Large Language Model

A large language model is trained on an enormous corpus of text with a deceptively simple objective: given some text, predict the next token. Do that at sufficient scale and something surprising emerges — to predict text well, the model has to encode grammar, facts, styles, reasoning patterns and a great deal of implicit world knowledge.

The result is a general-purpose text engine. The same model can summarize a contract, draft an email, explain a codebase or argue a position, because all of those are text-prediction problems in disguise. Modern models extend this to images, audio and video, at which point they are usually called multimodal.

The next-token framing also explains the failure modes. The model is optimizing for plausible continuations, not for truth, which is why a confidently wrong answer looks exactly like a correct one.

Nearly every AI product you encounter is a language model with an interface and some plumbing around it. Understanding that they predict rather than retrieve explains most of their behavior: why they are fluent but sometimes wrong, why phrasing a request differently changes the answer, and why giving them the right source material up front matters more than asking the question more forcefully.

That it looks things up. A language model has no database and performs no search. It predicts the next token from patterns learned in training, which is why the same question phrased two ways can give different answers, why it cannot cite a source unless you supply one, and why it produces a fluent answer about a topic it knows nothing about. Everything that feels like recall is reconstruction.

That "it's just autocomplete" settles the question. Mechanically accurate, and misleading as a conclusion. Predicting text well at sufficient scale required the model to encode grammar, factual associations, reasoning patterns and a great deal of implicit world knowledge — the capability is real even though the mechanism is humble. The dismissive framing and the overclaiming one fail the same way: both substitute a slogan about the mechanism for a look at what the thing can actually do.

That it learns from your conversation. The model is frozen; nothing you type changes its weights, and it forgets everything outside the current context window. Whether your data is retained or used to train a future version is a question about the vendor's terms, not about how the model works — and the answer usually differs between consumer and enterprise tiers. Check the contract, not the technology.

The practical skill is supplying context. A model asked to answer from its training alone is guessing from memory; the same model given the relevant document usually answers well. That gap is the entire reason retrieval-augmented generation exists.