A sunny beach
I’m going to a sunny beach. In my bag, I packed a towel and sunscreen.
In this example, a towel fits a trip to the beach.
Try a simple example
How does AI build an answer?
The details you give can change an AI answer. Try this simple example.
Prepared teaching exampleNo typing needed
What word might come next?
I’m going to a sunny beach. In my bag, I packed a
Pick a word to see the possibilities in this example.
One new detail changes the possibilities.
A useful detail can change what comes next. When you ask an AI for help, tell it what it needs to know.
Use that insight in a real request
An AI can use the details you include. Compare a vague request with one that gives it something to work with.
Authored comparison, not live AI output
Your request
Write an invitation.
Example response
You’re invited to a special event! Join us for a wonderful time with friends and neighbours. We look forward to seeing you.
The details are now available to the model. They improve the brief, but you still need to check the answer against them.
Try it on something you actually need
When you’re ready to try an AI chat, choose something you need to write. Say who it is for, what you want to say and which facts to keep. Check the reply before using it.
Read the full examples
These chances are invented for this example. They are not measurements from an AI.
A snowy walk
I’m going for a walk in the snow. In my bag, I packed a coat and gloves.
Change the trip to a walk in the snow, and this example suggests a coat first.
A quiet library
I’m going to read at the library. In my bag, I packed a book and notebook.
For a visit to the library, this example suggests a book first.
At the beach, this toy gives towel 70%, coat 10% and book 20%. In the snow, it gives towel 5%, coat 85% and book 10%. At the library, it gives towel 5%, coat 5% and book 90%. Choosing the highest value begins three different continuations. The numbers illustrate the mechanism; they are invented.
Use that insight in a real request
Authored comparison, not live AI output
Only the task
- Your request
- Write an invitation.
- Example response
- You’re invited to a special event! Join us for a wonderful time with friends and neighbours. We look forward to seeing you.
Add useful context
- Your request
- Write a short, friendly invitation to a free neighbourhood seed swap. Saturday, 24 October, 10 am–1 pm, at the library meeting room. Bring labelled seeds if you have them. Everyone is welcome. Do not invent details.
- Example response
- Come to our free neighbourhood seed swap on Saturday, 24 October, from 10 am to 1 pm in the library meeting room. Bring labelled seeds if you have them, or simply come along. Everyone is welcome. Meet neighbours, share what you grow, and find something new for your garden.
Try it on something you actually need
When you’re ready to try an AI chat, choose something you need to write. Say who it is for, what you want to say and which facts to keep. Check the reply before using it.
Go deeper
Prepared teaching example
A language model doesn’t pull one finished answer from a drawer. It generates a continuation from the context it has, one token at a time. A token is a piece of text; it is not always a whole word.
To make the sequence readable, this toy uses whole-word tokens and punctuation. It looks up authored probabilities and always takes the highest one. It has no neural network, training or live inference. Real token vocabularies are much larger, and generation can sample a less likely token. This illustration stops after three additions; real generation can stop on an end token, a length limit or another stopping condition.
At the beach, this toy gives towel 70%, coat 10% and book 20%. In the snow, it gives towel 5%, coat 85% and book 10%. At the library, it gives towel 5%, coat 5% and book 90%. Choosing the highest value begins three different continuations. The numbers illustrate the mechanism; they are invented.
What is the model actually calculating?
A decoder language model produces scores for possible next tokens from the tokens before them. A softmax turns scores into a distribution. The decoding procedure chooses a token, appends it, and repeats. This forward generation is different from filling a masked word using text on both sides.
Where do these patterns come from?
During training, model parameters are adjusted using examples. At generation time, the context helps determine the next-token scores using those learned parameters. Changing this prompt changes the computation for this answer; it does not, by itself, retrain the model.
Why can a likely continuation still be wrong?
The next-token distribution concerns what text follows in context. It is not a direct probability that a real-world claim is true. A coherent answer still needs evidence. Tool use, retrieval, evaluation and human review can add checks, with their own limits.
Explore the original sources
Primary references. These explain the mechanisms; the examples here are authored teaching simulations.
Try a conversation next
See how a small follow-up can change a familiar message. No typing needed.
Write a message with AI