What is a large language model (LLM)? A big AI model trained on huge amounts of text that writes by picking the next token

A big AI model trained on huge amounts of text that writes by picking the next token.

A large language model (LLM) is a big AI model trained on an enormous amount of text: web pages, books, code. What it does underneath is simple. It looks at the text so far, picks the next token, adds it, and repeats.

"Large" means billions to hundreds of billions of parameters, or more. Most are built on the Transformer design from 2017. In 2020 GPT-3 surprised people by picking up new tasks from a few examples in the prompt, and in 2022 ChatGPT put an LLM behind a chat window anyone could use.

Because it picks what sounds plausible next, it can confidently state false things. This is called hallucination, so check anything important.

Think of it like this

Think of the next-word suggestions on your phone keyboard: type "what's for" and it offers "lunch". An LLM is that feature grown on an unimaginable amount of text, until it can keep going for a paragraph or a whole essay.

For example

Replies seem to stream out bit by bit because the model really is picking and adding one token at a time.

Common mix-ups

Mix-upA large language model plans the whole answer first, then writes it out.
ActuallyIt can't go back and erase what it has already written; it can only keep adding. So if it sets off in the wrong direction, it tends to follow that path to the end — asking "please check that again" afterwards can sometimes help.
Mix-upIf it sounds confident, it's right.
ActuallyThe tone is just as smooth whether the answer is right or wrong, so it's no clue. Ask the same question two or three times in different words — if the answers disagree, be especially suspicious of that part.

Try it in class

For teachers at the start of a lesson — or for students working on their own.

Question: What does your phone keyboard's next-word suggestion have in common with a large language model, and what's different?
Activity: Build one sentence by tapping only your keyboard's suggested words, then compare with classmates. Next, put the 2003 (E0018), 2020 (E0108) and 2022 (E0181) records in order and write one line for each on how "guess the next word" went from a research idea to a chat window anyone can use.