How AI assistants actually work

You already use a language model every day. It lives in your keyboard. The auto-complete on your texting app is a tiny language model, and an LLM, a large language model, works the same way, just with much better prediction. Everything an LLM does, it does by predicting the next word. It weighs every word it knows, picks one, and then does that again.

That one idea explains more about your AI assistant than any feature list. This guide walks through the rest: what “thinking” means, who makes the models, and why the app you type into matters more than most people suspect. No technical background needed.

If it predicts well enough, you can call it thinking

These models do not actually think. When your assistant shows a “thinking” indicator, it is doing more of the same prediction, privately. Before writing the answer you see, the model predicts a hidden scratchpad of working-out. In that scratchpad it tries an approach in words, checks it in words, and discards it in words. More scratchpad usually means better answers, and it also costs more, because every predicted word is priced in tokens, the small chunks of text a model reads and writes.

Watch it happen in your own tool today: ask a question that needs a long answer and notice how the reply arrives word by word. It arrives that way because it is being predicted word by word.

The model is the auto-complete. The app around it makes it useful

A handful of providers make the models: Anthropic makes Claude, OpenAI makes the GPT models, Google makes Gemini. The app you type into is a separate thing, and sometimes a separate company. That app, the harness, decides what the model can see, what instructions it reads before your message, what tools it may use, and how long it may think.

This is why the same question gets noticeably different results in different apps running the same model. The brain matters, but the desk it sits at decides what the brain can actually do. If your company's assistant feels weaker than the demos you have seen, the harness is usually the reason, not the model.

How it learns: two phases, one ball

Training a model means rolling a ball around an error landscape until it finds the floor. Pre-training rolls it across the internet's valley: hundreds of thousands of books and a large share of the public web, on hardware that costs real money, for the equivalent of thousands of hours. Fine-tuning then rolls the same ball into a smaller notch: the behaviour a provider actually wants, helpful answers, refusals where needed, a particular voice.

You do not need the math. You need the consequence: a model knows what it read, it predicts rather than looks things up, and when prediction runs past what it read, it makes things up fluently. That is why judgment about what to trust is a skill your team can learn, not a mystery.

Five sentences to keep

An LLM is the auto-complete from your keyboard, running at an extraordinary scale. “Thinking” is more of the same prediction, done privately on a scratchpad, and every predicted word is priced in tokens. A handful of providers make the models, and the app you type into, the harness, is a separate thing. Training is a ball finding the floor of an error valley, first the internet's valley, then the notch that is your work. The same model inside a different harness will give you very different results.

This article is adapted from lesson one of The Brain and the Desk, our nine-lesson course on how AI assistants actually work. The full lesson, with the diagrams, is free and whole, no signup: read lesson one as a PDF.

We also teach the course live: corporate training for whole departments, taught inside the tools and policies your company already has. The curriculum, formats, and prices are on the training page.