A large language model (LLM) is a neural network trained to predict the next token in a sequence. A token is a small piece of text. During generation, the model reads the prompt, assigns a probability to each possible next token, chooses one, and repeats the process.

The model does not look up a complete answer stored in a database. Its training has encoded patterns about language, facts, code, and common reasoning into its weights. This lets it produce useful text, but it can also produce a fluent answer that is wrong. Chat models receive extra training after basic text prediction so they can follow instructions, use tools, and respond like assistants.

A token is a unit of text that a language model reads or writes. It may be a whole word, part of a word, punctuation, or whitespace. A tokenizer converts text into token IDs before the model processes it.

Tokens balance vocabulary size and sequence length. Character-based input would create very long sequences. Whole-word input would need a huge vocabulary and handle new words, typos, code, and many languages poorly. Subword tokens can represent any text while keeping sequences manageable.

Tokenization matters because context limits, cost, latency, and rate limits are usually measured in tokens. The exact count depends on the model's tokenizer, so estimates based only on characters can be inaccurate.

Token counts affect whether an LLM feature is affordable, fast, and reliable. Providers usually charge separately for input and output tokens. More input increases prompt-processing work, while more output increases generation time and often costs more per token.

Tokens also determine whether a request fits inside the model's context window. The prompt, conversation history, retrieved documents, tool descriptions, and expected response must all fit. Large requests can fail or lose important content through truncation.

Measure tokens on realistic traffic instead of using one average. Track input, output, cached, and reasoning tokens when the provider exposes them. Common savings include removing repeated instructions, retrieving fewer documents, limiting output length, and routing simple tasks to smaller models.

A context window is the maximum amount of information a model can process in one request. It normally includes system instructions, messages, tool definitions, documents, images represented as tokens, and the generated response.

The context window is not permanent memory. Most APIs require the application to resend any history that the model should consider. As a conversation grows, the application may need to drop old turns, summarize them, or retrieve only relevant facts.

Do not fill the entire window with input. Leave room for the answer and for any tool results added later. A large advertised window does not guarantee equal attention to every token. Test whether the model can find and use important details across realistic long prompts.

Temperature changes how strongly the model favors its most likely next tokens. A low value makes common choices more dominant, which usually produces more consistent output. A higher value gives less likely choices more chance, which can add variety but also increase mistakes or drift.

Use lower temperature for extraction, classification, and strict formatting. Moderate values can help with brainstorming or creative drafts. Temperature does not control truthfulness: the model's most likely answer can still be wrong.

Some models, especially reasoning models, do not expose temperature or support only a default value. Treat sampling settings as model-specific. Change one setting at a time and compare results on an evaluation set rather than choosing a value by intuition alone.

Pretraining teaches a model broad language patterns by asking it to predict missing or next tokens across a very large dataset. This stage builds most of the model's general knowledge and basic abilities. The result is often called a base model.

Post-training teaches that base model how to behave for particular uses. It can include examples of good instruction following, human or AI preference feedback, safety training, tool use, and reinforcement learning on tasks with verifiable answers.

The distinction matters because a knowledgeable base model may only continue text instead of answering a request. Post-training makes it more helpful and easier to control, but it can also introduce biases such as excessive agreement or refusal. Both stages shape the final behavior.

A hallucination is a claim produced by a model that sounds plausible but is unsupported or false. Examples include an invented citation, a nonexistent API, or a wrong date stated with confidence.

LLMs learn to generate likely text, not to verify every statement. Their training data can be incomplete, outdated, or conflicting. Knowledge stored in model weights is also an imperfect compression of that data. When a prompt requires missing information, the model may continue with a pattern that resembles a good answer.

Applications should assume hallucinations can occur. Useful controls include retrieval from trusted sources, calculation tools, and clear permission to admit uncertainty. Tie citations to source text and evaluate whether claims have factual support.