How does guessing the next word turn into an answer?
Give a model the words "The best month to plant tomatoes in Adelaide is" and it scores every possible next piece of text by how likely it is to follow. It picks one, adds it to the end, and scores again with the new word included. A 300-word email is a few hundred of those guesses in a row. Nothing is looked up. The answer is assembled from patterns the model absorbed in training.
The design that made this work at scale came from a 2017 paper by Google researchers, "Attention Is All You Need". It introduced the transformer, a way for the model to weigh every earlier word when choosing the next one, so that "it" near the end of a long sentence gets linked back to the right noun. The T in GPT is that transformer: generative pre-trained transformer. "Large" is the scale: billions of adjustable values inside the model, tuned on a big share of the public web, books and code (the entry on training data covers where it comes from).
Why does a chat assistant answer you instead of carrying on your sentence?
A model fresh from that first stage of training is a text completer. Ask it a question and it might reply with three more questions, because that is what a lot of text on the internet does. A second stage teaches it to act like an assistant: people write example answers, rate the model's attempts, and the model is trained towards the answers people preferred. OpenAI described the method in a 2022 paper, "Training language models to follow instructions with human feedback", and some version of it sits behind every chat assistant you have used.
That second stage shapes manners, tone and what the model refuses to do. It does not teach the model your business. An LLM knows nothing about your prices, your suburb or last week's sales unless something puts them in front of it.
What is an LLM good at, and where does it slip?
| The job | How an LLM does | What to check |
|---|---|---|
| Drafting in a tone you describe | Well, and fast | That it still sounds like you |
| Summarising a document you give it | Well, if the document fits in its context window | That nothing in the middle was skipped |
| Facts about your business | Only what you tell it; it fills gaps with plausible guesses | Every price, date, name and claim |
| Events after its training cut-off | Unaware of them unless the app searches the web | Anything that changed this year |
| Long sums done in its head | Better than it was, still unreliable | Do the maths in a spreadsheet |
What do 24 product descriptions cost, by hand and with an LLM? (example figures)
A bridal boutique in Bendigo has 24 new gowns to list on its website. Written by hand, each description takes the owner about 20 minutes. With an LLM, the owner pastes in the facts for each gown (fabric, neckline, sizes, price), spends 2 minutes on the request and 6 minutes checking and fixing the draft.
| By hand | With an LLM | |
|---|---|---|
| Time per gown | 20 minutes | 8 minutes |
| Time for 24 gowns | 8 hours | 3.2 hours |
| Owner's time at $70 an hour | $560 | $224 |
| Chat app subscription | None | $35 for the month |
| Total | $560 | $259 |
The checking is where the saving is earned. If the fabric line is missing for one gown, the model fills the gap with whatever sounds bridal, and silk sounds bridal. A polyester satin gown listed as silk is a refund and a complaint waiting to happen.
Is there a large language model inside Cora?
Yes: CoraVision is built on large language models and gives them the thing a bare model lacks, which is knowledge of your business. In the onboarding call (about 25 minutes, nothing to prepare) Cora asks for your goal in your own words and your average sale. Once Google Ads, Meta and Google Analytics 4 are connected, CoraVision reads them every night, so Cora talks about your campaigns, using your figures. You make every change yourself.
Questions
Is ChatGPT a large language model?
ChatGPT is an app, and large language models made by OpenAI are the engine inside it. The app adds the chat screen, memory, file uploads and web search around the model. Gemini and Claude are the same arrangement from Google and Anthropic.
Does an LLM look things up on the internet?
Not on its own. A bare model answers from what it absorbed in training, which stops at a cut-off date. Apps that search the web run the search first and hand the results to the model to read, a method called retrieval-augmented generation.
How large is large?
Large refers to the number of adjustable values inside the model, called parameters, and to the amount of text it trained on. OpenAI's GPT-3, described in a 2020 paper, had 175 billion parameters. Most model makers no longer publish the figure for their newest models.