How LLMs Work
Open the hood on large language models and see the simple idea that powers chat assistants.
A very good autocomplete
A large language model, or LLM, is the engine behind tools like chat assistants. At its core it does one thing: it reads the text so far and predicts what word is likely to come next. Type "The sky is" and it will lean toward "blue." Repeat that prediction thousands of times in a row and you get whole paragraphs, emails, and code. It is autocomplete taken to an astonishing extreme.
Learned from mountains of text
An LLM gets its skill from training, where it reads an enormous amount of writing from books, articles, and the web. While reading, it adjusts billions of internal dials so its next-word guesses get better and better. Nobody hand-codes grammar or facts into it; the patterns of language and knowledge are absorbed from all that text. This is why models can write fluently about so many topics.
It predicts, it does not look up
Here is a key surprise: by default an LLM is not searching a database or browsing the web. When you ask a question, it generates an answer one likely word at a time based on patterns it learned. That makes it creative and flexible, but also means it can sound confident while being wrong. Some products add real search or tools on top, but the raw model is a predictor, not a fact lookup.
Training vs. chatting
Training and using a model are two different events. Training happens once, takes huge computing power, and is where the model actually learns. Chatting with it afterward is called inference, and it does not teach the model anything new by default. So your everyday conversation usually does not update the model; it just runs the already-trained predictor on your words.
Key takeaways
- An LLM is autocomplete at an extreme: it predicts the next likely word over and over.
- It learns patterns from huge amounts of text during a one-time training phase.
- A raw model predicts rather than looks up, so confident answers can still be wrong.
4 questions · pass at 60% to earn XP