Text in, text out

About this video

You type a question into a chat assistant, and a clear, confident answer appears. What is on the other side? At the centre of every chat assistant and every AI agent is a language model: a program that takes text in and gives text out. It can't click, open your files or go on the internet. It reads text and writes text.

A language model starts as a huge set of random numbers. It is shown a vast amount of text, books, websites and code, billions of pages, and again and again asked to guess what comes next. Each wrong guess nudges its numbers a tiny bit, so the right continuation becomes a little more likely. This training takes months on thousands of computers. To guess the next words of a chemistry book well, it helps to pick up some chemistry, so the model ends up with a great deal of knowledge and skill, stored as hundreds of billions of numbers.

Nothing it read is kept as a page it can go back to. So it can't check: if it remembers something slightly wrong, it says the wrong thing in the same confident voice, because a fluent, confident sentence is what usually comes next.

When a chat assistant searches the web, the model isn't the one searching. The app around it is. The model writes a request to search, the app does the search and pastes the results back in as text, and the model reads them and writes its answer. Whatever an AI system does, the model at its centre did one thing: it read some text and wrote some text. So treat its facts as a good first guess, give it the text it needs, and when it does something, ask which part did it.

Transcript

1. 0:00You ask an AI for a poem about your own cat, and it writes one. It never saw that poem. So what is on the other side?
2. 0:18A language model. Text goes in, text comes out. It can't click, open your files or go online. It only reads and writes text.
3. 0:40It learned by reading billions of pages and guessing what comes next. Every wrong guess nudged its numbers a tiny bit. Again and again, for months, on thousands of computers.
4. 1:06To guess a chemistry book well, it helps to pick up some chemistry. So what it learned is kept as numbers, hundreds of billions of them. There is no library inside.
5. 1:30So it can't look anything up, and it can't check. When it remembers something wrong, it says it in the same confident voice.
6. 1:54When it searches the web, the app around it does the searching. The model writes a request, the app searches and pastes the results back in, and the model reads them and answers.
7. 2:22So treat its facts as a first guess. Give it the text it needs. And when it does something, ask which part did it.
8. 2:40A model reads text and writes text. Everything else is the app.