing up has a growing-up story of its own.
Dinosaurs look scary: all teeth and roar. AI can feel the same. Some care is sensible, and we will come back to that, but knowing how Rory grew up takes away a lot of the fear.

Big teeth can cast a scary shadow. Up close, Rory is still learning.
What happens when you press send
Five things happened in those few seconds.
1. Your message was cut into pieces. The AI split it into small chunks called tokens. A token is usually a word or part of a word. “Dinosaur” might be one token. “Bedtime” might be two.

Your message is cut into pieces called tokens.
2. It predicted the next piece, over and over. The AI looked at all your tokens together and asked: what is the most likely next word in a good reply? It picked one, then did the same again, one piece at a time. That is why the story appears word by word: it is being written as you watch. Hold on to this idea: even the cleverest AI today still works this way. Later we will see how something so simple can plan, write code and run whole projects.

Each time, it picks the likeliest next piece. The taller the bar, the more likely the card.
3. It used patterns it had already learned. During training, the AI was shown an enormous amount of text: books, websites, articles and more. From that, it learned how language works: that stories often start with “Once upon a time”, and that baby animals usually have a mum nearby. It is not copying a story it has read. It builds a new one from those patterns.

Patterns learned from an enormous amount of text.
4. It followed its training in good behaviour. After learning language, it had extra lessons in how to behave, a bit like a puppy at training classes. People rated its answers, and it was tuned to follow instructions and be helpful. That is why you got a gentle story with a happy ending, not a ramble.

Lessons in good behaviour, a bit like puppy training.
5. It only used what was in this chat. The AI only “remembers” what is in your current conversation. Tell it Rory is green and loves puddles, and it will use that. Start a new chat, and Rory is forgotten.

Tell it Rory is green, and it will remember, until the chat ends.
A note on memory: many chat tools now save useful details from past chats, such as your child’s name or that Rory is green, and quietly add them to the start of each new conversation. So the AI is not remembering the way you do. It is being handed a note before it starts. You can usually see, change or turn off these notes in the settings.
So how did a computer learn to do all that? Let’s go back to the very beginning, when Rory was still an egg.

Rory’s growing-up story, and AI’s: from wobbly teeth to the biggest teeth in the forest.
The egg: the 1950s
Nobody knows what will hatch, or how big its teeth will be.
In 1950, the British mathematician Alan Turing asked a simple question: “Can machines think?” He suggested a test. If you chatted with a machine and could not tell it apart from a person, perhaps it was thinking.

1950: “Can machines think?” Nobody knows what will hatch.
In 1956, a group of scientists met for a summer workshop at Dartmouth College in the USA. They gave the new field its name: artificial intelligence.
Soon after, the first machines that learned from examples appeared. In 1959, Arthur Samuel at IBM wrote a draughts program that got better by playing game after game. He called this “machine learning”.
Why it matters: Turing’s question is still the one people ask about chatbots, and Samuel’s idea of learning from examples is how Rory learned to write.
Two hatchlings and two hard winters: the 1960s to the 1980s
Rory hatches, and so does his twin brother, Rex. Rex does exactly what he is told, one rule at a time. Rory learns by watching and trying. Then two hard winters nearly finish them both.

Rory learns by watching and trying. His twin, Rex, follows the rulebook.
Rex, the rule-follower. One way to build AI was to write down the rules by hand: if this happens, do that. In 1966, a program called ELIZA chatted like a therapist using simple word tricks. It understood nothing, yet some people confided in it. In the 1980s, “expert systems” packed with thousands of rules helped companies with narrow jobs, such as spotting faults in machines.
Rex grew fast, then stopped. The world is too messy to write a rule for everything. Try writing the rules for a good bedtime story. You would never finish.
Rory, the learner. The other way was to learn from examples. Instead of rules, the computer was shown lots of cases and left to spot the pattern. Show it hundreds of pictures marked “cat” or “not cat”, and it slowly works out what a cat looks like, without anyone describing whiskers or ears. In 1986, a method called backpropagation gave it a better way to learn from its mistakes. Each time it guessed wrong, it nudged itself a little closer to the right answer.
But Rory grew slowly, for two reasons. First, computers had far less power than the phone in your pocket, and learning takes millions of small sums. Second, there was very little to learn from: no internet full of text and photos, just examples gathered and typed in by hand. It was like trying to learn a language from a single phrasebook.
The hard winters. In the mid-1970s and again in the late 1980s, promises ran ahead of results and the money dried up. People call these the “AI winters”. Like the real dinosaurs, AI nearly died out.

The AI winters: promises ran ahead of results, and the money dried up.
Why it matters: Rex was a rival that ran out of road. Rory survived, and it is Rory’s way of learning, from examples, that wrote the bedtime story.
Feeding time: the 2000s to 2013
At last there is plenty to eat, and somewhere warm to grow. Rory has a growth spurt. He starts to recognise shapes and faces, and to work out what words mean.
A growing dinosaur needs two things: food and warmth.
The food was data. The internet and smartphones meant billions of people were writing, posting and sharing every day. Suddenly there were mountains of text and pictures to learn from.
The warmth was computing power. Chips designed for video games, called GPUs, turned out to be very good at the sums that learning needs.

Food and warmth: mountains of data, and gaming chips to do the sums.
In 2012, a system called AlexNet used both to win an image recognition contest by a wide margin. Computers had learned to see. In 2013, a method called word2vec turned words into numbers that capture meaning.
To picture it, imagine a giant 3D grid, like a room full of invisible boxes. Every word gets its own spot in the room, set by three numbers: how far across, how far up and how far back. Words with similar meanings sit close together. “Puppy”, “dog” and “kitten” huddle in one corner. “Dinosaur”, “fossil” and “T. rex” gather in another. Even the directions mean something. The step from “man” to “woman” points the same way as the step from “king” to “queen”. Real AI uses hundreds of directions, not three, but the idea is the same: each word’s meaning becomes a place.

Every word gets its own place in the room: how far across, how far up and how far back.
Why it matters: this is how the AI knows that “dinosaur”, “mum” and “bedtime” belong in the same kind of story, and that “spreadsheet” does not.
A bigger brain: 2017
Rory’s brain has a growth spurt of its own. Now he can take in a whole story at once and keep track of who is who.
In 2017, eight researchers at Google published a paper with an odd title: “Attention Is All You Need”. It described a new design called the Transformer. The “T” in ChatGPT stands for it.
Older systems read text one word at a time, like a finger following a line, and often lost track of the start of a long sentence. They were also slow to train, because each step had to wait for the one before.
The Transformer takes in a whole passage at once. That does not mean it understands it all in one go. It runs lots of calculations side by side instead of one after another, which made it much faster to train on those gaming chips.
The key trick is called attention. For every word, the model works out which other words matter most to it. In “Rory asked his mum if his teeth would ever be big”, attention links “his” to Rory, and “big” to his teeth.

Attention links each word to the words that matter to it.
Why it matters: attention is how the AI kept track of Rory, his mum and his friends across your whole story, instead of losing the thread halfway through.
Learning to talk: 2018 to 2020
Rory starts to babble, then to talk. He is surprisingly fluent, but he does not listen. Ask him something and he just carries on talking.

Fluent, but not listening: a clever mimic, not yet a helper.
With a bigger brain, Rory could eat far more. Researchers found a simple recipe: take a Transformer, feed it a huge amount of text and train it to guess the next word. Then make it bigger, and feed it more.
In 2018, OpenAI released GPT-1 and Google released BERT. GPT stands for “generative pre-trained transformer”. Generative means it makes things. Pre-trained means it learns from a big pile of text first.
In February 2019, GPT-2 wrote paragraphs so fluent that OpenAI held back the full version at first, worried about fake news. In May 2020, GPT-3 arrived. It was more than 100 times bigger than GPT-2, and it could write articles, poems, stories and even simple computer code.
But it was hard to talk to. Ask it for a bedtime story and it might carry on your sentence or wander off topic. It was a clever mimic, not yet a helper.
Why it matters: fluent writing arrived before the chat window did. The skill that wrote Rory’s story was there by 2020. It just needed manners.
The first roar: 2022
Rory learns his manners. He listens, does what he is asked, and lets out his first real roar. It is loud, but it is friendly.
So who taught Rory his manners? People did. In early 2022, OpenAI showed that a model could learn to follow instructions if people wrote good example answers and ranked its replies. This is called reinforcement learning from human feedback.
Then, on 30 November 2022, OpenAI launched ChatGPT as a free “research preview”. It was the same kind of model, in a simple chat window, trained to be a helpful partner. Anyone could ask in plain English and get a sensible answer.
Where GPT-3 might have carried on your sentence or listed dinosaur names, ChatGPT simply writes the story. It starts with “Once upon a time”, gives Rory a mum and some friends, keeps it gentle and stops at a happy ending. Say “make Rory green and add a puddle”, and it rewrites the story to match. It still predicts the next word, but all that feedback from people has taught it that the best next words are the ones that do what you asked.

30 November 2022: the first roar, and the world noticed.
The world noticed. More than a million people tried ChatGPT within days. Teachers, coders, marketers and parents all started to experiment.
Why it matters: this is the roar you heard when Rory’s story appeared. A clever mimic became a helper you could simply ask.
Growing up fast: 2023 to 2026
Since that first roar, Rory has grown faster than at any time in his life. He has not grown a new brain. He is still predicting the next piece, one at a time. What has changed each year is what he can take in, how long he thinks, and what he is allowed to do.
2023: learning to see and draw
Rory learns to look closely at the world, name what he sees, and draw it in the mud with his claw.

2023: learning to see, and making pictures appear out of static.
The same prediction engine learned to handle pictures. A photo is cut into small squares, and each square is turned into numbers, just like a word. The model treats those squares as tokens alongside your words, so it can “read” a photo and describe what is in it. OpenAI’s GPT-4, released in March, was one of the first popular models to do this.
Making pictures works differently. Image tools such as Midjourney and DALL-E 3 start with a square of random static, like an untuned TV. Step by step, they remove the noise, guided by your words, until a picture appears. They learned how by studying millions of pictures paired with captions.
It was also the year of choice. Anthropic’s Claude and Google’s Bard (later Gemini) arrived as rivals to ChatGPT, and Meta shared its Llama models openly for anyone to build on.
Why it matters: you could show the AI your child’s crayon drawing of Rory and ask for a story to match, or ask it to draw Rory for the cover.
2024: hearing, remembering and thinking first
Rory finds his voice and learns to listen. He remembers more of what he is told, and starts to stop and think before he answers.

2024: listening, remembering more, and thinking before answering.
Listening and speaking. Older voice assistants turned your speech into text, answered in text, then read it out, which made them slow and flat. Newer models, such as OpenAI’s GPT-4o in May, handle sound directly as tokens. They reply in about the time a person would, and pick up on tone of voice.
A bigger short-term memory. The amount a model can hold in mind at once is called its context window. Early ChatGPT could hold a few thousand words. In 2024, Google’s Gemini could hold something as long as a whole book.
Thinking first. “Reasoning” models, such as OpenAI’s o1 in September, are trained to write out their working before they answer, like a pupil doing rough sums in the margin. In training they were rewarded when the working led to a right answer, so they learned to check, backtrack and try again. They are slower, but much better at maths, logic and planning.
Why it matters: you could ask for Rory’s story out loud and hear it read straight back, and the AI could work through a tricky question before answering.
2025: smarter training and doing jobs
Rory starts doing jobs for himself. He wanders off to find things out, and learns to use sticks and stones as tools.
Smarter, not just bigger. For years the recipe was simply more data and more chips. In January 2025, the Chinese company DeepSeek showed another way. Its R1 model learned to reason mostly by trial and reward, and it was trained for a fraction of the usual cost. It came close to the best American models and was shared openly. The news briefly rattled stock markets, and pushed the whole industry towards cheaper training.
From answers to actions. 2025 was the year of agents: AI that does a job for you, not just a reply. “Deep research” tools searched dozens of websites and wrote up a report with links. Coding tools such as Claude Code built and fixed software, testing their own work as they went. AI answers, built from the pages they found, became a normal part of Google search.

2025: Rory goes and checks, using tools to find things out.
Moving pictures. Video tools, such as Google’s Veo 3 in May, took the same remove-the-noise idea and applied it across many frames at once, with matching sound.
Why it matters: before telling your child that some dinosaurs had feathers, the AI could now go and check, and show you where it found out.
2026: finding friends
Rory finds friends just like him. Together they split up the big jobs, and each one does the part they are best at.

2026: a team of helpers, each doing the part they are best at.
Teams of helpers. In 2026, agents started working in teams. A lead agent breaks a big job into parts and hands each one to a helper. The helpers work at the same time, then the lead puts their results together and checks them. Anthropic brought in “agent teams” for Claude in February, and Microsoft rebuilt its Copilot assistant around agents in March.
Working for longer. The newest models, such as Anthropic’s Claude 5 family and OpenAI’s GPT-6, can keep going on a task for hours, not seconds, because they can plan, check and use tools along the way.
The right size for the job. Small, cheap models now handle quick everyday requests, and the biggest ones are saved for the hardest problems. That is one reason prices keep falling.
Why it matters: today, one request could get you Rory’s story written, illustrated, laid out as a picture book and saved to your files, with a team of helpers each doing a part.
From guessing words to doing jobs
Rory is all grown up, with the biggest teeth in the forest. But people keep asking him the same thing: are you really thinking, or just guessing?
Here is the puzzle. Underneath it all, Rory still does one thing: he predicts the next word. So how can that solve problems, write computer programs, give advice or run a whole project? The answer comes in three steps.
Step 1: predicting well is harder than it sounds. Finishing “The cat sat on the…” is easy. Now picture the last page of a detective novel: “The murderer was…”. To guess that word, you need to have followed the whole plot. The AI was trained on billions of pages like this, and it was nudged to improve after every wrong guess. To keep getting better, it had to pick up what lies underneath the words: facts, cause and effect, how a sum works, how people give advice, how a program fits together. So “just predicting the next word” ends up packing in a huge amount about how the world works.
Step 2: lots of small guesses make a chain of thought. One guess is one word. But every word the AI writes becomes part of what it reads next. So when a reasoning model writes out its working, each step gives it something to build on. It is like long division on paper: no single step is hard, but together they solve a problem that no single guess could. Computer code works the same way. Code is just another kind of text, so the AI writes a program one line at a time, just as it writes a story.

No single step is hard, but together they reach what one guess never could.
Step 3: add tools, and the guesses become actions. The AI’s next words do not have to be part of an answer. They can be an instruction, such as “search the web for feathered dinosaurs” or “run this code and show me what happens”. A tool carries out the instruction and hands back the result. The AI reads it and predicts the next step. Round and round it goes: plan, act, check, repeat. That loop is what turns a chatbot into an agent.

Plan, act, check, repeat: the loop that turns a chatbot into an agent.
Asking a question or setting a job?
Ask AI a question and you get one reply; you decide what happens next. Give an agent a job and it takes the steps for you. An agent usually takes longer, shows its progress as it goes (“searching 14 websites”, “running tests”), works with your files or apps, and often asks your permission before anything important, such as sending an email or making a booking. A plain chat just answers.

Ask a question and you get one reply. Set a job and the agent takes the steps.
So is it really thinking?
It depends what you mean by thinking. Nobody is in there having ideas, and it has no thoughts or feelings the way you do. But it can work through a problem step by step, notice its own mistakes and change course. From the outside, that looks a lot like thinking. A fair answer for now: it is not thinking like a person, but it is doing far more than guessing.

Is it really thinking? Not like a person, but far more than guessing.
Does it understand what I am saying?
Partly. Rory has never seen a real dinosaur, felt sleepy or tucked a child into bed, so he cannot understand those things the way you do. But he is not just matching surface patterns either. Remember the 3D grid? The AI has built a vast map of how words and ideas connect. That map lets it handle requests it has never seen before, like a story about a green dinosaur who is scared of puddles. Think of it as understanding the shape of an idea without ever having lived it.
If it is only predicting words, how can it be intelligent?
Turing’s old test asked whether you could tell a machine from a person just by chatting. Today’s AI can often fool people in a short chat. Experts still argue about whether that counts as real intelligence. What matters more is that it can be brilliant at some tasks and oddly silly at others, so it pays to know which is which.
Why did it suddenly get so good?
It was not really sudden. Four things came together in about a decade: mountains of data from the internet, cheap computing power from gaming chips, the Transformer design that could make use of both, and feedback from people that taught it manners. Researchers also found that bigger models trained on more data got better in a steady, predictable way, so companies kept making them bigger. Since 2024, a second boost has come from letting models think for longer before they answer.

Slow for decades, then a sudden spurt: data, computing power, the Transformer and feedback from people, all at once.
Why it matters: Rory is still predicting one word at a time. Give him room to think and tools to use, and those small guesses add up to real work.
Leaving the nest: what next for Rory?
Rory has his big teeth now, but he is still finding his feet, and still a bit clumsy. Soon he may leave the nest altogether.
Nobody knows for sure, and guesses about AI have often been wrong, but a few things look likely. AI helpers are starting to move off the screen and into the real world, inside robots, cars and household gadgets. They will probably remember you better, so Rory’s story already knows your child’s name and favourite colour. Teams of helpers may work for days on bigger jobs, checking in only when they need you to decide something.

Leaving the nest: nobody knows exactly which road he will take.
Remember the scary shadow? Some worries are real: AI can be confidently wrong, and private information needs care. A lot of effort is going into making Rory more honest about what he does not know, and he gets better every year. Until then, treat him like a clever friend who has read everything: worth listening to, and worth checking.
Soon you might ask for a bedtime story and have it read aloud by a cuddly toy Rory sitting at the end of the bed.
It took AI more than 75 years to get here. For most of that time it was small, slow and nearly forgotten. Then, in barely a decade, it learned to see, to talk, to think and to work in teams.
The child asleep in that bed will never know a world without it.
Rory is only just getting started. So is AI. The question is no longer what it can do. It is what we choose to ask of it.

The end, for now. Goodnight, Rory.
From the studio
What we do
Digital work for schools, charities, community groups and family businesses across Hertfordshire and Bedfordshire.

