How Models Think

You ask a chatbot a question and a paragraph appears, one word at a time, sounding like someone who knows. It helps to know what actually happened. The model read an enormous amount of text and learned which words tend to follow which. When you press send, it picks one of the likely continuations — not always the top one, which is why the same question can give you a different answer twice — appends it, then does the whole thing again with your question plus its own last word.

That loop is the entire trick. No database sits behind it, and no hidden reasoner checks the facts. It is predicting the next token — a token being a word or a piece of one — a few hundred times in a row.

Two moments get confused, and separating them explains most of the odd behavior. Training happened once, long before you opened the app: months of processing text into the patterns the model carries. Answering — inference — happens every time you press send, running those frozen patterns over your words.

Your chat teaches it nothing. Correcting it does change what it says next, because your correction is now part of the text it is predicting from, not because it learned anything. Close the conversation and the correction is gone. Some products have a memory feature that stores your text and pastes it back in later — still not learning, just re-reading.