01Prompter8 min
Brainstorm with AI without getting generic ideas
Give it constraints and your own examples, then ask for the unexpected.
Learn · 3 minTry it · 5 min
Start lessonUnderstand · Foundations
Foundations
What does a model actually do?
It predicts what comes next. That one idea explains most of what AI gets right and wrong.
5 min readPlain English
Read it in
In one sentence
A language model predicts what comes next: it reads your text, picks a likely next word, adds it, and repeats until the reply is done.
Think of it like…
Hum the first few notes to a musician who has played thousands of songs, and they’ll carry on in a way that sounds right, even for a tune they’ve never heard. Sounding right and being the actual song are not the same thing.
How it works
Everything a chat assistant does, from answering a question to drafting a policy, is built from one move repeated very fast: look at all the text so far and predict what comes next.
- 1ReadYour message, the chat so far and any instructions or files are split into tokens: words and pieces of words.
- 2WeighUsing patterns learned in training, the model rates how likely each possible next token is, given everything before it.
- 3PickIt chooses one of the likely options. A dash of chance in that choice is why the same question can get a different answer twice.
- 4RepeatIt adds that token to the text and predicts again, one token at a time, until the reply looks finished.
Unless the tool is connected to search or your files, nothing gets looked up. The answer comes from learned patterns, not a database of facts.
Good at, bad at
Good at
- Fluent writing in almost any tone
- Everyday emails, summaries and lists
- Following a pattern or example you give it
Bad at
- Rare facts, names and exact figures
- Original ideas, unless you push it
- Knowing when it’s guessing
Where you’ll see it at work
Routine emails. A plumbing firm’s quote follow-ups have a familiar shape, which is exactly what prediction handles well.
Changing the tone. A design studio turns a blunt note to a late-paying client into a firm but friendly one. Tone is a pattern the model knows well.
Product descriptions. An online shop gets a lively description in seconds, then checks every size and material, because plausible isn’t the same as true.
Myth vs reality
Myth: It looks answers up in a giant database.
Reality: It writes answers from patterns it learned. Some tools search the web or your files first, but the writing itself is still prediction.
Words to know
Test yourselfNo grades
Why can the same question get two different answers?