LLM (Large Language Model)

A large language model is a neural network — almost always a transformer — trained on trillions of words to do one narrow thing: predict the next token given everything before it. Scale is the defining trait. There is no formal cutoff, but the field informally reserves "large" for models with roughly ten billion parameters and up (IBM). Those parameters encode statistical patterns of language — grammar, style, reasoning shortcuts, and a compressed, lossy copy of whatever facts appeared often in the training data.

Mechanically, the model outputs a probability distribution over the next token and samples from it, one token at a time. Everything an LLM "knows" is expressed as the conditional probability of the next token given the preceding ones; there is no lookup table and no separate store of facts. In support, this is what drafts a reply, summarizes a thread, tags a ticket, or — wired to your knowledge base and allowed to send — resolves a contact on its own. Grounding the model in retrieved documents (see RAG) is what turns a plausible-sounding generator into something you can put in front of a customer.

What it hides: fluency is not accuracy. An LLM has no internal sense of true versus false — it produces the most likely continuation, which is often right and confidently wrong the rest of the time. On grounded summarization, the best models now hallucinate on only about 1–2% of documents (Vectara), but that is the easy case, where the answer sits in the passage you handed it. On harder, open-ended work the floor collapses: a Stanford study found purpose-built legal tools with retrieval still hallucinated on 17–33% of queries (Stanford HAI). Treat any unreviewed LLM output as a draft, not a fact.