Cortexa AI Glossary · The basics
What is an AI model?
From Cortexa Learn, by Cortexa Consulting. Last checked .
The word in every announcement, and the one thing it means.
The little menu
Open some chatbot apps and you'll find a small menu near the top that lets you pick a model. Or you'll read a headline saying a company has released a new model. The word turns up everywhere in artificial intelligence (AI), and it means one specific thing. Once you know what that is, the menu and the headlines both get easier to read.
What training leaves behind
An AI model is what's left after training. During training, a computer works through a huge number of examples and keeps adjusting a long list of numbers until its guesses come out right more often. When training ends, those numbers are saved. That saved set of learned numbers, together with the structure that puts them to use, is the model. You can think of it as one very large file.1
Something in, something out
A model does one kind of job. Something goes in, and something comes out. Give a photo model a picture, and out comes a label, like "dog" or "beach." Give a language model a question, and out comes text. Give a weather model today's readings, and out comes a forecast. The model doesn't change while it does this. It runs the same learned numbers on whatever you hand it.1
Model and app are different things
This is where people often get tangled. The model and the app are two different things. The model is the learned part. The app is the house built around it: the chat window, the buttons, your account. The same model can sit inside many different apps, and one app can switch between several models. That's what the little menu does. Same house, different engine in the back room. Companies often offer a few models side by side: a larger one that takes longer and costs more to run, and a smaller one that answers quickly. Picking from the menu is picking which file does the work.2
What the app adds
Much of what you notice in an AI tool comes from the app around the model. The app can add a search step so an answer includes today's news. It can let you upload files, remember your settings, and keep the history of your conversation. So two apps built on the same model can feel quite different. And when a tool gets a new feature, the model underneath may not have changed.3
A new model means new training
When a company announces a new model, it has usually trained a new set of numbers, often with more data, a different design, or both. It's a new file. And a model usually doesn't keep learning from each conversation as it happens. Its numbers were set when training ended, which is one reason an AI tool can be unaware of recent events unless the app looks them up.234
Reading an announcement
So when a headline says "new model," you can read it plainly: someone trained a new set of learned numbers, and every claim in the story is about what that file can do. A few fair questions follow. What was it trained on? What is it better at, and measured how? And which of the apps you use will run it?