Training data
The text, images and other material a model learned from before you ever used it.
Before a model answers anyone, it is trained on a huge collection of material: public web pages, books, code, licensed data and, for some products, conversations from users who allowed it. The model doesn't keep that collection as copies to look things up in. It keeps patterns: which words, facts and ideas tend to go together.
2 things follow. What was common in the training data comes out confidently and is usually right; what was rare, like a small company's founding year, can come out as a plausible guess. And the training data stops at a date, so anything after it is missing unless the app searches the web.
For example
A model can retell the plot of a famous novel well, because it was discussed everywhere, and still invent the page count of a self-published one.
Put it to use
Where you'll meet training data in ChatGPT, Gemini and Claude, how to use it there, and the mix-up most people make.
Daily Bytes Plus is in the Android app.