A large language model (LLM) is trained on large amounts of data to predict what comes next from context. It is good at language and code, but a fluent answer is not automatically a factual one.
Three basic ideas first
Token: a fragment the model uses to process text; it is not exactly the same as a character or a word. Different models and encodings produce different token counts. Context window: the maximum input and output capacity the model can consider in one interaction; a larger window does not guarantee that every detail will be used correctly. Prompt: the task, background, and output constraints you give the model.
What can it do, and what can’t it do?
Typical tasks include drafting email, extracting information, organizing documents, assisting with code, and explaining concepts. A model can hallucinate—producing plausible-looking errors when it lacks evidence. For important medical, legal, or financial decisions, check with a qualified professional.
Separate chat products from APIs
A chat product is designed for direct use in an interface. An API is an interface that programs call; it usually involves accounts, authentication, request parameters, error retries, and usage billing. A site claiming to provide an API does not necessarily support every model or feature you want.
A first exercise for beginners
Choose a public passage with no personal information and ask the model for both a “three-sentence summary” and a “list of facts to verify.” Compare both with the source and look for omissions or inventions. This teaches you more than simply asking whether the model is “smart.”
How does a model “learn” these abilities?
During training, many parameters are adjusted so the model can learn statistical patterns in its data. Instruction tuning and preference optimization may then make its output more useful in conversation. An everyday question is usually inference, not a new training run; whether a platform uses your data for later training depends on its terms and settings.
Inference, tools, and multimodality
Some models support more elaborate reasoning processes, while others accept image or audio input. Tool use usually means the model proposes a structured request and the application executes it; generating a function name does not by itself give a model permission to browse the web or make a payment. Verify each supported input and output type for the specific model.
Common misunderstandings
- Confidence does not prove that an answer was fact-checked.
- A longer context does not mean every page will be cited accurately.
- More parameters do not make a model better for every business task.
- Open weights do not mean use is free from license restrictions.