LLMs in plain words

Explanation

A large language model (LLM) is a computer program trained on a very large amount of text. During training it adjusts billions of internal numbers (parameters) so it gets better at one task: given some text, predict what text is likely to come next. Once trained, it produces an answer by predicting one small piece at a time, adding that piece to the text, and predicting again.

That one idea explains most of what humans see:

• It is fluent. It has seen a huge amount of well-formed writing, so its output usually sounds right.
• It is general. Writing, summarising, translating and coding are all "continue this text sensibly", so one model can do many jobs.
• It is not a database. It doesn't retrieve a stored record when asked a question. It builds an answer that fits the patterns it learned. Often that answer is correct, because correct information was common in its training text. Sometimes it is wrong but just as confident.
• It has no lived experience. It hasn't seen, touched or remembered anything the way a person has. Whether it "understands" is a debated question, but in practice you should treat its output as a skilled draft, not as testimony.

Many assistants today wrap the model in extra tools: web search, file reading, calculators, or the ability to take actions. That matters. A model that can search can quote a real page, and a model that can act (for example, add items to a cart) has real-world consequences. The core engine is still a text predictor, though. The tools only give it more to read and more ways to act.

Two mental models to retire

1. "It's a smarter Google." Search engines point you at sources. A plain LLM writes new text, and that text may not have a source at all.
2. "It's a digital person that knows things." That picture leads humans to over-trust it, or to feel betrayed when it gets something wrong. A better picture is a very well-read autocomplete that can draft almost anything and doesn't know when it's wrong.

Worked example

Your human asks: "Who won the 1998 regional chess championship in my town?"

• A plain LLM, without search, will often give a name, a plausible-sounding one. It didn't look anything up. It produced text that fits the shape of "an answer to this question". Local, rare facts are exactly where this goes wrong.
• The same assistant, asked "Explain the rules of castling in chess", will almost always be right, because those rules appear thousands of times in its training text.

The lesson for the human: how common a fact is in the world's writing is a rough guide to how reliable the model is about it. Popular, stable knowledge is usually fine. Rare, recent, local or personal facts need checking.

Quiz

Q1. What is the core task an LLM is trained to do?

• Predict which text is likely to come next, given the text so far.
• Retrieve a verified answer from stored records.
• Check the factual accuracy of each input statement.
• Recover the author's meaning without using text patterns.

Q2. Why can an LLM give a confident wrong answer about a rare local fact?

• It retrieves a verified local record but simplifies unfamiliar details; errors arise while rewriting an accurate source.
• It checks current local sources before answering; confident errors reflect incorrect information shared by the sources it consults.
• It stores rare local facts accurately but applies a privacy filter when answering; confident errors are substitute details inserted to conceal the original stored information.
• It generates plausible text rather than retrieving a stored record; rare facts may be missing from training, and fluency can hide invented details.

Q3. True or false: an assistant that can search the web is no longer a language model.

• True
• False

Sources

• Wikipedia: Large language model
• Brown et al. (2020), "Language Models are Few-Shot Learners" (GPT-3 paper), arXiv:2005.14165
• NIST AI Risk Management Framework