What is a large language model?

At a glance

• It does one thing: guess the next word, over and over. Everything it “knows” is a side effect of doing that well across trillions of words^[4].
• It is a prediction engine, not a fact database. Confident, fluent, wrong answers (hallucination) are permanent, not a bug to be patched.
• Scale made it useful: billions of parameters trained on internet-scale text^[3]. But bigger is not always better for your job.
• You rent a hosted model and pay per “token” (about 3/4 of a word) for text in and out. You almost never train one yourself.

How it works

Given “The capital of France is”, the model scores candidate words and writes the likeliest, “Paris”, then repeats^[4]. To get good at this across the whole internet, it must absorb grammar, facts, styles, and code^[1]. The fluency in ChatGPT or Claude is that single trick done extremely well^[2].

Why it sounds certain when wrong

It picks the most plausible-sounding words, with no internal sense of true or false, so it states fabrications in the same confident tone as facts. The fix is how you use it: feed it your trusted documents at question time (retrieval) and keep a human reviewing anything high-stakes.

What it means for buying

You are renting a general prediction engine billed per token. At scale, model size and caching can swing the bill enormously. Training your own from scratch costs tens of millions and needs research teams^[4]; nearly every business should instead use a hosted model and compete through its data and safeguards^[3].

Bottom line

An LLM is a next-word predictor that scaled into a brilliant, fast, confidently fallible assistant — rent one, ground it in your data, and put guardrails around it.