When people talk about AI, they usually mean an app: ChatGPT, Claude or Gemini, on a website or on your phone. The app is the part you see, with the text box and the chat history. The model is the part inside that reads your message and writes the reply.
Think of the app as a car and the model as its engine. You sit in the car and steer, but the engine does the work. The names can be confusing, because companies often use one name for both. Claude and Gemini are apps and also families of models, while ChatGPT mostly runs on models called GPT.
How a model learns
A model starts out knowing nothing. During training it reads an enormous amount of text, from books, websites and other sources, and plays one simple game over and over: guess the next word. After each guess, its inner settings are nudged slightly towards the right answer.
After countless rounds, it gets very good at the game. To predict the next word well, it has to pick up grammar, facts and how ideas connect. It's a bit like children learning to talk by listening, not from grammar books. A second stage of training then teaches it to follow instructions, hold a conversation and turn down harmful requests, often with people rating its answers.
Once training is finished, the model itself stays fixed: it doesn't change as you chat, although the app may remember things you've told it. Its knowledge stops at the point its training material was collected, so it won't know recent news unless the app lets it search the web.
What parameters are
Inside a model is a vast set of numbers called Parameters: The numbers a model learns during training: a rough measure of its size. Bigger usually means more capable, but slower and pricier to run. 7B means 7 billion., often billions of them. They're the settings that training adjusts. Picture a mixing desk in a recording studio with billions of tiny dials: training turns each one a fraction at a time until the output sounds right.
The number of parameters is a rough measure of a model's size. You'll see it written as 8B or 70B, where B stands for billion. Bigger models usually know more and handle harder problems, but they need more computing power, so they're slower and cost more to run. Size isn't everything, though: a well-trained small model can beat an older, bigger one.
One model, many apps
Labs don't only put their models in their own apps. Any company can plug a model into its own product and pay for what it uses. This works through an API: How developers plug a model into their own apps and pay per use. A model that's "in the API" can be used by programmers, even if it isn't in a chat app yet., a way for one piece of software to use another. So the same model might sit behind a chatbot, a writing tool and a customer service assistant.
This is why apps change. When a lab releases a better model, apps can switch to it, often without changing anything you see, like fitting a new engine to the same car. ChatGPT has moved to newer models many times since it launched in November 2022.
Usually the switch makes an app quicker or more capable overnight, though sometimes it just feels different for a while. Either way, a new model is often the first sign that the apps you use are about to change.