Large Language Model (LLM)

Everything an LLM does is next-token prediction over a context. Training establishes a distribution over what typically follows what; a later alignment stage — instruction tuning and preference optimisation — reshapes that distribution towards answers people judge helpful and safe. At inference the model produces a probability distribution over the vocabulary at each step, and a sampling strategy picks the next token from it. There is no lookup table and no database inside the weights, which is why a model can be fluent about a topic it has no reliable information on.

For engineers, the practical consequences follow directly from that mechanism. Output is stochastic unless you constrain sampling, so identical inputs need not yield identical outputs. Cost and latency scale with tokens in and tokens out, which makes prompt size an architectural concern rather than a formatting detail. And the model’s knowledge has a boundary in time; anything after it must arrive through retrieval or a tool call.

The misconception that causes the most damage in production is treating the model as a knowledge base. It is a language engine that has absorbed a great deal of knowledge as a side effect. When correctness matters, the reliable pattern is to supply the facts in the prompt or fetch them with a tool, and use the model for what it is genuinely good at: understanding the request, and shaping the response.