what is an LLM? large language models, explained without the hype

ai · Oct 6, 2026 · 4 min read

LLM stands for Large Language Model, and the name is more literal than it sounds: a model of language, large.

the mechanism in one paragraph

during training the model reads a staggering amount of text with one task: predict the next word. it gets this wrong billions of times, and every mistake nudges billions of internal parameters a little. what emerges is not a database of facts — it is a compressed statistical map of how ideas, sentences and arguments fit together. when you ask it something, it navigates that map and writes the most plausible path through it.

that is the whole trick. everything else — chatbots, coding, translation — is that one mechanism pointed at different tasks.

what it is genuinely good at

  • transforming text you give it: summarize this, translate that, rewrite in plain language, extract the dates
  • drafting when a human will review: first versions, outlines, emails you will edit
  • explaining at the level you ask for: the same topic for a child or for an engineer
  • code, when the task is well-specified: it has read a lot of code, and code has strict syntax to be wrong against

what it is genuinely bad at

  • knowing anything recent (its knowledge stopped at training time)
  • counting, precisely citing, doing arithmetic with certainty
  • staying silent: it will always produce an answer, even when the honest answer is "there is no way to know"
  • your specific context: it has never met your customers, your codebase, your invoices

the three questions i ask before using one for anything real

  1. what happens when it is wrong? if the cost is a laugh, go wild. if the cost is a legal filing, put a human in the loop
  2. can i check the output faster than writing it myself? if not, the llm is a toy here, not a tool
  3. is the task transformation or knowledge? transformation (rewrite, summarize, translate) is where they shine. knowledge is where they hallucinate

i build websites, not models, and this post is a user's notes, not a researcher's. but that is exactly the perspective most explanations skip: the interesting question for most people is not "how does it work inside" — it is "what is it for, and where does it break".