From training data to the sentence it just wrote — a clear explanation for curious people who have never written a line of code.
A model like Claude didn't start smart. It started as a system with billions of adjustable dials — called weights — all set to random values. Then it read text. An enormous amount of it: books, articles, conversations, and more.
For each piece of text, the model tried to predict what word came next. It guessed. It was wrong. Then its dials were adjusted, just slightly, to make a better guess next time. This happened billions of times across trillions of words — a process called gradient descent, where the model is nudged, again and again, toward being less wrong.
By the end, those dials encode something remarkable: a compressed statistical map of how language works, and by extension, how much of human knowledge is expressed in words.
Modern AI models are built on a design called the transformer. Its key innovation is attention: when the model processes a word, it doesn't look at it in isolation. It asks, "Which other words in this passage are most relevant to understanding this one, right now?"
Take the sentence: "The trophy didn't fit in the suitcase because it was too big." To figure out what "it" refers to, your brain scans back. You decide "trophy" makes more sense than "suitcase" given context. A transformer does exactly this — in parallel, for every word, across many layers of analysis at once.
Your prompt is broken into tokens, each token "attends" to the others across many layers, and the model predicts one word at a time.
When you send a message, the model doesn't look it up in a database. It converts your text into a sequence of tokens — roughly word-sized chunks — and passes them through dozens of transformer layers. At each layer, every token updates its understanding based on every other token.
By the final layer, the model has built up a nuanced picture of your question, and it uses that picture to predict the most fitting next word. Then it appends that word and runs through the whole process again — and again — until the response is complete.
This is why models can reason through problems they've never seen before. They aren't retrieving; they're constructing — building an answer token by token, guided by patterns baked in during training.
There are real limits. The model has no memory of past conversations unless you provide them. It can be confidently wrong — it generates plausible text, not verified truth. And its knowledge cuts off whenever training stopped. Understanding this helps you use it well: treat it as a very well-read collaborator who reasons carefully but can still be mistaken.