A language model writes the next chunk of text that fits the text you already gave it. That is the mechanism, the basic move. Chat windows, agents, and copilots are products wrapped around that move. They add buttons, files, and a bill. They do not change the basic move.
IBM’s public explainer puts large language models in the family of models trained on a lot of text to produce text. Google’s crash course walks the same idea, from data to a model that generalizes. Hugging Face’s free courses walk the implementation. The model generates text. The product decides what information and tools it can use. The course homes linked above list the lectures.
The reply is a continuation
You send a prompt, the text of your request. The model assigns likelihoods to possible next pieces and picks one, then does it again. “Understanding” is a useful metaphor and a bad specification. The model does not look up a fact unless the product added a lookup. If the lookup is missing, a smooth sentence can still be wrong. The model was trained to write fluent text. A citation exists only if the product put a source in the request or the reply.
Two chats can answer the same question differently. Temperature and similar sampling settings, when the product exposes them, change how often a less likely piece gets picked. A higher setting is not a research method. It is a wider roll of the dice.
The product is not the model
When a page says “ChatGPT” or “Claude,” it usually means a hosted product: an app, a set of tools, a usage policy, and a price. The model name inside that product can change while the logo stays. The tool profiles on this site date the public pricing page they used. They summarize dated public information. They are not hands-on benchmarks, and this course does not add a comparison test.
A local runner such as Ollama is the other shape. The app runs on your machine. The weights you download come with their own license. Running it locally does not make the weights yours to relicense.
What to ask before you trust a paragraph
Before you trust an answer, ask these three questions.
- What text was the model allowed to see? If your handbook was not in the prompt and nothing retrieved it, the model is guessing from training, not reading your policy.
- What tools was it allowed to call? A pure chat cannot file a refund. A product with a refund tool might, and it might do it when you did not want that.
- What would count as wrong, in a way you can check? “Sounds smart” is not a check. A date, a price, a function name, or a quoted sentence from a page you can open is a check.
Take one sentence a model gave you this week. Underline any part you could verify by opening a page, and any part you believed because the sentence was smooth. If the second list is longer, you have a draft, and you do not yet have a source you can open.
Sources & limits: The definitions here follow the IBM, Google, and Hugging Face pages linked above. The tool profiles are dated summaries of public pages, not a test we ran.