™ChatGPT, Claude, Gemini and Copilot are all built on something called a large language model, usually shortened to LLM. You don't need the math to use one well, but knowing roughly how it works explains a lot of its strengths and quirks.
It learned from an enormous amount of text
An LLM is trained on a huge collection of writing: books, articles, websites, code and more. During training, it plays one game billions of times. It sees some text and guesses what comes next, then adjusts itself whenever the guess is wrong.
Over time it picks up grammar, facts, writing styles, how arguments are structured, and how to follow instructions. Nobody programs those rules in. They emerge from the patterns in the text.
When you use it, it writes one piece at a time
When you send a message, the model reads it and predicts the most fitting next word. Then the next. Then the next. (Technically it works in small chunks called tokens, which are pieces of words.)
That sounds like fancy autocomplete, and in a sense it is. But with enough patterns learned, predicting the next word well means it can summarize a report, draft a polite reply to an angry customer, or work through a problem step by step.
Why it sounds so confident
The model is trained to produce text that fits. It is not checking facts against a database. Fluent, confident writing is what fits, so that's what you get, even when the underlying fact is wrong.
This is the single most important thing to understand about these tools. Fluent is not the same as accurate.
Training happens once. Your chat is the working space.
The big learning happens during training, which takes months and happens before you ever use the model. After that:
- The model has a knowledge cutoff. It knows about the world up to the date its training data ends. Many tools now search the web to fill the gap, but not all of them do it every time.
- Your conversation is its working space. Everything you've typed or attached in the current chat is what it can see right now. This is called the context, and it has a size limit.
- It doesn't learn from your chat the way a person would. Some tools have a separate "memory" feature that saves notes about you between chats, but the model itself isn't retrained by your conversation.
Different companies, similar machinery
OpenAI makes the models behind ChatGPT. Anthropic makes Claude. Google makes Gemini. Microsoft's Copilot draws on models from other AI labs and connects them to Microsoft's apps. They differ in style, strengths and features, and the lead changes often. Under the hood, the basic idea is the same.